text
stringlengths 23
371k
| source
stringlengths 32
152
|
|---|---|
--
language: en
license: mit
library_name: timm
tags:
- pytorch
- image-classification
datasets:
- beans
metrics:
- accuracy
model-index:
- name: my-cool-model
results:
- task:
type: image-classification
dataset:
type: beans
name: Beans
metrics:
- type: accuracy
value: 0.9
- task:
type: image-classification
dataset:
type: beans
name: Beans
config: default
split: test
revision: 5503434ddd753f426f4b38109466949a1217c2bb
args:
date: 20220120
metrics:
- type: f1
value: 0.66
---
# my-cool-model
## Model description
This is a test model card with multiple evaluations across different (dataset, metric) configurations.
| huggingface/huggingface_hub/blob/main/tests/fixtures/cards/sample_simple_model_index.md |
--
license: mit
language: eo
thumbnail: https://huggingface.co/blog/assets/01_how-to-train/EsperBERTo-thumbnail-v2.png
widget:
- text: "Jen la komenco de bela <mask>."
- text: "Uno du <mask>"
- text: "Jen finiĝas bela <mask>."
---
# Hello old Windows line breaks
| huggingface/huggingface_hub/blob/main/tests/fixtures/cards/sample_windows_line_breaks.md |
ResNeXt
A **ResNeXt** repeats a [building block](https://paperswithcode.com/method/resnext-block) that aggregates a set of transformations with the same topology. Compared to a [ResNet](https://paperswithcode.com/method/resnet), it exposes a new dimension, *cardinality* (the size of the set of transformations) $C$, as an essential factor in addition to the dimensions of depth and width.
## How do I use this model on an image?
To load a pretrained model:
```python
import timm
model = timm.create_model('resnext101_32x8d', pretrained=True)
model.eval()
```
To load and preprocess the image:
```python
import urllib
from PIL import Image
from timm.data import resolve_data_config
from timm.data.transforms_factory import create_transform
config = resolve_data_config({}, model=model)
transform = create_transform(**config)
url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
urllib.request.urlretrieve(url, filename)
img = Image.open(filename).convert('RGB')
tensor = transform(img).unsqueeze(0) # transform and add batch dimension
```
To get the model predictions:
```python
import torch
with torch.no_grad():
out = model(tensor)
probabilities = torch.nn.functional.softmax(out[0], dim=0)
print(probabilities.shape)
# prints: torch.Size([1000])
```
To get the top-5 predictions class names:
```python
# Get imagenet class mappings
url, filename = ("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt", "imagenet_classes.txt")
urllib.request.urlretrieve(url, filename)
with open("imagenet_classes.txt", "r") as f:
categories = [s.strip() for s in f.readlines()]
# Print top categories per image
top5_prob, top5_catid = torch.topk(probabilities, 5)
for i in range(top5_prob.size(0)):
print(categories[top5_catid[i]], top5_prob[i].item())
# prints class names and probabilities like:
# [('Samoyed', 0.6425196528434753), ('Pomeranian', 0.04062102362513542), ('keeshond', 0.03186424449086189), ('white wolf', 0.01739676296710968), ('Eskimo dog', 0.011717947199940681)]
```
Replace the model name with the variant you want to use, e.g. `resnext101_32x8d`. You can find the IDs in the model summaries at the top of this page.
To extract image features with this model, follow the [timm feature extraction examples](https://rwightman.github.io/pytorch-image-models/feature_extraction/), just change the name of the model you want to use.
## How do I finetune this model?
You can finetune any of the pre-trained models just by changing the classifier (the last layer).
```python
model = timm.create_model('resnext101_32x8d', pretrained=True, num_classes=NUM_FINETUNE_CLASSES)
```
To finetune on your own dataset, you have to write a training loop or adapt [timm's training
script](https://github.com/rwightman/pytorch-image-models/blob/master/train.py) to use your dataset.
## How do I train this model?
You can follow the [timm recipe scripts](https://rwightman.github.io/pytorch-image-models/scripts/) for training a new model afresh.
## Citation
```BibTeX
@article{DBLP:journals/corr/XieGDTH16,
author = {Saining Xie and
Ross B. Girshick and
Piotr Doll{\'{a}}r and
Zhuowen Tu and
Kaiming He},
title = {Aggregated Residual Transformations for Deep Neural Networks},
journal = {CoRR},
volume = {abs/1611.05431},
year = {2016},
url = {http://arxiv.org/abs/1611.05431},
archivePrefix = {arXiv},
eprint = {1611.05431},
timestamp = {Mon, 13 Aug 2018 16:45:58 +0200},
biburl = {https://dblp.org/rec/journals/corr/XieGDTH16.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
```
<!--
Type: model-index
Collections:
- Name: ResNeXt
Paper:
Title: Aggregated Residual Transformations for Deep Neural Networks
URL: https://paperswithcode.com/paper/aggregated-residual-transformations-for-deep
Models:
- Name: resnext101_32x8d
In Collection: ResNeXt
Metadata:
FLOPs: 21180417024
Parameters: 88790000
File Size: 356082095
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: resnext101_32x8d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bilinear
Code: https://github.com/rwightman/pytorch-image-models/blob/b9843f954b0457af2db4f9dea41a8538f51f5d78/timm/models/resnet.py#L877
Weights: https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.3%
Top 5 Accuracy: 94.53%
- Name: resnext50_32x4d
In Collection: ResNeXt
Metadata:
FLOPs: 5472648192
Parameters: 25030000
File Size: 100435887
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: resnext50_32x4d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/b9843f954b0457af2db4f9dea41a8538f51f5d78/timm/models/resnet.py#L851
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnext50_32x4d_ra-d733960d.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.79%
Top 5 Accuracy: 94.61%
- Name: resnext50d_32x4d
In Collection: ResNeXt
Metadata:
FLOPs: 5781119488
Parameters: 25050000
File Size: 100515304
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: resnext50d_32x4d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/b9843f954b0457af2db4f9dea41a8538f51f5d78/timm/models/resnet.py#L869
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnext50d_32x4d-103e99f8.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.67%
Top 5 Accuracy: 94.87%
- Name: tv_resnext50_32x4d
In Collection: ResNeXt
Metadata:
FLOPs: 5472648192
Parameters: 25030000
File Size: 100441675
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Techniques:
- SGD with Momentum
- Weight Decay
Training Data:
- ImageNet
ID: tv_resnext50_32x4d
LR: 0.1
Epochs: 90
Crop Pct: '0.875'
LR Gamma: 0.1
Momentum: 0.9
Batch Size: 32
Image Size: '224'
LR Step Size: 30
Weight Decay: 0.0001
Interpolation: bilinear
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/resnet.py#L842
Weights: https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 77.61%
Top 5 Accuracy: 93.68%
--> | huggingface/pytorch-image-models/blob/main/docs/models/resnext.md |
!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Utilities for Tokenizers
This page lists all the utility functions used by the tokenizers, mainly the class
[`~tokenization_utils_base.PreTrainedTokenizerBase`] that implements the common methods between
[`PreTrainedTokenizer`] and [`PreTrainedTokenizerFast`] and the mixin
[`~tokenization_utils_base.SpecialTokensMixin`].
Most of those are only useful if you are studying the code of the tokenizers in the library.
## PreTrainedTokenizerBase
[[autodoc]] tokenization_utils_base.PreTrainedTokenizerBase
- __call__
- all
## SpecialTokensMixin
[[autodoc]] tokenization_utils_base.SpecialTokensMixin
## Enums and namedtuples
[[autodoc]] tokenization_utils_base.TruncationStrategy
[[autodoc]] tokenization_utils_base.CharSpan
[[autodoc]] tokenization_utils_base.TokenSpan
| huggingface/transformers/blob/main/docs/source/en/internal/tokenization_utils.md |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# IA3
This conceptual guide gives a brief overview of [IA3](https://arxiv.org/abs/2205.05638), a parameter-efficient fine tuning technique that is
intended to improve over [LoRA](./lora).
To make fine-tuning more efficient, IA3 (Infused Adapter by Inhibiting and Amplifying Inner Activations)
rescales inner activations with learned vectors. These learned vectors are injected in the attention and feedforward modules
in a typical transformer-based architecture. These learned vectors are the only trainable parameters during fine-tuning, and thus the original
weights remain frozen. Dealing with learned vectors (as opposed to learned low-rank updates to a weight matrix like LoRA)
keeps the number of trainable parameters much smaller.
Being similar to LoRA, IA3 carries many of the same advantages:
* IA3 makes fine-tuning more efficient by drastically reducing the number of trainable parameters. (For T0, an IA3 model only has about 0.01% trainable parameters, while even LoRA has > 0.1%)
* The original pre-trained weights are kept frozen, which means you can have multiple lightweight and portable IA3 models for various downstream tasks built on top of them.
* Performance of models fine-tuned using IA3 is comparable to the performance of fully fine-tuned models.
* IA3 does not add any inference latency because adapter weights can be merged with the base model.
In principle, IA3 can be applied to any subset of weight matrices in a neural network to reduce the number of trainable
parameters. Following the authors' implementation, IA3 weights are added to the key, value and feedforward layers
of a Transformer model. To be specific, for transformer models, IA3 weights are added to the outputs of key and value layers, and to the input of the second feedforward layer
in each transformer block.
Given the target layers for injecting IA3 parameters, the number of trainable parameters
can be determined based on the size of the weight matrices.
## Common IA3 parameters in PEFT
As with other methods supported by PEFT, to fine-tune a model using IA3, you need to:
1. Instantiate a base model.
2. Create a configuration (`IA3Config`) where you define IA3-specific parameters.
3. Wrap the base model with `get_peft_model()` to get a trainable `PeftModel`.
4. Train the `PeftModel` as you normally would train the base model.
`IA3Config` allows you to control how IA3 is applied to the base model through the following parameters:
- `target_modules`: The modules (for example, attention blocks) to apply the IA3 vectors.
- `feedforward_modules`: The list of modules to be treated as feedforward layers in `target_modules`. While learned vectors are multiplied with
the output activation for attention blocks, the vectors are multiplied with the input for classic feedforward layers. Note that `feedforward_modules` must be a subset of `target_modules`.
- `modules_to_save`: List of modules apart from IA3 layers to be set as trainable and saved in the final checkpoint. These typically include model's custom head that is randomly initialized for the fine-tuning task.
## Example Usage
For the task of sequence classification, one can initialize the IA3 config for a Llama model as follows:
```py
peft_config = IA3Config(
task_type=TaskType.SEQ_CLS, target_modules=["k_proj", "v_proj", "down_proj"], feedforward_modules=["down_proj"]
)
``` | huggingface/peft/blob/main/docs/source/conceptual_guides/ia3.md |
Sign in with Hugging Face
You can use the HF OAuth / OpenID connect flow to create a **"Sign in with HF"** flow in any website or App.
This will allow users to sign in to your website or app using their HF account, by clicking a button similar to this one:
After clicking this button your users will be presented with a permissions modal to authorize your app:
## Creating an oauth app
You can create your application in your [settings](https://huggingface.co/settings/applications/new):
### If you are hosting in Spaces
<Tip>
If you host your app on Spaces, then the flow will be even easier to implement (and built-in to Gradio directly); Check our [Spaces OAuth guide](https://huggingface.co/docs/hub/spaces-oauth).
</Tip>
## Currently supported scopes
The currently supported scopes are:
- `openid`: Get the ID token in addition to the access token.
- `profile`: Get the user's profile information (username, avatar, etc.)
- `email`: Get the user's email address.
- `read-repos`: Get read access to the user's personal repos.
- `write-repos`: Get write access to the user's personal repos. Does not grant read access on its own, you need to include `read-repos` as well.
- `manage-repos`: Get access to a repo's settings. Also grants repo creation and deletion.
- `inference-api`: Get access to the [Inference API](https://huggingface.co/docs/api-inference/index), you will be able to make inference requests on behalf of the user.
All other information is available in the [OpenID metadata](https://huggingface.co/.well-known/openid-configuration).
<Tip warning={true}>
Please contact us if you need any extra scopes.
</Tip>
## Branding
You are free to use your own design for the button. Below are some SVG images helpfully provided.
Check out [our badges](https://huggingface.co/datasets/huggingface/badges#sign-in-with-hugging-face) with explanations for integrating them in markdown or HTML.
[](https://huggingface.co/oauth/authorize?client_id=CLIENT_ID&redirect_uri=REDIRECT_URI&scope=openid%20profile&state=STATE)
[](https://huggingface.co/oauth/authorize?client_id=CLIENT_ID&redirect_uri=REDIRECT_URI&scope=openid%20profile&state=STATE)
[](https://huggingface.co/oauth/authorize?client_id=CLIENT_ID&redirect_uri=REDIRECT_URI&scope=openid%20profile&state=STATE)
[](https://huggingface.co/oauth/authorize?client_id=CLIENT_ID&redirect_uri=REDIRECT_URI&scope=openid%20profile&state=STATE)
[](https://huggingface.co/oauth/authorize?client_id=CLIENT_ID&redirect_uri=REDIRECT_URI&scope=openid%20profile&state=STATE)
[](https://huggingface.co/oauth/authorize?client_id=CLIENT_ID&redirect_uri=REDIRECT_URI&scope=openid%20profile&state=STATE)
[](https://huggingface.co/oauth/authorize?client_id=CLIENT_ID&redirect_uri=REDIRECT_URI&scope=openid%20profile&state=STATE)
[](https://huggingface.co/oauth/authorize?client_id=CLIENT_ID&redirect_uri=REDIRECT_URI&scope=openid%20profile&state=STATE)
| huggingface/hub-docs/blob/main/docs/hub/oauth.md |
FrameworkSwitchCourse {fw} />
# Token classification[[token-classification]]
{#if fw === 'pt'}
<CourseFloatingBanner chapter={7}
classNames="absolute z-10 right-0 top-0"
notebooks={[
{label: "Google Colab", value: "https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/en/chapter7/section2_pt.ipynb"},
{label: "Aws Studio", value: "https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/master/course/en/chapter7/section2_pt.ipynb"},
]} />
{:else}
<CourseFloatingBanner chapter={7}
classNames="absolute z-10 right-0 top-0"
notebooks={[
{label: "Google Colab", value: "https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/en/chapter7/section2_tf.ipynb"},
{label: "Aws Studio", value: "https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/master/course/en/chapter7/section2_tf.ipynb"},
]} />
{/if}
The first application we'll explore is token classification. This generic task encompasses any problem that can be formulated as "attributing a label to each token in a sentence," such as:
- **Named entity recognition (NER)**: Find the entities (such as persons, locations, or organizations) in a sentence. This can be formulated as attributing a label to each token by having one class per entity and one class for "no entity."
- **Part-of-speech tagging (POS)**: Mark each word in a sentence as corresponding to a particular part of speech (such as noun, verb, adjective, etc.).
- **Chunking**: Find the tokens that belong to the same entity. This task (which can be combined with POS or NER) can be formulated as attributing one label (usually `B-`) to any tokens that are at the beginning of a chunk, another label (usually `I-`) to tokens that are inside a chunk, and a third label (usually `O`) to tokens that don't belong to any chunk.
<Youtube id="wVHdVlPScxA"/>
Of course, there are many other types of token classification problem; those are just a few representative examples. In this section, we will fine-tune a model (BERT) on a NER task, which will then be able to compute predictions like this one:
<iframe src="https://course-demos-bert-finetuned-ner.hf.space" frameBorder="0" height="350" title="Gradio app" class="block dark:hidden container p-0 flex-grow space-iframe" allow="accelerometer; ambient-light-sensor; autoplay; battery; camera; document-domain; encrypted-media; fullscreen; geolocation; gyroscope; layout-animations; legacy-image-formats; magnetometer; microphone; midi; oversized-images; payment; picture-in-picture; publickey-credentials-get; sync-xhr; usb; vr ; wake-lock; xr-spatial-tracking" sandbox="allow-forms allow-modals allow-popups allow-popups-to-escape-sandbox allow-same-origin allow-scripts allow-downloads"></iframe>
<a class="flex justify-center" href="/huggingface-course/bert-finetuned-ner">
<img class="block dark:hidden lg:w-3/5" src="https://huggingface.co/datasets/huggingface-course/documentation-images/resolve/main/en/chapter7/model-eval-bert-finetuned-ner.png" alt="One-hot encoded labels for question answering."/>
<img class="hidden dark:block lg:w-3/5" src="https://huggingface.co/datasets/huggingface-course/documentation-images/resolve/main/en/chapter7/model-eval-bert-finetuned-ner-dark.png" alt="One-hot encoded labels for question answering."/>
</a>
You can find the model we'll train and upload to the Hub and double-check its predictions [here](https://huggingface.co/huggingface-course/bert-finetuned-ner?text=My+name+is+Sylvain+and+I+work+at+Hugging+Face+in+Brooklyn).
## Preparing the data[[preparing-the-data]]
First things first, we need a dataset suitable for token classification. In this section we will use the [CoNLL-2003 dataset](https://huggingface.co/datasets/conll2003), which contains news stories from Reuters.
<Tip>
💡 As long as your dataset consists of texts split into words with their corresponding labels, you will be able to adapt the data processing procedures described here to your own dataset. Refer back to [Chapter 5](/course/chapter5) if you need a refresher on how to load your own custom data in a `Dataset`.
</Tip>
### The CoNLL-2003 dataset[[the-conll-2003-dataset]]
To load the CoNLL-2003 dataset, we use the `load_dataset()` method from the 🤗 Datasets library:
```py
from datasets import load_dataset
raw_datasets = load_dataset("conll2003")
```
This will download and cache the dataset, like we saw in [Chapter 3](/course/chapter3) for the GLUE MRPC dataset. Inspecting this object shows us the columns present and the split between the training, validation, and test sets:
```py
raw_datasets
```
```python out
DatasetDict({
train: Dataset({
features: ['chunk_tags', 'id', 'ner_tags', 'pos_tags', 'tokens'],
num_rows: 14041
})
validation: Dataset({
features: ['chunk_tags', 'id', 'ner_tags', 'pos_tags', 'tokens'],
num_rows: 3250
})
test: Dataset({
features: ['chunk_tags', 'id', 'ner_tags', 'pos_tags', 'tokens'],
num_rows: 3453
})
})
```
In particular, we can see the dataset contains labels for the three tasks we mentioned earlier: NER, POS, and chunking. A big difference from other datasets is that the input texts are not presented as sentences or documents, but lists of words (the last column is called `tokens`, but it contains words in the sense that these are pre-tokenized inputs that still need to go through the tokenizer for subword tokenization).
Let's have a look at the first element of the training set:
```py
raw_datasets["train"][0]["tokens"]
```
```python out
['EU', 'rejects', 'German', 'call', 'to', 'boycott', 'British', 'lamb', '.']
```
Since we want to perform named entity recognition, we will look at the NER tags:
```py
raw_datasets["train"][0]["ner_tags"]
```
```python out
[3, 0, 7, 0, 0, 0, 7, 0, 0]
```
Those are the labels as integers ready for training, but they're not necessarily useful when we want to inspect the data. Like for text classification, we can access the correspondence between those integers and the label names by looking at the `features` attribute of our dataset:
```py
ner_feature = raw_datasets["train"].features["ner_tags"]
ner_feature
```
```python out
Sequence(feature=ClassLabel(num_classes=9, names=['O', 'B-PER', 'I-PER', 'B-ORG', 'I-ORG', 'B-LOC', 'I-LOC', 'B-MISC', 'I-MISC'], names_file=None, id=None), length=-1, id=None)
```
So this column contains elements that are sequences of `ClassLabel`s. The type of the elements of the sequence is in the `feature` attribute of this `ner_feature`, and we can access the list of names by looking at the `names` attribute of that `feature`:
```py
label_names = ner_feature.feature.names
label_names
```
```python out
['O', 'B-PER', 'I-PER', 'B-ORG', 'I-ORG', 'B-LOC', 'I-LOC', 'B-MISC', 'I-MISC']
```
We already saw these labels when digging into the `token-classification` pipeline in [Chapter 6](/course/chapter6/3), but for a quick refresher:
- `O` means the word doesn't correspond to any entity.
- `B-PER`/`I-PER` means the word corresponds to the beginning of/is inside a *person* entity.
- `B-ORG`/`I-ORG` means the word corresponds to the beginning of/is inside an *organization* entity.
- `B-LOC`/`I-LOC` means the word corresponds to the beginning of/is inside a *location* entity.
- `B-MISC`/`I-MISC` means the word corresponds to the beginning of/is inside a *miscellaneous* entity.
Now decoding the labels we saw earlier gives us this:
```python
words = raw_datasets["train"][0]["tokens"]
labels = raw_datasets["train"][0]["ner_tags"]
line1 = ""
line2 = ""
for word, label in zip(words, labels):
full_label = label_names[label]
max_length = max(len(word), len(full_label))
line1 += word + " " * (max_length - len(word) + 1)
line2 += full_label + " " * (max_length - len(full_label) + 1)
print(line1)
print(line2)
```
```python out
'EU rejects German call to boycott British lamb .'
'B-ORG O B-MISC O O O B-MISC O O'
```
And for an example mixing `B-` and `I-` labels, here's what the same code gives us on the element of the training set at index 4:
```python out
'Germany \'s representative to the European Union \'s veterinary committee Werner Zwingmann said on Wednesday consumers should buy sheepmeat from countries other than Britain until the scientific advice was clearer .'
'B-LOC O O O O B-ORG I-ORG O O O B-PER I-PER O O O O O O O O O O O B-LOC O O O O O O O'
```
As we can see, entities spanning two words, like "European Union" and "Werner Zwingmann," are attributed a `B-` label for the first word and an `I-` label for the second.
<Tip>
✏️ **Your turn!** Print the same two sentences with their POS or chunking labels.
</Tip>
### Processing the data[[processing-the-data]]
<Youtube id="iY2AZYdZAr0"/>
As usual, our texts need to be converted to token IDs before the model can make sense of them. As we saw in [Chapter 6](/course/chapter6/), a big difference in the case of token classification tasks is that we have pre-tokenized inputs. Fortunately, the tokenizer API can deal with that pretty easily; we just need to warn the `tokenizer` with a special flag.
To begin, let's create our `tokenizer` object. As we said before, we will be using a BERT pretrained model, so we'll start by downloading and caching the associated tokenizer:
```python
from transformers import AutoTokenizer
model_checkpoint = "bert-base-cased"
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
```
You can replace the `model_checkpoint` with any other model you prefer from the [Hub](https://huggingface.co/models), or with a local folder in which you've saved a pretrained model and a tokenizer. The only constraint is that the tokenizer needs to be backed by the 🤗 Tokenizers library, so there's a "fast" version available. You can see all the architectures that come with a fast version in [this big table](https://huggingface.co/transformers/#supported-frameworks), and to check that the `tokenizer` object you're using is indeed backed by 🤗 Tokenizers you can look at its `is_fast` attribute:
```py
tokenizer.is_fast
```
```python out
True
```
To tokenize a pre-tokenized input, we can use our `tokenizer` as usual and just add `is_split_into_words=True`:
```py
inputs = tokenizer(raw_datasets["train"][0]["tokens"], is_split_into_words=True)
inputs.tokens()
```
```python out
['[CLS]', 'EU', 'rejects', 'German', 'call', 'to', 'boycott', 'British', 'la', '##mb', '.', '[SEP]']
```
As we can see, the tokenizer added the special tokens used by the model (`[CLS]` at the beginning and `[SEP]` at the end) and left most of the words untouched. The word `lamb`, however, was tokenized into two subwords, `la` and `##mb`. This introduces a mismatch between our inputs and the labels: the list of labels has only 9 elements, whereas our input now has 12 tokens. Accounting for the special tokens is easy (we know they are at the beginning and the end), but we also need to make sure we align all the labels with the proper words.
Fortunately, because we're using a fast tokenizer we have access to the 🤗 Tokenizers superpowers, which means we can easily map each token to its corresponding word (as seen in [Chapter 6](/course/chapter6/3)):
```py
inputs.word_ids()
```
```python out
[None, 0, 1, 2, 3, 4, 5, 6, 7, 7, 8, None]
```
With a tiny bit of work, we can then expand our label list to match the tokens. The first rule we'll apply is that special tokens get a label of `-100`. This is because by default `-100` is an index that is ignored in the loss function we will use (cross entropy). Then, each token gets the same label as the token that started the word it's inside, since they are part of the same entity. For tokens inside a word but not at the beginning, we replace the `B-` with `I-` (since the token does not begin the entity):
```python
def align_labels_with_tokens(labels, word_ids):
new_labels = []
current_word = None
for word_id in word_ids:
if word_id != current_word:
# Start of a new word!
current_word = word_id
label = -100 if word_id is None else labels[word_id]
new_labels.append(label)
elif word_id is None:
# Special token
new_labels.append(-100)
else:
# Same word as previous token
label = labels[word_id]
# If the label is B-XXX we change it to I-XXX
if label % 2 == 1:
label += 1
new_labels.append(label)
return new_labels
```
Let's try it out on our first sentence:
```py
labels = raw_datasets["train"][0]["ner_tags"]
word_ids = inputs.word_ids()
print(labels)
print(align_labels_with_tokens(labels, word_ids))
```
```python out
[3, 0, 7, 0, 0, 0, 7, 0, 0]
[-100, 3, 0, 7, 0, 0, 0, 7, 0, 0, 0, -100]
```
As we can see, our function added the `-100` for the two special tokens at the beginning and the end, and a new `0` for our word that was split into two tokens.
<Tip>
✏️ **Your turn!** Some researchers prefer to attribute only one label per word, and assign `-100` to the other subtokens in a given word. This is to avoid long words that split into lots of subtokens contributing heavily to the loss. Change the previous function to align labels with input IDs by following this rule.
</Tip>
To preprocess our whole dataset, we need to tokenize all the inputs and apply `align_labels_with_tokens()` on all the labels. To take advantage of the speed of our fast tokenizer, it's best to tokenize lots of texts at the same time, so we'll write a function that processes a list of examples and use the `Dataset.map()` method with the option `batched=True`. The only thing that is different from our previous example is that the `word_ids()` function needs to get the index of the example we want the word IDs of when the inputs to the tokenizer are lists of texts (or in our case, list of lists of words), so we add that too:
```py
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples["tokens"], truncation=True, is_split_into_words=True
)
all_labels = examples["ner_tags"]
new_labels = []
for i, labels in enumerate(all_labels):
word_ids = tokenized_inputs.word_ids(i)
new_labels.append(align_labels_with_tokens(labels, word_ids))
tokenized_inputs["labels"] = new_labels
return tokenized_inputs
```
Note that we haven't padded our inputs yet; we'll do that later, when creating the batches with a data collator.
We can now apply all that preprocessing in one go on the other splits of our dataset:
```py
tokenized_datasets = raw_datasets.map(
tokenize_and_align_labels,
batched=True,
remove_columns=raw_datasets["train"].column_names,
)
```
We've done the hardest part! Now that the data has been preprocessed, the actual training will look a lot like what we did in [Chapter 3](/course/chapter3).
{#if fw === 'pt'}
## Fine-tuning the model with the `Trainer` API[[fine-tuning-the-model-with-the-trainer-api]]
The actual code using the `Trainer` will be the same as before; the only changes are the way the data is collated into a batch and the metric computation function.
{:else}
## Fine-tuning the model with Keras[[fine-tuning-the-model-with-keras]]
The actual code using Keras will be very similar to before; the only changes are the way the data is collated into a batch and the metric computation function.
{/if}
### Data collation[[data-collation]]
We can't just use a `DataCollatorWithPadding` like in [Chapter 3](/course/chapter3) because that only pads the inputs (input IDs, attention mask, and token type IDs). Here our labels should be padded the exact same way as the inputs so that they stay the same size, using `-100` as a value so that the corresponding predictions are ignored in the loss computation.
This is all done by a [`DataCollatorForTokenClassification`](https://huggingface.co/transformers/main_classes/data_collator.html#datacollatorfortokenclassification). Like the `DataCollatorWithPadding`, it takes the `tokenizer` used to preprocess the inputs:
{#if fw === 'pt'}
```py
from transformers import DataCollatorForTokenClassification
data_collator = DataCollatorForTokenClassification(tokenizer=tokenizer)
```
{:else}
```py
from transformers import DataCollatorForTokenClassification
data_collator = DataCollatorForTokenClassification(
tokenizer=tokenizer, return_tensors="tf"
)
```
{/if}
To test this on a few samples, we can just call it on a list of examples from our tokenized training set:
```py
batch = data_collator([tokenized_datasets["train"][i] for i in range(2)])
batch["labels"]
```
```python out
tensor([[-100, 3, 0, 7, 0, 0, 0, 7, 0, 0, 0, -100],
[-100, 1, 2, -100, -100, -100, -100, -100, -100, -100, -100, -100]])
```
Let's compare this to the labels for the first and second elements in our dataset:
```py
for i in range(2):
print(tokenized_datasets["train"][i]["labels"])
```
```python out
[-100, 3, 0, 7, 0, 0, 0, 7, 0, 0, 0, -100]
[-100, 1, 2, -100]
```
{#if fw === 'pt'}
As we can see, the second set of labels has been padded to the length of the first one using `-100`s.
{:else}
Our data collator is ready to go! Now let's use it to make a `tf.data.Dataset` with the `to_tf_dataset()` method. You can also use `model.prepare_tf_dataset()` to do this with a bit less boilerplate code - you'll see this in some of the other sections of this chapter.
```py
tf_train_dataset = tokenized_datasets["train"].to_tf_dataset(
columns=["attention_mask", "input_ids", "labels", "token_type_ids"],
collate_fn=data_collator,
shuffle=True,
batch_size=16,
)
tf_eval_dataset = tokenized_datasets["validation"].to_tf_dataset(
columns=["attention_mask", "input_ids", "labels", "token_type_ids"],
collate_fn=data_collator,
shuffle=False,
batch_size=16,
)
```
Next stop: the model itself.
{/if}
{#if fw === 'tf'}
### Defining the model[[defining-the-model]]
Since we are working on a token classification problem, we will use the `TFAutoModelForTokenClassification` class. The main thing to remember when defining this model is to pass along some information on the number of labels we have. The easiest way to do this is to pass that number with the `num_labels` argument, but if we want a nice inference widget working like the one we saw at the beginning of this section, it's better to set the correct label correspondences instead.
They should be set by two dictionaries, `id2label` and `label2id`, which contain the mapping from ID to label and vice versa:
```py
id2label = {i: label for i, label in enumerate(label_names)}
label2id = {v: k for k, v in id2label.items()}
```
Now we can just pass them to the `TFAutoModelForTokenClassification.from_pretrained()` method, and they will be set in the model's configuration, then properly saved and uploaded to the Hub:
```py
from transformers import TFAutoModelForTokenClassification
model = TFAutoModelForTokenClassification.from_pretrained(
model_checkpoint,
id2label=id2label,
label2id=label2id,
)
```
Like when we defined our `TFAutoModelForSequenceClassification` in [Chapter 3](/course/chapter3), creating the model issues a warning that some weights were not used (the ones from the pretraining head) and some other weights are randomly initialized (the ones from the new token classification head), and that this model should be trained. We will do that in a minute, but first let's double-check that our model has the right number of labels:
```python
model.config.num_labels
```
```python out
9
```
<Tip warning={true}>
⚠️ If you have a model with the wrong number of labels, you will get an obscure error when calling `model.fit()` later. This can be annoying to debug, so make sure you do this check to confirm you have the expected number of labels.
</Tip>
### Fine-tuning the model[[fine-tuning-the-model]]
We are now ready to train our model! We have just a little more housekeeping to do first, though: we should log in to Hugging Face and define our training hyperparameters. If you're working in a notebook, there's a convenience function to help you with this:
```python
from huggingface_hub import notebook_login
notebook_login()
```
This will display a widget where you can enter your Hugging Face login credentials.
If you aren't working in a notebook, just type the following line in your terminal:
```bash
huggingface-cli login
```
After logging in, we can prepare everything we need to compile our model. 🤗 Transformers provides a convenient `create_optimizer()` function that will give you an `AdamW` optimizer with appropriate settings for the weight decay and learning rate decay, both of which will improve your model's performance compared to the built-in `Adam` optimizer:
```python
from transformers import create_optimizer
import tensorflow as tf
# Train in mixed-precision float16
# Comment this line out if you're using a GPU that will not benefit from this
tf.keras.mixed_precision.set_global_policy("mixed_float16")
# The number of training steps is the number of samples in the dataset, divided by the batch size then multiplied
# by the total number of epochs. Note that the tf_train_dataset here is a batched tf.data.Dataset,
# not the original Hugging Face Dataset, so its len() is already num_samples // batch_size.
num_epochs = 3
num_train_steps = len(tf_train_dataset) * num_epochs
optimizer, schedule = create_optimizer(
init_lr=2e-5,
num_warmup_steps=0,
num_train_steps=num_train_steps,
weight_decay_rate=0.01,
)
model.compile(optimizer=optimizer)
```
Note also that we don't supply a `loss` argument to `compile()`. This is because the models can actually compute loss internally -- if you compile without a loss and supply your labels in the input dictionary (as we do in our datasets), then the model will train using that internal loss, which will be appropriate for the task and model type you have chosen.
Next, we define a `PushToHubCallback` to upload our model to the Hub during training, and fit the model with that callback:
```python
from transformers.keras_callbacks import PushToHubCallback
callback = PushToHubCallback(output_dir="bert-finetuned-ner", tokenizer=tokenizer)
model.fit(
tf_train_dataset,
validation_data=tf_eval_dataset,
callbacks=[callback],
epochs=num_epochs,
)
```
You can specify the full name of the repository you want to push to with the `hub_model_id` argument (in particular, you will have to use this argument to push to an organization). For instance, when we pushed the model to the [`huggingface-course` organization](https://huggingface.co/huggingface-course), we added `hub_model_id="huggingface-course/bert-finetuned-ner"`. By default, the repository used will be in your namespace and named after the output directory you set, for example `"cool_huggingface_user/bert-finetuned-ner"`.
<Tip>
💡 If the output directory you are using already exists, it needs to be a local clone of the repository you want to push to. If it isn't, you'll get an error when calling `model.fit()` and will need to set a new name.
</Tip>
Note that while the training happens, each time the model is saved (here, every epoch) it is uploaded to the Hub in the background. This way, you will be able to to resume your training on another machine if necessary.
At this stage, you can use the inference widget on the Model Hub to test your model and share it with your friends. You have successfully fine-tuned a model on a token classification task -- congratulations! But how good is our model, really? We should evaluate some metrics to find out.
{/if}
### Metrics[[metrics]]
{#if fw === 'pt'}
To have the `Trainer` compute a metric every epoch, we will need to define a `compute_metrics()` function that takes the arrays of predictions and labels, and returns a dictionary with the metric names and values.
The traditional framework used to evaluate token classification prediction is [*seqeval*](https://github.com/chakki-works/seqeval). To use this metric, we first need to install the *seqeval* library:
```py
!pip install seqeval
```
We can then load it via the `evaluate.load()` function like we did in [Chapter 3](/course/chapter3):
{:else}
The traditional framework used to evaluate token classification prediction is [*seqeval*](https://github.com/chakki-works/seqeval). To use this metric, we first need to install the *seqeval* library:
```py
!pip install seqeval
```
We can then load it via the `evaluate.load()` function like we did in [Chapter 3](/course/chapter3):
{/if}
```py
import evaluate
metric = evaluate.load("seqeval")
```
This metric does not behave like the standard accuracy: it will actually take the lists of labels as strings, not integers, so we will need to fully decode the predictions and labels before passing them to the metric. Let's see how it works. First, we'll get the labels for our first training example:
```py
labels = raw_datasets["train"][0]["ner_tags"]
labels = [label_names[i] for i in labels]
labels
```
```python out
['B-ORG', 'O', 'B-MISC', 'O', 'O', 'O', 'B-MISC', 'O', 'O']
```
We can then create fake predictions for those by just changing the value at index 2:
```py
predictions = labels.copy()
predictions[2] = "O"
metric.compute(predictions=[predictions], references=[labels])
```
Note that the metric takes a list of predictions (not just one) and a list of labels. Here's the output:
```python out
{'MISC': {'precision': 1.0, 'recall': 0.5, 'f1': 0.67, 'number': 2},
'ORG': {'precision': 1.0, 'recall': 1.0, 'f1': 1.0, 'number': 1},
'overall_precision': 1.0,
'overall_recall': 0.67,
'overall_f1': 0.8,
'overall_accuracy': 0.89}
```
{#if fw === 'pt'}
This is sending back a lot of information! We get the precision, recall, and F1 score for each separate entity, as well as overall. For our metric computation we will only keep the overall score, but feel free to tweak the `compute_metrics()` function to return all the metrics you would like reported.
This `compute_metrics()` function first takes the argmax of the logits to convert them to predictions (as usual, the logits and the probabilities are in the same order, so we don't need to apply the softmax). Then we have to convert both labels and predictions from integers to strings. We remove all the values where the label is `-100`, then pass the results to the `metric.compute()` method:
```py
import numpy as np
def compute_metrics(eval_preds):
logits, labels = eval_preds
predictions = np.argmax(logits, axis=-1)
# Remove ignored index (special tokens) and convert to labels
true_labels = [[label_names[l] for l in label if l != -100] for label in labels]
true_predictions = [
[label_names[p] for (p, l) in zip(prediction, label) if l != -100]
for prediction, label in zip(predictions, labels)
]
all_metrics = metric.compute(predictions=true_predictions, references=true_labels)
return {
"precision": all_metrics["overall_precision"],
"recall": all_metrics["overall_recall"],
"f1": all_metrics["overall_f1"],
"accuracy": all_metrics["overall_accuracy"],
}
```
Now that this is done, we are almost ready to define our `Trainer`. We just need a `model` to fine-tune!
{:else}
This is sending back a lot of information! We get the precision, recall, and F1 score for each separate entity, as well as overall. Now let's see what happens if we try using our actual model predictions to compute some real scores.
TensorFlow doesn't like concatenating our predictions together, because they have variable sequence lengths. This means we can't just use `model.predict()` -- but that's not going to stop us. We'll get some predictions a batch at a time and concatenate them into one big long list as we go, dropping the `-100` tokens that indicate masking/padding, then compute metrics on the list at the end:
```py
import numpy as np
all_predictions = []
all_labels = []
for batch in tf_eval_dataset:
logits = model.predict_on_batch(batch)["logits"]
labels = batch["labels"]
predictions = np.argmax(logits, axis=-1)
for prediction, label in zip(predictions, labels):
for predicted_idx, label_idx in zip(prediction, label):
if label_idx == -100:
continue
all_predictions.append(label_names[predicted_idx])
all_labels.append(label_names[label_idx])
metric.compute(predictions=[all_predictions], references=[all_labels])
```
```python out
{'LOC': {'precision': 0.91, 'recall': 0.92, 'f1': 0.91, 'number': 1668},
'MISC': {'precision': 0.70, 'recall': 0.79, 'f1': 0.74, 'number': 702},
'ORG': {'precision': 0.85, 'recall': 0.90, 'f1': 0.88, 'number': 1661},
'PER': {'precision': 0.95, 'recall': 0.95, 'f1': 0.95, 'number': 1617},
'overall_precision': 0.87,
'overall_recall': 0.91,
'overall_f1': 0.89,
'overall_accuracy': 0.97}
```
How did your model do, compared to ours? If you got similar numbers, your training was a success!
{/if}
{#if fw === 'pt'}
### Defining the model[[defining-the-model]]
Since we are working on a token classification problem, we will use the `AutoModelForTokenClassification` class. The main thing to remember when defining this model is to pass along some information on the number of labels we have. The easiest way to do this is to pass that number with the `num_labels` argument, but if we want a nice inference widget working like the one we saw at the beginning of this section, it's better to set the correct label correspondences instead.
They should be set by two dictionaries, `id2label` and `label2id`, which contain the mappings from ID to label and vice versa:
```py
id2label = {i: label for i, label in enumerate(label_names)}
label2id = {v: k for k, v in id2label.items()}
```
Now we can just pass them to the `AutoModelForTokenClassification.from_pretrained()` method, and they will be set in the model's configuration and then properly saved and uploaded to the Hub:
```py
from transformers import AutoModelForTokenClassification
model = AutoModelForTokenClassification.from_pretrained(
model_checkpoint,
id2label=id2label,
label2id=label2id,
)
```
Like when we defined our `AutoModelForSequenceClassification` in [Chapter 3](/course/chapter3), creating the model issues a warning that some weights were not used (the ones from the pretraining head) and some other weights are randomly initialized (the ones from the new token classification head), and that this model should be trained. We will do that in a minute, but first let's double-check that our model has the right number of labels:
```python
model.config.num_labels
```
```python out
9
```
<Tip warning={true}>
⚠️ If you have a model with the wrong number of labels, you will get an obscure error when calling the `Trainer.train()` method later on (something like "CUDA error: device-side assert triggered"). This is the number one cause of bugs reported by users for such errors, so make sure you do this check to confirm that you have the expected number of labels.
</Tip>
### Fine-tuning the model[[fine-tuning-the-model]]
We are now ready to train our model! We just need to do two last things before we define our `Trainer`: log in to Hugging Face and define our training arguments. If you're working in a notebook, there's a convenience function to help you with this:
```python
from huggingface_hub import notebook_login
notebook_login()
```
This will display a widget where you can enter your Hugging Face login credentials.
If you aren't working in a notebook, just type the following line in your terminal:
```bash
huggingface-cli login
```
Once this is done, we can define our `TrainingArguments`:
```python
from transformers import TrainingArguments
args = TrainingArguments(
"bert-finetuned-ner",
evaluation_strategy="epoch",
save_strategy="epoch",
learning_rate=2e-5,
num_train_epochs=3,
weight_decay=0.01,
push_to_hub=True,
)
```
You've seen most of those before: we set some hyperparameters (like the learning rate, the number of epochs to train for, and the weight decay), and we specify `push_to_hub=True` to indicate that we want to save the model and evaluate it at the end of every epoch, and that we want to upload our results to the Model Hub. Note that you can specify the name of the repository you want to push to with the `hub_model_id` argument (in particular, you will have to use this argument to push to an organization). For instance, when we pushed the model to the [`huggingface-course` organization](https://huggingface.co/huggingface-course), we added `hub_model_id="huggingface-course/bert-finetuned-ner"` to `TrainingArguments`. By default, the repository used will be in your namespace and named after the output directory you set, so in our case it will be `"sgugger/bert-finetuned-ner"`.
<Tip>
💡 If the output directory you are using already exists, it needs to be a local clone of the repository you want to push to. If it isn't, you'll get an error when defining your `Trainer` and will need to set a new name.
</Tip>
Finally, we just pass everything to the `Trainer` and launch the training:
```python
from transformers import Trainer
trainer = Trainer(
model=model,
args=args,
train_dataset=tokenized_datasets["train"],
eval_dataset=tokenized_datasets["validation"],
data_collator=data_collator,
compute_metrics=compute_metrics,
tokenizer=tokenizer,
)
trainer.train()
```
Note that while the training happens, each time the model is saved (here, every epoch) it is uploaded to the Hub in the background. This way, you will be able to to resume your training on another machine if necessary.
Once the training is complete, we use the `push_to_hub()` method to make sure we upload the most recent version of the model:
```py
trainer.push_to_hub(commit_message="Training complete")
```
This command returns the URL of the commit it just did, if you want to inspect it:
```python out
'https://huggingface.co/sgugger/bert-finetuned-ner/commit/26ab21e5b1568f9afeccdaed2d8715f571d786ed'
```
The `Trainer` also drafts a model card with all the evaluation results and uploads it. At this stage, you can use the inference widget on the Model Hub to test your model and share it with your friends. You have successfully fine-tuned a model on a token classification task -- congratulations!
If you want to dive a bit more deeply into the training loop, we will now show you how to do the same thing using 🤗 Accelerate.
## A custom training loop[[a-custom-training-loop]]
Let's now take a look at the full training loop, so you can easily customize the parts you need. It will look a lot like what we did in [Chapter 3](/course/chapter3/4), with a few changes for the evaluation.
### Preparing everything for training[[preparing-everything-for-training]]
First we need to build the `DataLoader`s from our datasets. We'll reuse our `data_collator` as a `collate_fn` and shuffle the training set, but not the validation set:
```py
from torch.utils.data import DataLoader
train_dataloader = DataLoader(
tokenized_datasets["train"],
shuffle=True,
collate_fn=data_collator,
batch_size=8,
)
eval_dataloader = DataLoader(
tokenized_datasets["validation"], collate_fn=data_collator, batch_size=8
)
```
Next we reinstantiate our model, to make sure we're not continuing the fine-tuning from before but starting from the BERT pretrained model again:
```py
model = AutoModelForTokenClassification.from_pretrained(
model_checkpoint,
id2label=id2label,
label2id=label2id,
)
```
Then we will need an optimizer. We'll use the classic `AdamW`, which is like `Adam`, but with a fix in the way weight decay is applied:
```py
from torch.optim import AdamW
optimizer = AdamW(model.parameters(), lr=2e-5)
```
Once we have all those objects, we can send them to the `accelerator.prepare()` method:
```py
from accelerate import Accelerator
accelerator = Accelerator()
model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader
)
```
<Tip>
🚨 If you're training on a TPU, you'll need to move all the code starting from the cell above into a dedicated training function. See [Chapter 3](/course/chapter3) for more details.
</Tip>
Now that we have sent our `train_dataloader` to `accelerator.prepare()`, we can use its length to compute the number of training steps. Remember that we should always do this after preparing the dataloader, as that method will change its length. We use a classic linear schedule from the learning rate to 0:
```py
from transformers import get_scheduler
num_train_epochs = 3
num_update_steps_per_epoch = len(train_dataloader)
num_training_steps = num_train_epochs * num_update_steps_per_epoch
lr_scheduler = get_scheduler(
"linear",
optimizer=optimizer,
num_warmup_steps=0,
num_training_steps=num_training_steps,
)
```
Lastly, to push our model to the Hub, we will need to create a `Repository` object in a working folder. First log in to Hugging Face, if you're not logged in already. We'll determine the repository name from the model ID we want to give our model (feel free to replace the `repo_name` with your own choice; it just needs to contain your username, which is what the function `get_full_repo_name()` does):
```py
from huggingface_hub import Repository, get_full_repo_name
model_name = "bert-finetuned-ner-accelerate"
repo_name = get_full_repo_name(model_name)
repo_name
```
```python out
'sgugger/bert-finetuned-ner-accelerate'
```
Then we can clone that repository in a local folder. If it already exists, this local folder should be an existing clone of the repository we are working with:
```py
output_dir = "bert-finetuned-ner-accelerate"
repo = Repository(output_dir, clone_from=repo_name)
```
We can now upload anything we save in `output_dir` by calling the `repo.push_to_hub()` method. This will help us upload the intermediate models at the end of each epoch.
### Training loop[[training-loop]]
We are now ready to write the full training loop. To simplify its evaluation part, we define this `postprocess()` function that takes predictions and labels and converts them to lists of strings, like our `metric` object expects:
```py
def postprocess(predictions, labels):
predictions = predictions.detach().cpu().clone().numpy()
labels = labels.detach().cpu().clone().numpy()
# Remove ignored index (special tokens) and convert to labels
true_labels = [[label_names[l] for l in label if l != -100] for label in labels]
true_predictions = [
[label_names[p] for (p, l) in zip(prediction, label) if l != -100]
for prediction, label in zip(predictions, labels)
]
return true_labels, true_predictions
```
Then we can write the training loop. After defining a progress bar to follow how training goes, the loop has three parts:
- The training in itself, which is the classic iteration over the `train_dataloader`, forward pass through the model, then backward pass and optimizer step.
- The evaluation, in which there is a novelty after getting the outputs of our model on a batch: since two processes may have padded the inputs and labels to different shapes, we need to use `accelerator.pad_across_processes()` to make the predictions and labels the same shape before calling the `gather()` method. If we don't do this, the evaluation will either error out or hang forever. Then we send the results to `metric.add_batch()` and call `metric.compute()` once the evaluation loop is over.
- Saving and uploading, where we first save the model and the tokenizer, then call `repo.push_to_hub()`. Notice that we use the argument `blocking=False` to tell the 🤗 Hub library to push in an asynchronous process. This way, training continues normally and this (long) instruction is executed in the background.
Here's the complete code for the training loop:
```py
from tqdm.auto import tqdm
import torch
progress_bar = tqdm(range(num_training_steps))
for epoch in range(num_train_epochs):
# Training
model.train()
for batch in train_dataloader:
outputs = model(**batch)
loss = outputs.loss
accelerator.backward(loss)
optimizer.step()
lr_scheduler.step()
optimizer.zero_grad()
progress_bar.update(1)
# Evaluation
model.eval()
for batch in eval_dataloader:
with torch.no_grad():
outputs = model(**batch)
predictions = outputs.logits.argmax(dim=-1)
labels = batch["labels"]
# Necessary to pad predictions and labels for being gathered
predictions = accelerator.pad_across_processes(predictions, dim=1, pad_index=-100)
labels = accelerator.pad_across_processes(labels, dim=1, pad_index=-100)
predictions_gathered = accelerator.gather(predictions)
labels_gathered = accelerator.gather(labels)
true_predictions, true_labels = postprocess(predictions_gathered, labels_gathered)
metric.add_batch(predictions=true_predictions, references=true_labels)
results = metric.compute()
print(
f"epoch {epoch}:",
{
key: results[f"overall_{key}"]
for key in ["precision", "recall", "f1", "accuracy"]
},
)
# Save and upload
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(output_dir, save_function=accelerator.save)
if accelerator.is_main_process:
tokenizer.save_pretrained(output_dir)
repo.push_to_hub(
commit_message=f"Training in progress epoch {epoch}", blocking=False
)
```
In case this is the first time you're seeing a model saved with 🤗 Accelerate, let's take a moment to inspect the three lines of code that go with it:
```py
accelerator.wait_for_everyone()
unwrapped_model = accelerator.unwrap_model(model)
unwrapped_model.save_pretrained(output_dir, save_function=accelerator.save)
```
The first line is self-explanatory: it tells all the processes to wait until everyone is at that stage before continuing. This is to make sure we have the same model in every process before saving. Then we grab the `unwrapped_model`, which is the base model we defined. The `accelerator.prepare()` method changes the model to work in distributed training, so it won't have the `save_pretrained()` method anymore; the `accelerator.unwrap_model()` method undoes that step. Lastly, we call `save_pretrained()` but tell that method to use `accelerator.save()` instead of `torch.save()`.
Once this is done, you should have a model that produces results pretty similar to the one trained with the `Trainer`. You can check the model we trained using this code at [*huggingface-course/bert-finetuned-ner-accelerate*](https://huggingface.co/huggingface-course/bert-finetuned-ner-accelerate). And if you want to test out any tweaks to the training loop, you can directly implement them by editing the code shown above!
{/if}
## Using the fine-tuned model[[using-the-fine-tuned-model]]
We've already shown you how you can use the model we fine-tuned on the Model Hub with the inference widget. To use it locally in a `pipeline`, you just have to specify the proper model identifier:
```py
from transformers import pipeline
# Replace this with your own checkpoint
model_checkpoint = "huggingface-course/bert-finetuned-ner"
token_classifier = pipeline(
"token-classification", model=model_checkpoint, aggregation_strategy="simple"
)
token_classifier("My name is Sylvain and I work at Hugging Face in Brooklyn.")
```
```python out
[{'entity_group': 'PER', 'score': 0.9988506, 'word': 'Sylvain', 'start': 11, 'end': 18},
{'entity_group': 'ORG', 'score': 0.9647625, 'word': 'Hugging Face', 'start': 33, 'end': 45},
{'entity_group': 'LOC', 'score': 0.9986118, 'word': 'Brooklyn', 'start': 49, 'end': 57}]
```
Great! Our model is working as well as the default one for this pipeline!
| huggingface/course/blob/main/chapters/en/chapter7/2.mdx |
n this video we take a look at setting up a custom loss function for training. In the default loss functions all samples such as these code snippets are treated the same irrespective of their content, but there are scenarios where you it could make sense to weight the samples differently. If for example one sample contains a lot of tokens that or of interest to us or it has a favourable diversity of tokens. We can also think of other heuristics we can implement with pattern matching or other rules. For each sample we get a loss value during training and we can combine that loss with a weight. Then we can for example create a weighted sum to get the final loss for a batch. Let’s have a look at a specific example: we want to setup a language model that helps us autocomplete complete common data science code. For that task we would like to weight samples stronger where tokens related to the data science stack, such as pd or np, occur more frequently. Here you see a loss function that does exactly that for causal language modeling. It takes the models it takes the model’s inputs and predicted logits as well as the key tokens as input. First the inputs and logits are aligned, then the loss per sample is calculate followed by the weights. Finally the loss and weights are combined and returned. This is a pretty big function so let’s take a closer look at the loss and weight blocks. During the calculation of the standard loss the logits and labels are flattened over the batch. With the view we unflatten the tensor to get a matrix with a row for each sample in the batch and a column for each position in the sequence of the samples. We don’t need the loss per position so we average the loss over all positions for each sample. For the weights we use boolean logic to get a tensor with 1s where a keyword occurred and 0s where not. This tensor has an additional dimension as the loss tensor we just saw because we get the information for each keyword in a separate matrix. Only want to know how many times keywords occurred per sample so we can sum over all keywords and all positions per sample. Now we are almost there, we only need to combine the loss with the weight per sample. We do this with element wise multiplication and then average over all samples in the batch. In the end we have exactly one loss value for the whole batch. And this is the whole necessary logic to create a custom weighted loss. Let’s see how we can make use of that custom loss with Accelerate and the Trainer In Accelerate we just pass the input_ids to the models to get the logits and can then call the custom loss function. After that we continue with the normal training loop by for example calling backward. For the Trainer we can overwrite the compute loss function of the standard trainer. We just need to make sure that that we return the loss and the model outputs in the same format. With that you can integrate your own awesome loss function with both the trainer and accelerates. | huggingface/course/blob/main/subtitles/en/raw/chapter7/06b_custom-loss.md |
--
title: "Zero-shot image-to-text generation with BLIP-2"
thumbnail: /blog/assets/blip-2/thumbnail.png
authors:
- user: MariaK
- user: JunnanLi
---
# Zero-shot image-to-text generation with BLIP-2
This guide introduces [BLIP-2](https://huggingface.co/docs/transformers/main/en/model_doc/blip-2) from Salesforce Research
that enables a suite of state-of-the-art visual-language models that are now available in [🤗 Transformers](https://huggingface.co/transformers).
We'll show you how to use it for image captioning, prompted image captioning, visual question-answering, and chat-based prompting.
## Table of contents
1. [Introduction](#introduction)
2. [What's under the hood in BLIP-2?](#whats-under-the-hood-in-blip-2)
3. [Using BLIP-2 with Hugging Face Transformers](#using-blip-2-with-hugging-face-transformers)
1. [Image Captioning](#image-captioning)
2. [Prompted image captioning](#prompted-image-captioning)
3. [Visual question answering](#visual-question-answering)
4. [Chat-based prompting](#chat-based-prompting)
4. [Conclusion](#conclusion)
5. [Acknowledgments](#acknowledgments)
## Introduction
Recent years have seen rapid advancements in computer vision and natural language processing. Still, many real-world
problems are inherently multimodal - they involve several distinct forms of data, such as images and text.
Visual-language models face the challenge of combining modalities so that they can open the door to a wide range of
applications. Some of the image-to-text tasks that visual language models can tackle include image captioning, image-text
retrieval, and visual question answering. Image captioning can aid the visually impaired, create useful product descriptions,
identify inappropriate content beyond text, and more. Image-text retrieval can be applied in multimodal search, as well
as in applications such as autonomous driving. Visual question-answering can aid in education, enable multimodal chatbots,
and assist in various domain-specific information retrieval applications.
Modern computer vision and natural language models have become more capable; however, they have also significantly
grown in size compared to their predecessors. While pre-training a single-modality model is resource-consuming and expensive,
the cost of end-to-end vision-and-language pre-training has become increasingly prohibitive.
[BLIP-2](https://arxiv.org/pdf/2301.12597.pdf) tackles this challenge by introducing a new visual-language pre-training paradigm that can potentially leverage
any combination of pre-trained vision encoder and LLM without having to pre-train the whole architecture end to end.
This enables achieving state-of-the-art results on multiple visual-language tasks while significantly reducing the number
of trainable parameters and pre-training costs. Moreover, this approach paves the way for a multimodal ChatGPT-like model.
## What's under the hood in BLIP-2?
BLIP-2 bridges the modality gap between vision and language models by adding a lightweight Querying Transformer (Q-Former)
between an off-the-shelf frozen pre-trained image encoder and a frozen large language model. Q-Former is the only
trainable part of BLIP-2; both the image encoder and language model remain frozen.
<p align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/blip-2/q-former-1.png" alt="Overview of BLIP-2's framework" width=500>
</p>
Q-Former is a transformer model that consists of two submodules that share the same self-attention layers:
* an image transformer that interacts with the frozen image encoder for visual feature extraction
* a text transformer that can function as both a text encoder and a text decoder
<p align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/blip-2/q-former-2.png" alt="Q-Former architecture" width=500>
</p>
The image transformer extracts a fixed number of output features from the image encoder, independent of input image resolution,
and receives learnable query embeddings as input. The queries can additionally interact with the text through the same self-attention layers.
Q-Former is pre-trained in two stages. In the first stage, the image encoder is frozen, and Q-Former is trained with three losses:
* Image-text contrastive loss: pairwise similarity between each query output and text output's CLS token is calculated, and the highest one is picked. Query embeddings and text don't “see” each other.
* Image-grounded text generation: queries can attend to each other but not to the text tokens, and text has a causal mask and can attend to all of the queries.
* Image-text matching loss: queries and text can see others, and a logit is obtained to indicate whether the text matches the image or not. To obtain negative examples, hard negative mining is used.
In the second pre-training stage, the query embeddings now have the relevant visual information to the text as it has
passed through an information bottleneck. These embeddings are now used as a visual prefix to the input to the LLM. This
pre-training phase effectively involves an image-ground text generation task using the causal LM loss.
As a visual encoder, BLIP-2 uses ViT, and for an LLM, the paper authors used OPT and Flan T5 models. You can find
pre-trained checkpoints for both OPT and Flan T5 on [Hugging Face Hub](https://huggingface.co/models?other=blip-2).
However, as mentioned before, the introduced pre-training approach allows combining any visual backbone with any LLM.
## Using BLIP-2 with Hugging Face Transformers
Using Hugging Face Transformers, you can easily download and run a pre-trained BLIP-2 model on your images. Make sure to use a GPU environment with high RAM if you'd like to follow along with the examples in this blog post.
Let's start by installing Transformers. As this model has been added to Transformers very recently, we need to install Transformers from the source:
```bash
pip install git+https://github.com/huggingface/transformers.git
```
Next, we'll need an input image. Every week The New Yorker runs a [cartoon captioning contest](https://www.newyorker.com/cartoons/contest#thisweek)
among its readers, so let's take one of these cartoons to put BLIP-2 to the test.
```
import requests
from PIL import Image
url = 'https://media.newyorker.com/cartoons/63dc6847be24a6a76d90eb99/master/w_1160,c_limit/230213_a26611_838.jpg'
image = Image.open(requests.get(url, stream=True).raw).convert('RGB')
display(image.resize((596, 437)))
```
<p align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/blip-2/cartoon.jpeg" alt="New Yorker Cartoon" width=500>
</p>
We have an input image. Now we need a pre-trained BLIP-2 model and corresponding preprocessor to prepare the inputs. You
can find the list of all available pre-trained checkpoints on [Hugging Face Hub](https://huggingface.co/models?other=blip-2).
Here, we'll load a BLIP-2 checkpoint that leverages the pre-trained OPT model by Meta AI, which has 2.7 billion parameters.
```
from transformers import AutoProcessor, Blip2ForConditionalGeneration
import torch
processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b")
model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16)
```
Notice that BLIP-2 is a rare case where you cannot load the model with Auto API (e.g. AutoModelForXXX), and you need to
explicitly use `Blip2ForConditionalGeneration`. However, you can use `AutoProcessor` to fetch the appropriate processor
class - `Blip2Processor` in this case.
Let's use GPU to make text generation faster:
```
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
```
### Image Captioning
Let's find out if BLIP-2 can caption a New Yorker cartoon in a zero-shot manner. To caption an image, we do not have to
provide any text prompt to the model, only the preprocessed input image. Without any text prompt, the model will start
generating text from the BOS (beginning-of-sequence) token thus creating a caption.
```
inputs = processor(image, return_tensors="pt").to(device, torch.float16)
generated_ids = model.generate(**inputs, max_new_tokens=20)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
print(generated_text)
```
```
"two cartoon monsters sitting around a campfire"
```
This is an impressively accurate description for a model that wasn't trained on New Yorker style cartoons!
### Prompted image captioning
We can extend image captioning by providing a text prompt, which the model will continue given the image.
```
prompt = "this is a cartoon of"
inputs = processor(image, text=prompt, return_tensors="pt").to(device, torch.float16)
generated_ids = model.generate(**inputs, max_new_tokens=20)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
print(generated_text)
```
```
"two monsters sitting around a campfire"
```
```
prompt = "they look like they are"
inputs = processor(image, text=prompt, return_tensors="pt").to(device, torch.float16)
generated_ids = model.generate(**inputs, max_new_tokens=20)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
print(generated_text)
```
```
"having a good time"
```
### Visual question answering
For visual question answering the prompt has to follow a specific format:
"Question: {} Answer:"
```
prompt = "Question: What is a dinosaur holding? Answer:"
inputs = processor(image, text=prompt, return_tensors="pt").to(device, torch.float16)
generated_ids = model.generate(**inputs, max_new_tokens=10)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
print(generated_text)
```
```
"A torch"
```
### Chat-based prompting
Finally, we can create a ChatGPT-like interface by concatenating each generated response to the conversation. We prompt
the model with some text (like "What is a dinosaur holding?"), the model generates an answer for it "a torch"), which we
can concatenate to the conversation. Then we do it again, building up the context.
However, make sure that the context does not exceed 512 tokens, as this is the context length of the language models used by BLIP-2 (OPT and T5).
```
context = [
("What is a dinosaur holding?", "a torch"),
("Where are they?", "In the woods.")
]
question = "What for?"
template = "Question: {} Answer: {}."
prompt = " ".join([template.format(context[i][0], context[i][1]) for i in range(len(context))]) + " Question: " + question + " Answer:"
print(prompt)
```
```
Question: What is a dinosaur holding? Answer: a torch. Question: Where are they? Answer: In the woods.. Question: What for? Answer:
```
```
inputs = processor(image, text=prompt, return_tensors="pt").to(device, torch.float16)
generated_ids = model.generate(**inputs, max_new_tokens=10)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
print(generated_text)
```
```
To light a fire.
```
## Conclusion
BLIP-2 is a zero-shot visual-language model that can be used for multiple image-to-text tasks with image and image and
text prompts. It is an effective and efficient approach that can be applied to image understanding in numerous scenarios,
especially when examples are scarce.
The model bridges the gap between vision and natural language modalities by adding a transformer between pre-trained models.
The new pre-training paradigm allows this model to keep up with the advances in both individual modalities.
If you'd like to learn how to fine-tune BLIP-2 models for various vision-language tasks, check out [LAVIS library by Salesforce](https://github.com/salesforce/LAVIS)
that offers comprehensive support for model training.
To see BLIP-2 in action, try its demo on [Hugging Face Spaces](https://huggingface.co/spaces/Salesforce/BLIP2).
## Acknowledgments
Many thanks to the Salesforce Research team for working on BLIP-2, Niels Rogge for adding BLIP-2 to 🤗 Transformers, and
to Omar Sanseviero for reviewing this blog post.
| huggingface/blog/blob/main/blip-2.md |
Using `Transformers.js` at Hugging Face
Transformers.js is a JavaScript library for running 🤗 Transformers directly in your browser, with no need for a server! It is designed to be functionally equivalent to the original [Python library](https://github.com/huggingface/transformers), meaning you can run the same pretrained models using a very similar API.
## Exploring `transformers.js` in the Hub
You can find `transformers.js` models by filtering by library in the [models page](https://huggingface.co/models?library=transformers.js).
## Quick tour
It's super simple to translate from existing code! Just like the Python library, we support the `pipeline` API. Pipelines group together a pretrained model with preprocessing of inputs and postprocessing of outputs, making it the easiest way to run models with the library.
<table>
<tr>
<th width="440px" align="center"><b>Python (original)</b></th>
<th width="440px" align="center"><b>Javascript (ours)</b></th>
</tr>
<tr>
<td>
```python
from transformers import pipeline
# Allocate a pipeline for sentiment-analysis
pipe = pipeline('sentiment-analysis')
out = pipe('I love transformers!')
# [{'label': 'POSITIVE', 'score': 0.999806941}]
```
</td>
<td>
```javascript
import { pipeline } from '@xenova/transformers';
// Allocate a pipeline for sentiment-analysis
let pipe = await pipeline('sentiment-analysis');
let out = await pipe('I love transformers!');
// [{'label': 'POSITIVE', 'score': 0.999817686}]
```
</td>
</tr>
</table>
You can also use a different model by specifying the model id or path as the second argument to the `pipeline` function. For example:
```javascript
// Use a different model for sentiment-analysis
let pipe = await pipeline('sentiment-analysis', 'nlptown/bert-base-multilingual-uncased-sentiment');
```
Refer to the [documentation](https://huggingface.co/docs/transformers.js) for the full list of supported tasks and models.
## Installation
To install via [NPM](https://www.npmjs.com/package/@xenova/transformers), run:
```bash
npm i @xenova/transformers
```
For more information, including how to use it in vanilla JS (without any bundler) via a CDN or static hosting, refer to the [README](https://github.com/xenova/transformers.js/blob/main/README.md#installation).
## Additional resources
* Transformers.js [repository](https://github.com/xenova/transformers.js)
* Transformers.js [docs](https://huggingface.co/docs/transformers.js)
* Transformers.js [demo](https://xenova.github.io/transformers.js/)
| huggingface/hub-docs/blob/main/docs/hub/transformers-js.md |
Time to slice and dice[[time-to-slice-and-dice]]
<CourseFloatingBanner chapter={5}
classNames="absolute z-10 right-0 top-0"
notebooks={[
{label: "Google Colab", value: "https://colab.research.google.com/github/huggingface/notebooks/blob/master/course/en/chapter5/section3.ipynb"},
{label: "Aws Studio", value: "https://studiolab.sagemaker.aws/import/github/huggingface/notebooks/blob/master/course/en/chapter5/section3.ipynb"},
]} />
Most of the time, the data you work with won't be perfectly prepared for training models. In this section we'll explore the various features that 🤗 Datasets provides to clean up your datasets.
<Youtube id="tqfSFcPMgOI"/>
## Slicing and dicing our data[[slicing-and-dicing-our-data]]
Similar to Pandas, 🤗 Datasets provides several functions to manipulate the contents of `Dataset` and `DatasetDict` objects. We already encountered the `Dataset.map()` method in [Chapter 3](/course/chapter3), and in this section we'll explore some of the other functions at our disposal.
For this example we'll use the [Drug Review Dataset](https://archive.ics.uci.edu/ml/datasets/Drug+Review+Dataset+%28Drugs.com%29) that's hosted on the [UC Irvine Machine Learning Repository](https://archive.ics.uci.edu/ml/index.php), which contains patient reviews on various drugs, along with the condition being treated and a 10-star rating of the patient's satisfaction.
First we need to download and extract the data, which can be done with the `wget` and `unzip` commands:
```py
!wget "https://archive.ics.uci.edu/ml/machine-learning-databases/00462/drugsCom_raw.zip"
!unzip drugsCom_raw.zip
```
Since TSV is just a variant of CSV that uses tabs instead of commas as the separator, we can load these files by using the `csv` loading script and specifying the `delimiter` argument in the `load_dataset()` function as follows:
```py
from datasets import load_dataset
data_files = {"train": "drugsComTrain_raw.tsv", "test": "drugsComTest_raw.tsv"}
# \t is the tab character in Python
drug_dataset = load_dataset("csv", data_files=data_files, delimiter="\t")
```
A good practice when doing any sort of data analysis is to grab a small random sample to get a quick feel for the type of data you're working with. In 🤗 Datasets, we can create a random sample by chaining the `Dataset.shuffle()` and `Dataset.select()` functions together:
```py
drug_sample = drug_dataset["train"].shuffle(seed=42).select(range(1000))
# Peek at the first few examples
drug_sample[:3]
```
```python out
{'Unnamed: 0': [87571, 178045, 80482],
'drugName': ['Naproxen', 'Duloxetine', 'Mobic'],
'condition': ['Gout, Acute', 'ibromyalgia', 'Inflammatory Conditions'],
'review': ['"like the previous person mention, I'm a strong believer of aleve, it works faster for my gout than the prescription meds I take. No more going to the doctor for refills.....Aleve works!"',
'"I have taken Cymbalta for about a year and a half for fibromyalgia pain. It is great\r\nas a pain reducer and an anti-depressant, however, the side effects outweighed \r\nany benefit I got from it. I had trouble with restlessness, being tired constantly,\r\ndizziness, dry mouth, numbness and tingling in my feet, and horrible sweating. I am\r\nbeing weaned off of it now. Went from 60 mg to 30mg and now to 15 mg. I will be\r\noff completely in about a week. The fibro pain is coming back, but I would rather deal with it than the side effects."',
'"I have been taking Mobic for over a year with no side effects other than an elevated blood pressure. I had severe knee and ankle pain which completely went away after taking Mobic. I attempted to stop the medication however pain returned after a few days."'],
'rating': [9.0, 3.0, 10.0],
'date': ['September 2, 2015', 'November 7, 2011', 'June 5, 2013'],
'usefulCount': [36, 13, 128]}
```
Note that we've fixed the seed in `Dataset.shuffle()` for reproducibility purposes. `Dataset.select()` expects an iterable of indices, so we've passed `range(1000)` to grab the first 1,000 examples from the shuffled dataset. From this sample we can already see a few quirks in our dataset:
* The `Unnamed: 0` column looks suspiciously like an anonymized ID for each patient.
* The `condition` column includes a mix of uppercase and lowercase labels.
* The reviews are of varying length and contain a mix of Python line separators (`\r\n`) as well as HTML character codes like `&\#039;`.
Let's see how we can use 🤗 Datasets to deal with each of these issues. To test the patient ID hypothesis for the `Unnamed: 0` column, we can use the `Dataset.unique()` function to verify that the number of IDs matches the number of rows in each split:
```py
for split in drug_dataset.keys():
assert len(drug_dataset[split]) == len(drug_dataset[split].unique("Unnamed: 0"))
```
This seems to confirm our hypothesis, so let's clean up the dataset a bit by renaming the `Unnamed: 0` column to something a bit more interpretable. We can use the `DatasetDict.rename_column()` function to rename the column across both splits in one go:
```py
drug_dataset = drug_dataset.rename_column(
original_column_name="Unnamed: 0", new_column_name="patient_id"
)
drug_dataset
```
```python out
DatasetDict({
train: Dataset({
features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount'],
num_rows: 161297
})
test: Dataset({
features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount'],
num_rows: 53766
})
})
```
<Tip>
✏️ **Try it out!** Use the `Dataset.unique()` function to find the number of unique drugs and conditions in the training and test sets.
</Tip>
Next, let's normalize all the `condition` labels using `Dataset.map()`. As we did with tokenization in [Chapter 3](/course/chapter3), we can define a simple function that can be applied across all the rows of each split in `drug_dataset`:
```py
def lowercase_condition(example):
return {"condition": example["condition"].lower()}
drug_dataset.map(lowercase_condition)
```
```python out
AttributeError: 'NoneType' object has no attribute 'lower'
```
Oh no, we've run into a problem with our map function! From the error we can infer that some of the entries in the `condition` column are `None`, which cannot be lowercased as they're not strings. Let's drop these rows using `Dataset.filter()`, which works in a similar way to `Dataset.map()` and expects a function that receives a single example of the dataset. Instead of writing an explicit function like:
```py
def filter_nones(x):
return x["condition"] is not None
```
and then running `drug_dataset.filter(filter_nones)`, we can do this in one line using a _lambda function_. In Python, lambda functions are small functions that you can define without explicitly naming them. They take the general form:
```
lambda <arguments> : <expression>
```
where `lambda` is one of Python's special [keywords](https://docs.python.org/3/reference/lexical_analysis.html#keywords), `<arguments>` is a list/set of comma-separated values that define the inputs to the function, and `<expression>` represents the operations you wish to execute. For example, we can define a simple lambda function that squares a number as follows:
```
lambda x : x * x
```
To apply this function to an input, we need to wrap it and the input in parentheses:
```py
(lambda x: x * x)(3)
```
```python out
9
```
Similarly, we can define lambda functions with multiple arguments by separating them with commas. For example, we can compute the area of a triangle as follows:
```py
(lambda base, height: 0.5 * base * height)(4, 8)
```
```python out
16.0
```
Lambda functions are handy when you want to define small, single-use functions (for more information about them, we recommend reading the excellent [Real Python tutorial](https://realpython.com/python-lambda/) by Andre Burgaud). In the 🤗 Datasets context, we can use lambda functions to define simple map and filter operations, so let's use this trick to eliminate the `None` entries in our dataset:
```py
drug_dataset = drug_dataset.filter(lambda x: x["condition"] is not None)
```
With the `None` entries removed, we can normalize our `condition` column:
```py
drug_dataset = drug_dataset.map(lowercase_condition)
# Check that lowercasing worked
drug_dataset["train"]["condition"][:3]
```
```python out
['left ventricular dysfunction', 'adhd', 'birth control']
```
It works! Now that we've cleaned up the labels, let's take a look at cleaning up the reviews themselves.
## Creating new columns[[creating-new-columns]]
Whenever you're dealing with customer reviews, a good practice is to check the number of words in each review. A review might be just a single word like "Great!" or a full-blown essay with thousands of words, and depending on the use case you'll need to handle these extremes differently. To compute the number of words in each review, we'll use a rough heuristic based on splitting each text by whitespace.
Let's define a simple function that counts the number of words in each review:
```py
def compute_review_length(example):
return {"review_length": len(example["review"].split())}
```
Unlike our `lowercase_condition()` function, `compute_review_length()` returns a dictionary whose key does not correspond to one of the column names in the dataset. In this case, when `compute_review_length()` is passed to `Dataset.map()`, it will be applied to all the rows in the dataset to create a new `review_length` column:
```py
drug_dataset = drug_dataset.map(compute_review_length)
# Inspect the first training example
drug_dataset["train"][0]
```
```python out
{'patient_id': 206461,
'drugName': 'Valsartan',
'condition': 'left ventricular dysfunction',
'review': '"It has no side effect, I take it in combination of Bystolic 5 Mg and Fish Oil"',
'rating': 9.0,
'date': 'May 20, 2012',
'usefulCount': 27,
'review_length': 17}
```
As expected, we can see a `review_length` column has been added to our training set. We can sort this new column with `Dataset.sort()` to see what the extreme values look like:
```py
drug_dataset["train"].sort("review_length")[:3]
```
```python out
{'patient_id': [103488, 23627, 20558],
'drugName': ['Loestrin 21 1 / 20', 'Chlorzoxazone', 'Nucynta'],
'condition': ['birth control', 'muscle spasm', 'pain'],
'review': ['"Excellent."', '"useless"', '"ok"'],
'rating': [10.0, 1.0, 6.0],
'date': ['November 4, 2008', 'March 24, 2017', 'August 20, 2016'],
'usefulCount': [5, 2, 10],
'review_length': [1, 1, 1]}
```
As we suspected, some reviews contain just a single word, which, although it may be okay for sentiment analysis, would not be informative if we want to predict the condition.
<Tip>
🙋 An alternative way to add new columns to a dataset is with the `Dataset.add_column()` function. This allows you to provide the column as a Python list or NumPy array and can be handy in situations where `Dataset.map()` is not well suited for your analysis.
</Tip>
Let's use the `Dataset.filter()` function to remove reviews that contain fewer than 30 words. Similarly to what we did with the `condition` column, we can filter out the very short reviews by requiring that the reviews have a length above this threshold:
```py
drug_dataset = drug_dataset.filter(lambda x: x["review_length"] > 30)
print(drug_dataset.num_rows)
```
```python out
{'train': 138514, 'test': 46108}
```
As you can see, this has removed around 15% of the reviews from our original training and test sets.
<Tip>
✏️ **Try it out!** Use the `Dataset.sort()` function to inspect the reviews with the largest numbers of words. See the [documentation](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.sort) to see which argument you need to use sort the reviews by length in descending order.
</Tip>
The last thing we need to deal with is the presence of HTML character codes in our reviews. We can use Python's `html` module to unescape these characters, like so:
```py
import html
text = "I'm a transformer called BERT"
html.unescape(text)
```
```python out
"I'm a transformer called BERT"
```
We'll use `Dataset.map()` to unescape all the HTML characters in our corpus:
```python
drug_dataset = drug_dataset.map(lambda x: {"review": html.unescape(x["review"])})
```
As you can see, the `Dataset.map()` method is quite useful for processing data -- and we haven't even scratched the surface of everything it can do!
## The `map()` method's superpowers[[the-map-methods-superpowers]]
The `Dataset.map()` method takes a `batched` argument that, if set to `True`, causes it to send a batch of examples to the map function at once (the batch size is configurable but defaults to 1,000). For instance, the previous map function that unescaped all the HTML took a bit of time to run (you can read the time taken from the progress bars). We can speed this up by processing several elements at the same time using a list comprehension.
When you specify `batched=True` the function receives a dictionary with the fields of the dataset, but each value is now a _list of values_, and not just a single value. The return value of `Dataset.map()` should be the same: a dictionary with the fields we want to update or add to our dataset, and a list of values. For example, here is another way to unescape all HTML characters, but using `batched=True`:
```python
new_drug_dataset = drug_dataset.map(
lambda x: {"review": [html.unescape(o) for o in x["review"]]}, batched=True
)
```
If you're running this code in a notebook, you'll see that this command executes way faster than the previous one. And it's not because our reviews have already been HTML-unescaped -- if you re-execute the instruction from the previous section (without `batched=True`), it will take the same amount of time as before. This is because list comprehensions are usually faster than executing the same code in a `for` loop, and we also gain some performance by accessing lots of elements at the same time instead of one by one.
Using `Dataset.map()` with `batched=True` will be essential to unlock the speed of the "fast" tokenizers that we'll encounter in [Chapter 6](/course/chapter6), which can quickly tokenize big lists of texts. For instance, to tokenize all the drug reviews with a fast tokenizer, we could use a function like this:
```python
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
def tokenize_function(examples):
return tokenizer(examples["review"], truncation=True)
```
As you saw in [Chapter 3](/course/chapter3), we can pass one or several examples to the tokenizer, so we can use this function with or without `batched=True`. Let's take this opportunity to compare the performance of the different options. In a notebook, you can time a one-line instruction by adding `%time` before the line of code you wish to measure:
```python no-format
%time tokenized_dataset = drug_dataset.map(tokenize_function, batched=True)
```
You can also time a whole cell by putting `%%time` at the beginning of the cell. On the hardware we executed this on, it showed 10.8s for this instruction (it's the number written after "Wall time").
<Tip>
✏️ **Try it out!** Execute the same instruction with and without `batched=True`, then try it with a slow tokenizer (add `use_fast=False` in the `AutoTokenizer.from_pretrained()` method) so you can see what numbers you get on your hardware.
</Tip>
Here are the results we obtained with and without batching, with a fast and a slow tokenizer:
Options | Fast tokenizer | Slow tokenizer
:--------------:|:--------------:|:-------------:
`batched=True` | 10.8s | 4min41s
`batched=False` | 59.2s | 5min3s
This means that using a fast tokenizer with the `batched=True` option is 30 times faster than its slow counterpart with no batching -- this is truly amazing! That's the main reason why fast tokenizers are the default when using `AutoTokenizer` (and why they are called "fast"). They're able to achieve such a speedup because behind the scenes the tokenization code is executed in Rust, which is a language that makes it easy to parallelize code execution.
Parallelization is also the reason for the nearly 6x speedup the fast tokenizer achieves with batching: you can't parallelize a single tokenization operation, but when you want to tokenize lots of texts at the same time you can just split the execution across several processes, each responsible for its own texts.
`Dataset.map()` also has some parallelization capabilities of its own. Since they are not backed by Rust, they won't let a slow tokenizer catch up with a fast one, but they can still be helpful (especially if you're using a tokenizer that doesn't have a fast version). To enable multiprocessing, use the `num_proc` argument and specify the number of processes to use in your call to `Dataset.map()`:
```py
slow_tokenizer = AutoTokenizer.from_pretrained("bert-base-cased", use_fast=False)
def slow_tokenize_function(examples):
return slow_tokenizer(examples["review"], truncation=True)
tokenized_dataset = drug_dataset.map(slow_tokenize_function, batched=True, num_proc=8)
```
You can experiment a little with timing to determine the optimal number of processes to use; in our case 8 seemed to produce the best speed gain. Here are the numbers we got with and without multiprocessing:
Options | Fast tokenizer | Slow tokenizer
:--------------:|:--------------:|:-------------:
`batched=True` | 10.8s | 4min41s
`batched=False` | 59.2s | 5min3s
`batched=True`, `num_proc=8` | 6.52s | 41.3s
`batched=False`, `num_proc=8` | 9.49s | 45.2s
Those are much more reasonable results for the slow tokenizer, but the performance of the fast tokenizer was also substantially improved. Note, however, that won't always be the case -- for values of `num_proc` other than 8, our tests showed that it was faster to use `batched=True` without that option. In general, we don't recommend using Python multiprocessing for fast tokenizers with `batched=True`.
<Tip>
Using `num_proc` to speed up your processing is usually a great idea, as long as the function you are using is not already doing some kind of multiprocessing of its own.
</Tip>
All of this functionality condensed into a single method is already pretty amazing, but there's more! With `Dataset.map()` and `batched=True` you can change the number of elements in your dataset. This is super useful in many situations where you want to create several training features from one example, and we will need to do this as part of the preprocessing for several of the NLP tasks we'll undertake in [Chapter 7](/course/chapter7).
<Tip>
💡 In machine learning, an _example_ is usually defined as the set of _features_ that we feed to the model. In some contexts, these features will be the set of columns in a `Dataset`, but in others (like here and for question answering), multiple features can be extracted from a single example and belong to a single column.
</Tip>
Let's have a look at how it works! Here we will tokenize our examples and truncate them to a maximum length of 128, but we will ask the tokenizer to return *all* the chunks of the texts instead of just the first one. This can be done with `return_overflowing_tokens=True`:
```py
def tokenize_and_split(examples):
return tokenizer(
examples["review"],
truncation=True,
max_length=128,
return_overflowing_tokens=True,
)
```
Let's test this on one example before using `Dataset.map()` on the whole dataset:
```py
result = tokenize_and_split(drug_dataset["train"][0])
[len(inp) for inp in result["input_ids"]]
```
```python out
[128, 49]
```
So, our first example in the training set became two features because it was tokenized to more than the maximum number of tokens we specified: the first one of length 128 and the second one of length 49. Now let's do this for all elements of the dataset!
```py
tokenized_dataset = drug_dataset.map(tokenize_and_split, batched=True)
```
```python out
ArrowInvalid: Column 1 named condition expected length 1463 but got length 1000
```
Oh no! That didn't work! Why not? Looking at the error message will give us a clue: there is a mismatch in the lengths of one of the columns, one being of length 1,463 and the other of length 1,000. If you've looked at the `Dataset.map()` [documentation](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map), you may recall that it's the number of samples passed to the function that we are mapping; here those 1,000 examples gave 1,463 new features, resulting in a shape error.
The problem is that we're trying to mix two different datasets of different sizes: the `drug_dataset` columns will have a certain number of examples (the 1,000 in our error), but the `tokenized_dataset` we are building will have more (the 1,463 in the error message; it is more than 1,000 because we are tokenizing long reviews into more than one example by using `return_overflowing_tokens=True`). That doesn't work for a `Dataset`, so we need to either remove the columns from the old dataset or make them the same size as they are in the new dataset. We can do the former with the `remove_columns` argument:
```py
tokenized_dataset = drug_dataset.map(
tokenize_and_split, batched=True, remove_columns=drug_dataset["train"].column_names
)
```
Now this works without error. We can check that our new dataset has many more elements than the original dataset by comparing the lengths:
```py
len(tokenized_dataset["train"]), len(drug_dataset["train"])
```
```python out
(206772, 138514)
```
We mentioned that we can also deal with the mismatched length problem by making the old columns the same size as the new ones. To do this, we will need the `overflow_to_sample_mapping` field the tokenizer returns when we set `return_overflowing_tokens=True`. It gives us a mapping from a new feature index to the index of the sample it originated from. Using this, we can associate each key present in our original dataset with a list of values of the right size by repeating the values of each example as many times as it generates new features:
```py
def tokenize_and_split(examples):
result = tokenizer(
examples["review"],
truncation=True,
max_length=128,
return_overflowing_tokens=True,
)
# Extract mapping between new and old indices
sample_map = result.pop("overflow_to_sample_mapping")
for key, values in examples.items():
result[key] = [values[i] for i in sample_map]
return result
```
We can see it works with `Dataset.map()` without us needing to remove the old columns:
```py
tokenized_dataset = drug_dataset.map(tokenize_and_split, batched=True)
tokenized_dataset
```
```python out
DatasetDict({
train: Dataset({
features: ['attention_mask', 'condition', 'date', 'drugName', 'input_ids', 'patient_id', 'rating', 'review', 'review_length', 'token_type_ids', 'usefulCount'],
num_rows: 206772
})
test: Dataset({
features: ['attention_mask', 'condition', 'date', 'drugName', 'input_ids', 'patient_id', 'rating', 'review', 'review_length', 'token_type_ids', 'usefulCount'],
num_rows: 68876
})
})
```
We get the same number of training features as before, but here we've kept all the old fields. If you need them for some post-processing after applying your model, you might want to use this approach.
You've now seen how 🤗 Datasets can be used to preprocess a dataset in various ways. Although the processing functions of 🤗 Datasets will cover most of your model training needs,
there may be times when you'll need to switch to Pandas to access more powerful features, like `DataFrame.groupby()` or high-level APIs for visualization. Fortunately, 🤗 Datasets is designed to be interoperable with libraries such as Pandas, NumPy, PyTorch, TensorFlow, and JAX. Let's take a look at how this works.
## From `Dataset`s to `DataFrame`s and back[[from-datasets-to-dataframes-and-back]]
<Youtube id="tfcY1067A5Q"/>
To enable the conversion between various third-party libraries, 🤗 Datasets provides a `Dataset.set_format()` function. This function only changes the _output format_ of the dataset, so you can easily switch to another format without affecting the underlying _data format_, which is Apache Arrow. The formatting is done in place. To demonstrate, let's convert our dataset to Pandas:
```py
drug_dataset.set_format("pandas")
```
Now when we access elements of the dataset we get a `pandas.DataFrame` instead of a dictionary:
```py
drug_dataset["train"][:3]
```
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>patient_id</th>
<th>drugName</th>
<th>condition</th>
<th>review</th>
<th>rating</th>
<th>date</th>
<th>usefulCount</th>
<th>review_length</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>95260</td>
<td>Guanfacine</td>
<td>adhd</td>
<td>"My son is halfway through his fourth week of Intuniv..."</td>
<td>8.0</td>
<td>April 27, 2010</td>
<td>192</td>
<td>141</td>
</tr>
<tr>
<th>1</th>
<td>92703</td>
<td>Lybrel</td>
<td>birth control</td>
<td>"I used to take another oral contraceptive, which had 21 pill cycle, and was very happy- very light periods, max 5 days, no other side effects..."</td>
<td>5.0</td>
<td>December 14, 2009</td>
<td>17</td>
<td>134</td>
</tr>
<tr>
<th>2</th>
<td>138000</td>
<td>Ortho Evra</td>
<td>birth control</td>
<td>"This is my first time using any form of birth control..."</td>
<td>8.0</td>
<td>November 3, 2015</td>
<td>10</td>
<td>89</td>
</tr>
</tbody>
</table>
Let's create a `pandas.DataFrame` for the whole training set by selecting all the elements of `drug_dataset["train"]`:
```py
train_df = drug_dataset["train"][:]
```
<Tip>
🚨 Under the hood, `Dataset.set_format()` changes the return format for the dataset's `__getitem__()` dunder method. This means that when we want to create a new object like `train_df` from a `Dataset` in the `"pandas"` format, we need to slice the whole dataset to obtain a `pandas.DataFrame`. You can verify for yourself that the type of `drug_dataset["train"]` is `Dataset`, irrespective of the output format.
</Tip>
From here we can use all the Pandas functionality that we want. For example, we can do fancy chaining to compute the class distribution among the `condition` entries:
```py
frequencies = (
train_df["condition"]
.value_counts()
.to_frame()
.reset_index()
.rename(columns={"index": "condition", "condition": "frequency"})
)
frequencies.head()
```
<table border="1" class="dataframe">
<thead>
<tr style="text-align: right;">
<th></th>
<th>condition</th>
<th>frequency</th>
</tr>
</thead>
<tbody>
<tr>
<th>0</th>
<td>birth control</td>
<td>27655</td>
</tr>
<tr>
<th>1</th>
<td>depression</td>
<td>8023</td>
</tr>
<tr>
<th>2</th>
<td>acne</td>
<td>5209</td>
</tr>
<tr>
<th>3</th>
<td>anxiety</td>
<td>4991</td>
</tr>
<tr>
<th>4</th>
<td>pain</td>
<td>4744</td>
</tr>
</tbody>
</table>
And once we're done with our Pandas analysis, we can always create a new `Dataset` object by using the `Dataset.from_pandas()` function as follows:
```py
from datasets import Dataset
freq_dataset = Dataset.from_pandas(frequencies)
freq_dataset
```
```python out
Dataset({
features: ['condition', 'frequency'],
num_rows: 819
})
```
<Tip>
✏️ **Try it out!** Compute the average rating per drug and store the result in a new `Dataset`.
</Tip>
This wraps up our tour of the various preprocessing techniques available in 🤗 Datasets. To round out the section, let's create a validation set to prepare the dataset for training a classifier on. Before doing so, we'll reset the output format of `drug_dataset` from `"pandas"` to `"arrow"`:
```python
drug_dataset.reset_format()
```
## Creating a validation set[[creating-a-validation-set]]
Although we have a test set we could use for evaluation, it's a good practice to leave the test set untouched and create a separate validation set during development. Once you are happy with the performance of your models on the validation set, you can do a final sanity check on the test set. This process helps mitigate the risk that you'll overfit to the test set and deploy a model that fails on real-world data.
🤗 Datasets provides a `Dataset.train_test_split()` function that is based on the famous functionality from `scikit-learn`. Let's use it to split our training set into `train` and `validation` splits (we set the `seed` argument for reproducibility):
```py
drug_dataset_clean = drug_dataset["train"].train_test_split(train_size=0.8, seed=42)
# Rename the default "test" split to "validation"
drug_dataset_clean["validation"] = drug_dataset_clean.pop("test")
# Add the "test" set to our `DatasetDict`
drug_dataset_clean["test"] = drug_dataset["test"]
drug_dataset_clean
```
```python out
DatasetDict({
train: Dataset({
features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount', 'review_length', 'review_clean'],
num_rows: 110811
})
validation: Dataset({
features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount', 'review_length', 'review_clean'],
num_rows: 27703
})
test: Dataset({
features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount', 'review_length', 'review_clean'],
num_rows: 46108
})
})
```
Great, we've now prepared a dataset that's ready for training some models on! In [section 5](/course/chapter5/5) we'll show you how to upload datasets to the Hugging Face Hub, but for now let's cap off our analysis by looking at a few ways you can save datasets on your local machine.
## Saving a dataset[[saving-a-dataset]]
<Youtube id="blF9uxYcKHo"/>
Although 🤗 Datasets will cache every downloaded dataset and the operations performed on it, there are times when you'll want to save a dataset to disk (e.g., in case the cache gets deleted). As shown in the table below, 🤗 Datasets provides three main functions to save your dataset in different formats:
| Data format | Function |
| :---------: | :--------------------: |
| Arrow | `Dataset.save_to_disk()` |
| CSV | `Dataset.to_csv()` |
| JSON | `Dataset.to_json()` |
For example, let's save our cleaned dataset in the Arrow format:
```py
drug_dataset_clean.save_to_disk("drug-reviews")
```
This will create a directory with the following structure:
```
drug-reviews/
├── dataset_dict.json
├── test
│ ├── dataset.arrow
│ ├── dataset_info.json
│ └── state.json
├── train
│ ├── dataset.arrow
│ ├── dataset_info.json
│ ├── indices.arrow
│ └── state.json
└── validation
├── dataset.arrow
├── dataset_info.json
├── indices.arrow
└── state.json
```
where we can see that each split is associated with its own *dataset.arrow* table, and some metadata in *dataset_info.json* and *state.json*. You can think of the Arrow format as a fancy table of columns and rows that is optimized for building high-performance applications that process and transport large datasets.
Once the dataset is saved, we can load it by using the `load_from_disk()` function as follows:
```py
from datasets import load_from_disk
drug_dataset_reloaded = load_from_disk("drug-reviews")
drug_dataset_reloaded
```
```python out
DatasetDict({
train: Dataset({
features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount', 'review_length'],
num_rows: 110811
})
validation: Dataset({
features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount', 'review_length'],
num_rows: 27703
})
test: Dataset({
features: ['patient_id', 'drugName', 'condition', 'review', 'rating', 'date', 'usefulCount', 'review_length'],
num_rows: 46108
})
})
```
For the CSV and JSON formats, we have to store each split as a separate file. One way to do this is by iterating over the keys and values in the `DatasetDict` object:
```py
for split, dataset in drug_dataset_clean.items():
dataset.to_json(f"drug-reviews-{split}.jsonl")
```
This saves each split in [JSON Lines format](https://jsonlines.org), where each row in the dataset is stored as a single line of JSON. Here's what the first example looks like:
```py
!head -n 1 drug-reviews-train.jsonl
```
```python out
{"patient_id":141780,"drugName":"Escitalopram","condition":"depression","review":"\"I seemed to experience the regular side effects of LEXAPRO, insomnia, low sex drive, sleepiness during the day. I am taking it at night because my doctor said if it made me tired to take it at night. I assumed it would and started out taking it at night. Strange dreams, some pleasant. I was diagnosed with fibromyalgia. Seems to be helping with the pain. Have had anxiety and depression in my family, and have tried quite a few other medications that haven't worked. Only have been on it for two weeks but feel more positive in my mind, want to accomplish more in my life. Hopefully the side effects will dwindle away, worth it to stick with it from hearing others responses. Great medication.\"","rating":9.0,"date":"May 29, 2011","usefulCount":10,"review_length":125}
```
We can then use the techniques from [section 2](/course/chapter5/2) to load the JSON files as follows:
```py
data_files = {
"train": "drug-reviews-train.jsonl",
"validation": "drug-reviews-validation.jsonl",
"test": "drug-reviews-test.jsonl",
}
drug_dataset_reloaded = load_dataset("json", data_files=data_files)
```
And that's it for our excursion into data wrangling with 🤗 Datasets! Now that we have a cleaned dataset for training a model on, here are a few ideas that you could try out:
1. Use the techniques from [Chapter 3](/course/chapter3) to train a classifier that can predict the patient condition based on the drug review.
2. Use the `summarization` pipeline from [Chapter 1](/course/chapter1) to generate summaries of the reviews.
Next, we'll take a look at how 🤗 Datasets can enable you to work with huge datasets without blowing up your laptop!
| huggingface/course/blob/main/chapters/en/chapter5/3.mdx |
Hub methods
Methods for using the Hugging Face Hub:
## Push to hub
[[autodoc]] evaluate.push_to_hub
| huggingface/evaluate/blob/main/docs/source/package_reference/hub_methods.mdx |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Speed up inference
There are several ways to optimize 🤗 Diffusers for inference speed. As a general rule of thumb, we recommend using either [xFormers](xformers) or `torch.nn.functional.scaled_dot_product_attention` in PyTorch 2.0 for their memory-efficient attention.
<Tip>
In many cases, optimizing for speed or memory leads to improved performance in the other, so you should try to optimize for both whenever you can. This guide focuses on inference speed, but you can learn more about preserving memory in the [Reduce memory usage](memory) guide.
</Tip>
The results below are obtained from generating a single 512x512 image from the prompt `a photo of an astronaut riding a horse on mars` with 50 DDIM steps on a Nvidia Titan RTX, demonstrating the speed-up you can expect.
| | latency | speed-up |
| ---------------- | ------- | ------- |
| original | 9.50s | x1 |
| fp16 | 3.61s | x2.63 |
| channels last | 3.30s | x2.88 |
| traced UNet | 3.21s | x2.96 |
| memory efficient attention | 2.63s | x3.61 |
## Use TensorFloat-32
On Ampere and later CUDA devices, matrix multiplications and convolutions can use the [TensorFloat-32 (TF32)](https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/) mode for faster, but slightly less accurate computations. By default, PyTorch enables TF32 mode for convolutions but not matrix multiplications. Unless your network requires full float32 precision, we recommend enabling TF32 for matrix multiplications. It can significantly speeds up computations with typically negligible loss in numerical accuracy.
```python
import torch
torch.backends.cuda.matmul.allow_tf32 = True
```
You can learn more about TF32 in the [Mixed precision training](https://huggingface.co/docs/transformers/en/perf_train_gpu_one#tf32) guide.
## Half-precision weights
To save GPU memory and get more speed, try loading and running the model weights directly in half-precision or float16:
```Python
import torch
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
use_safetensors=True,
)
pipe = pipe.to("cuda")
prompt = "a photo of an astronaut riding a horse on mars"
image = pipe(prompt).images[0]
```
<Tip warning={true}>
Don't use [`torch.autocast`](https://pytorch.org/docs/stable/amp.html#torch.autocast) in any of the pipelines as it can lead to black images and is always slower than pure float16 precision.
</Tip>
| huggingface/diffusers/blob/main/docs/source/en/optimization/fp16.md |
Using SetFit with Hugging Face
SetFit is an efficient and prompt-free framework for few-shot fine-tuning of [Sentence Transformers](https://sbert.net/). It achieves high accuracy with little labeled data - for instance, with only 8 labeled examples per class on the Customer Reviews sentiment dataset, SetFit is competitive with fine-tuning RoBERTa Large on the full training set of 3k examples 🤯!
Compared to other few-shot learning methods, SetFit has several unique features:
* 🗣 **No prompts or verbalizers:** Current techniques for few-shot fine-tuning require handcrafted prompts or verbalizers to convert examples into a format suitable for the underlying language model. SetFit dispenses with prompts altogether by generating rich embeddings directly from text examples.
* 🏎 **Fast to train:** SetFit doesn't require large-scale models like [T0](https://huggingface.co/bigscience/T0) or GPT-3 to achieve high accuracy. As a result, it is typically an order of magnitude (or more) faster to train and run inference with.
* 🌎 **Multilingual support**: SetFit can be used with any [Sentence Transformer](https://huggingface.co/models?library=sentence-transformers&sort=downloads) on the Hub, which means you can classify text in multiple languages by simply fine-tuning a multilingual checkpoint.
## Exploring SetFit on the Hub
You can find SetFit models by filtering at the left of the [models page](https://huggingface.co/models?library=setfit).
All models on the Hub come with these useful features:
1. An automatically generated model card with a brief description.
2. An interactive widget you can use to play with the model directly in the browser.
3. An Inference API that allows you to make inference requests.
## Installation
To get started, you can follow the [SetFit installation guide](https://huggingface.co/docs/setfit/installation). You can also use the following one-line install through pip:
```
pip install -U setfit
```
## Using existing models
All `setfit` models can easily be loaded from the Hub.
```py
from setfit import SetFitModel
model = SetFitModel.from_pretrained("tomaarsen/setfit-paraphrase-mpnet-base-v2-sst2-8-shot")
```
Once loaded, you can use [`SetFitModel.predict`](https://huggingface.co/docs/setfit/reference/main#setfit.SetFitModel.predict) to perform inference.
```py
model.predict("Amelia Earhart flew her single engine Lockheed Vega 5B across the Atlantic to Paris.")
```
```bash
['positive', 'negative']
```
If you want to load a specific SetFit model, you can click `Use in SetFit` and you will be given a working snippet!
## Additional resources
* [All SetFit models available on the Hub](https://huggingface.co/models?library=setfit)
* SetFit [repository](https://github.com/huggingface/setfit)
* SetFit [docs](https://huggingface.co/docs/setfit)
* SetFit [paper](https://arxiv.org/abs/2209.11055)
| huggingface/hub-docs/blob/main/docs/hub/setfit.md |
--
title: "Scaling-up BERT Inference on CPU (Part 1)"
thumbnail: /blog/assets/21_bert_cpu_scaling_part_1/imgs/numa_set.png
authors:
- user: mfuntowicz
---
<style>
.centered {
display: block;
margin: 0 auto;
}
figure {
text-align: center;
display: table;
max-width: 85%; /* demo; set some amount (px or %) if you can */
margin: 10px auto; /* not needed unless you want centered */
}
</style>
# Scaling up BERT-like model Inference on modern CPU - Part 1
## 1. Context and Motivations
Back in October 2019, my colleague Lysandre Debut published a comprehensive _(at the time)_ [inference performance
benchmarking blog (1)](https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2).
Since then, [🤗 transformers (2)](https://github.com/huggingface/transformers) welcomed a tremendous number
of new architectures and thousands of new models were added to the [🤗 hub (3)](https://huggingface.co/models)
which now counts more than 9,000 of them as of first quarter of 2021.
As the NLP landscape keeps trending towards more and more BERT-like models being used in production, it
remains challenging to efficiently deploy and run these architectures at scale.
This is why we recently introduced our [🤗 Inference API](https://api-inference.huggingface.co/docs/python/html/index.html):
to let you focus on building value for your users and customers, rather than digging into all the highly
technical aspects of running such models.
This blog post is the first part of a series which will cover most of the hardware and software optimizations to better
leverage CPUs for BERT model inference.
For this initial blog post, we will cover the hardware part:
- Setting up a baseline - Out of the box results
- Practical & technical considerations when leveraging modern CPUs for CPU-bound tasks
- Core count scaling - Does increasing the number of cores actually give better performance?
- Batch size scaling - Increasing throughput with multiple parallel & independent model instances
We decided to focus on the most famous Transformer model architecture,
[BERT (Delvin & al. 2018) (4)](https://arxiv.org/abs/1810.04805v1). While we focus this blog post on BERT-like
models to keep the article concise, all the described techniques
can be applied to any architecture on the Hugging Face model hub.
In this blog post we will not describe in detail the Transformer architecture - to learn about that I can't
recommend enough the
[Illustrated Transformer blogpost from Jay Alammar (5)](https://jalammar.github.io/illustrated-transformer/).
Today's goals are to give you an idea of where we are from an Open Source perspective using BERT-like
models for inference on PyTorch and TensorFlow, and also what you can easily leverage to speedup inference.
## 2. Benchmarking methodology
When it comes to leveraging BERT-like models from Hugging Face's model hub, there are many knobs which can
be tuned to make things faster.
Also, in order to quantify what "faster" means, we will rely on widely adopted metrics:
- **Latency**: Time it takes for a single execution of the model (i.e. forward call)
- **Throughput**: Number of executions performed in a fixed amount of time
These two metrics will help us understand the benefits and tradeoffs along this blog post.
The benchmarking methodology was reimplemented from scratch in order to integrate the latest features provided by transformers
and also to let the community run and share benchmarks in an __hopefully easier__ way.
The whole framework is now based on [Facebook AI & Research's Hydra configuration library](https://hydra.cc/) allowing us to easily report
and track all the items involved while running the benchmark, hence increasing the overall reproducibility.
You can find the whole structure of the project [here](https://github.com/huggingface/tune)
On the 2021 version, we kept the ability to run inference workloads through PyTorch and Tensorflow as in the
previous blog [(1)](https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2) along with their traced counterpart
[TorchScript (6)](https://pytorch.org/docs/stable/jit.html), [Google Accelerated Linear Algebra (XLA) (7)](https://www.tensorflow.org/xla).
Also, we decided to include support for [ONNX Runtime (8)](https://www.onnxruntime.ai/) as it provides many optimizations
specifically targeting transformers based models which makes it a strong candidate to consider when discussing
performance.
Last but not least, this new unified benchmarking environment will allow us to easily run inference for different scenarios
such as [Quantized Models (Zafrir & al.) (9)](https://arxiv.org/abs/1910.06188)
using less precise number representations (`float16`, `int8`, `int4`).
This method known as **quantization** has seen an increased adoption among all major hardware providers.
In the near future, we would like to integrate additional methods we are actively working on at Hugging Face, namely Distillation, Pruning & Sparsificaton.
## 3. Baselines
All the results below were run on [Amazon Web Services (AWS) c5.metal instance](https://aws.amazon.com/ec2/instance-types/c5)
leveraging an Intel Xeon Platinum 8275 CPU (48 cores/96 threads).
The choice of this instance provides all the useful CPU features to speedup Deep Learning workloads such as:
- AVX512 instructions set (_which might not be leveraged out-of-the-box by the various frameworks_)
- Intel Deep Learning Boost (also known as Vector Neural Network Instruction - VNNI) which provides specialized
CPU instructions for running quantized networks (_using int8 data type_)
The choice of using _metal_ instance is to avoid any virtualization issue which can arise when using cloud providers.
This gives us full control of the hardware, especially while targeting the NUMA (Non-Unified Memory Architecture) controller, which
we will cover later in this post.
_The operating system was Ubuntu 20.04 (LTS) and all the experiments were conducted using Hugging Face transformers version 4.5.0, PyTorch 1.8.1 & Google TensorFlow 2.4.0_
## 4. Out of the box results
<br>
<figure class="image">
<img alt="pytorch versus tensorflow out of the box" src="assets/21_bert_cpu_scaling_part_1/imgs/pytorch_vs_tf_oob.svg" />
<figcaption>Figure 1. PyTorch (1.8.1) vs Google TensorFlow (2.4.1) out of the box</figcaption>
</figure>
<br>
<br>
<figure class="image">
<img alt="pytorch versus tensorflow out of the box bigger batch sizes" src="assets/21_bert_cpu_scaling_part_1/imgs/pytorch_vs_tf_oob_big_batch.svg" />
<figcaption>Figure 2. PyTorch (1.8.1) vs Google TensorFlow (2.4.1) out of the box - (Bigger Batch Size)</figcaption>
</figure>
<br>
Straigh to the point, out-of-the-box, PyTorch shows better inference results over TensorFlow for all the configurations tested here.
It is important to note the results out-of-the-box might not reflect the "optimal" setup for both PyTorch and TensorFlow and thus it can look deceiving here.
One possible way to explain such difference between the two frameworks might be the underlying technology to
execute parallel sections within operators.
PyTorch internally uses [OpenMP (10)](https://www.openmp.org/) along with [Intel MKL (now oneDNN) (11)](https://software.intel.com/content/www/us/en/develop/documentation/oneapi-programming-guide/top/api-based-programming/intel-oneapi-deep-neural-network-library-onednn.html) for efficient linear algebra computations whereas TensorFlow relies on Eigen and its own threading implementation.
## 5. Scaling BERT Inference to increase overall throughput on modern CPU
### 5.1. Introduction
There are multiple ways to improve the latency and throughput for tasks such as BERT inference.
Improvements and tuning can be performed at various levels from enabling Operating System features, swapping dependent
libraries with more performant ones, carefully tuning framework properties and, last but not least,
using parallelization logic leveraging all the cores on the CPU(s).
For the remainder of this blog post we will focus on the latter, also known as **Multiple Inference Stream**.
The idea is simple: Allocate **multiple instances** of the same model and assign the execution of each instance to a
**dedicated, non-overlapping subset of the CPU cores** in order to have truly parallel instances.
### 5.2. Cores and Threads on Modern CPUs
On our way towards optimizing CPU inference for better usage of the CPU cores you might have already seen -_at least for the
past 20 years_- modern CPUs specifications report "cores" and "hardware threads" or "physical" and "logical" numbers.
These notions refer to a mechanism called **Simultaneous Multi-Threading** (SMT) or **Hyper-Threading** on Intel's platforms.
To illustrate this, imagine two tasks **A** and **B**, executing in parallel, each on its own software thread.
At some point, there is a high probability these two tasks will have to wait for some resources to be fetched from main memory, SSD, HDD
or even the network.
If the threads are scheduled on different physical cores, with no hyper-threading,
during these periods the core executing the task is in an **Idle** state waiting for the resources to arrive, and effectively doing nothing... and hence not getting fully utilized
Now, with **SMT**, the **two software threads for task A and B** can be scheduled on the same **physical core**,
such that their execution is interleaved on that physical core:
Task A and Task B will execute simultaneously on the physical core and when one task is halted, the other task can still continue execution
on the core thereby increasing the utilization of that core.
<br>
<figure class="image">
<img class="centered" alt="Intel Hyper Threading technology" src="assets/21_bert_cpu_scaling_part_1/imgs/hyper_threading_explained.png" />
<figcaption>Figure 3. Illustration of Intel Hyper Threading technology (SMT)</figcaption>
</figure>
<br>
The figure 3. above simplifies the situation by assuming single core setup. If you want some more details on how SMT works on multi-cores CPUs, please
refer to these two articles with very deep technical explanations of the behavior:
- [Intel® Hyper-Threading Technology - Technical User Guide (12)](http://www.cslab.ece.ntua.gr/courses/advcomparch/2007/material/readings/Intel%20Hyper-Threading%20Technology.pdf)
- [Introduction to Hyper-Threading Technology (13)](https://software.intel.com/content/www/us/en/develop/articles/introduction-to-hyper-threading-technology.html)
Back to our model inference workload... If you think about it, in a perfect world with a fully optimized setup, computations take the majority of time.
In this context, using the logical cores shouldn't bring us any performance benefit because both logical cores (hardware threads) compete for the core’s execution resources.
As a result, the tasks being a majority of general matrix multiplications (_[gemms (14)](https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms#Level_3)_), they are inherently CPU bounds and **does not benefits** from SMT.
### 5.3. Leveraging Multi-Socket servers and CPU affinity
Nowadays servers bring many cores, some of them even support multi-socket setups (_i.e. multiple CPUs on the motherboard_).
On Linux, the command `lscpu` reports all the specifications and topology of the CPUs present on the system:
```shell
ubuntu@some-ec2-machine:~$ lscpu
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 46 bits physical, 48 bits virtual
CPU(s): 96
On-line CPU(s) list: 0-95
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz
Stepping: 7
CPU MHz: 1200.577
CPU max MHz: 3900.0000
CPU min MHz: 1200.0000
BogoMIPS: 6000.00
Virtualization: VT-x
L1d cache: 1.5 MiB
L1i cache: 1.5 MiB
L2 cache: 48 MiB
L3 cache: 71.5 MiB
NUMA node0 CPU(s): 0-23,48-71
NUMA node1 CPU(s): 24-47,72-95
```
In our case we have a machine with **2 sockets**, each socket providing **24 physical cores** with **2 threads per cores** (SMT).
Another interesting characteristic is the notion of **NUMA** node (0, 1) which represents how cores and memory are being
mapped on the system.
Non-Uniform Memory Access (**NUMA**) is the opposite of Uniform Memory Access (**UMA**) where the whole memory pool
is accessible by all the cores through a single unified bus between sockets and the main memory.
**NUMA** on the other hand splits the memory pool and each CPU socket is responsible to address a subset of the memory,
reducing the congestion on the bus.
<br>
<figure class="image">
<img class="centered" alt="Non-Uniform Memory Access and Uniform Memory Access architectures" src="assets/21_bert_cpu_scaling_part_1/imgs/UMA_NUMA.png" />
<figcaption>Figure 5. Difference illustration of UMA and NUMA architectures <a href="https://software.intel.com/content/www/us/en/develop/articles/optimizing-applications-for-numa.html">(source (15))</a></figcaption>
</figure>
<br>
In order to fully utilize the potential of such a beefy machine, we need to ensure our model instances are correctly
dispatched across all the **physical** cores on all sockets along with enforcing memory allocation to be "NUMA-aware".
On Linux, NUMA's process configuration can be tuned through [`numactl`](https://linux.die.net/man/8/numactl) which provides an interface to bind a process to a
set of CPU cores (referred as **Thread Affinity**).
Also, it allows tuning the memory allocation policy, making sure the memory allocated for the process
is as close as possible to the cores' memory pool (referred as **Explicit Memory Allocation Directives**).
_Note: Setting both cores and memory affinities is important here. Having computations done on socket 0 and memory allocated
on socket 1 would ask the system to go over the sockets shared bus to exchange memory, thus leading to an undesired overhead._
### 5.4. Tuning Thread Affinity & Memory Allocation Policy
Now that we have all the knobs required to control the resources' allocation of our model instances we go further and see how to
effectively deploy those and see the impact on latency and throughput.
Let's go gradually to get a sense of what is the impact of each command and parameter.
First, we start by launching our inference model without any tuning, and we observe how the computations are being dispatched on CPU cores (_Left_).
```shell
python3 src/main.py model=bert-base-cased backend.name=pytorch batch_size=1 sequence_length=128
```
Then we specify the core and memory affinity through `numactl` using all the **physical** cores and only a single thread (thread 0) per core (_Right_):
```shell
numactl -C 0-47 -m 0,1 python3 src/main.py model=bert-base-cased backend.name=pytorch batch_size=1 sequence_length=128
```
<br>
<figure class="image">
<img class="centered" alt="htop CPU usage without and with numactl thread affinity set" src="assets/21_bert_cpu_scaling_part_1/imgs/numa_combined.svg" />
<figcaption>Figure 6. Linux htop command side-by-side results without & with Thread Affinity set</figcaption>
</figure>
<br>
As you can see, without any specific tuning, PyTorch and TensorFlow dispatch the work on a single socket, using all the logical cores in that socket (both threads on 24 cores).
Also, as we highlighted earlier, we do not want to leverage the **SMT** feature in our case, so we set the process' thread affinity to target only 1 hardware thread.
_Note, this is specific to this run and can vary depending on individual setups. Hence, it is recommended to check thread affinity settings for each specific use-case._
Let's take sometime from here to highlight what we did with `numactl`:
- `-C 0-47` indicates to `numactl` what is the thread affinity (cores 0 to 47).
- `-m 0,1` indicates to `numactl` to allocate memory on both CPU sockets
If you wonder why we are binding the process to cores [0...47], you need to go back to look at the output of `lscpu`.
From there you will find the section `NUMA node0` and `NUMA node1` which has the form `NUMA node<X> <logical ids>`
In our case, each socket is one NUMA node and there are 2 NUMA nodes.
Each socket or each NUMA node has 24 physical cores and 2 hardware threads per core, so 48 logical cores.
For NUMA node 0, 0-23 are hardware thread 0 and 24-47 are hardware thread 1 on the 24 physical cores in socket 0.
Likewise, for NUMA node 1, 48-71 are hardware thread 0 and 72-95 are hardware thread 1 on the 24 physical cores in socket 1.
As we are targeting just 1 thread per physical core, as explained earlier, we pick only thread 0 on each core and hence logical processors 0-47.
Since we are using both sockets, we need to also bind the memory allocations accordingly (0,1).
_Please note that using both sockets may not always give the best results, particularly for small problem sizes.
The benefit of using compute resources across both sockets might be reduced or even negated by cross-socket communication overhead._
## 6. Core count scaling - Does using more cores actually improve performance?
When thinking about possible ways to improve our model inference performances, the first rational solution might be to
throw some more resources to do the same amount of work.
Through the rest of this blog series, we will refer to this setup as **Core Count Scaling** meaning, only the number
of cores used on the system to achieve the task will vary. This is also often referred as Strong Scaling in the HPC world.
At this stage, you may wonder what is the point of allocating only a subset of the cores rather than throwing
all the horses at the task to achieve minimum latency.
Indeed, depending on the problem-size, throwing more resources to the task might give better results.
It is also possible that for small problems putting more CPU cores at work doesn't improve the final latency.
In order to illustrate this, the figure 6. below takes different problem sizes (`batch_size = 1, sequence length = {32, 128, 512}`)
and reports the latencies with respect to the number of CPU cores used for running
computations for both PyTorch and TensorFlow.
Limiting the number of resources involved in computation is done by limiting the CPU cores involved in
**intra** operations (_**intra** here means inside an operator doing computation, also known as "kernel"_).
This is achieved through the following APIs:
- PyTorch: `torch.set_num_threads(x)`
- TensorFlow: `tf.config.threading.set_intra_op_parallelism_threads(x)`
<br>
<figure class="image">
<img alt="" src="assets/21_bert_cpu_scaling_part_1/imgs/core_count_scaling.svg" />
<figcaption>Figure 7. Latency measurements</figcaption>
</figure>
<br>
As you can see, depending on the problem size, the number of threads involved in the computations has a positive impact
on the latency measurements.
For small-sized problems & medium-sized problems using only one socket would give the best performance.
For large-sized problems, the overhead of the cross-socket communication is covered by the computations cost, thus benefiting from
using all the cores available on the both sockets.
## 7. Multi-Stream Inference - Using multiple instances in parallel
If you're still reading this, you should now be in good shape to set up parallel inference workloads on CPU.
Now, we are going to highlight some possibilities offered by the powerful hardware we have, and tuning the knobs described before,
to scale our inference as linearly as possible.
In the following section we will explore another possible scaling solution **Batch Size Scaling**, but before diving into this, let's
take a look at how we can leverage Linux tools in order to assign Thread Affinity allowing effective model instance parallelism.
Instead of throwing more cores to the task as you would do in the core count scaling setup, now we will be using more model instances.
Each instance will run independently on its own subset of the hardware resources in a truly parallel fashion on a subset of the CPU cores.
### 7.1. How-to allocate multiple independent instances
Let's start simple, if we want to spawn 2 instances, one on each socket with 24 cores assigned:
```shell
numactl -C 0-23 -m 0 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=128 backend.name=pytorch backend.num_threads=24
numactl -C 24-47 -m 1 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=128 backend.name=pytorch backend.num_threads=24
```
Starting from here, each instance does not share any resource with the other, and everything is operating at maximum efficiency from a
hardware perspective.
The latency measurements are identical to what a single instance would achieve, but throughput is actually 2x higher
as the two instances operate in a truly parallel way.
We can further increase the number of instances, lowering the number of cores assigned for each instance.
Let's run 4 independent instances, each of them effectively bound to 12 CPU cores.
```shell
numactl -C 0-11 -m 0 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=128 backend.name=pytorch backend.num_threads=12
numactl -C 12-23 -m 0 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=128 backend.name=pytorch backend.num_threads=12
numactl -C 24-35 -m 1 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=128 backend.name=pytorch backend.num_threads=12
numactl -C 36-47 -m 1 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=128 backend.name=pytorch backend.num_threads=12
```
The outcomes remain the same, our 4 instances are effectively running in a truly parallel manner.
The latency will be slightly higher than the example before (2x less cores being used), but the throughput will be again 2x higher.
### 7.2. Smart dispatching - Allocating different model instances for different problem sizes
One another possibility offered by this setup is to have multiple instances carefully tuned for various problem sizes.
With a smart dispatching approach, one can redirect incoming requests to the right configuration giving the best latency depending on the request workload.
```shell
# Small-sized problems (sequence length <= 32) use only 8 cores (on socket 0 - 8/24 cores used)
numactl -C 0-7 -m 0 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=32 backend.name=pytorch backend.num_threads=8
# Medium-sized problems (32 > sequence <= 384) use remaining 16 cores (on socket 0 - (8+16)/24 cores used)
numactl -C 8-23 -m 0 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=128 backend.name=pytorch backend.num_threads=16
# Large sized problems (sequence >= 384) use the entire CPU (on socket 1 - 24/24 cores used)
numactl -C 24-37 -m 1 python3 src/main.py model=bert-base-cased batch_size=1 sequence_length=384 backend.name=pytorch backend.num_threads=24
```
## 8. Batch size scaling - Improving throughput and latency with multiple parallel & independent model instances
One another very interesting direction for scaling up inference is to actually put some more model instances into the pool
along with reducing the actual workload each instance receives proportionally.
This method actually changes both the size of the problem (_batch size_), and the resources involved in the computation (_cores_).
To illustrate, imagine you have a server with `C` CPU cores, and you want to run a workload containing B samples with S tokens.
You can represent this workload as a tensor of shape `[B, S]`, B being the size of the batch and S being the maximum sequence length within the B samples.
For all the instances (`N`), each of them executes on `C / N` cores and would receive a subset of the task `[B / N, S]`.
Each instance doesn't receive the global batch but instead, they all receive a subset of it `[B / N, S]` thus the name **Batch Size Scaling**.
In order to highlight the benefits of such scaling method, the charts below reports both the latencies when scaling up model instances along with the effects on the throughput.
When looking at the results, let's focus on the latency and the throughput aspects:
On one hand, we are taking the maximum latency over the pool of instances to reflect the time it takes to process all the samples in the batch.
Putting it differently, as instances operate in a truly parallel fashion, the time it takes to gather all the batch chunks from all the instances
is driven by the longest time it takes for individual instance in the pool to get their chunk done.
As you can see below on Figure 7., the actual latency gain when increasing the number of instances is really dependent of the problem size.
In all cases, we can find an optimal resource allocation (batch size & number of instances) to minimize our latency but, there is no specific pattern on the number of cores to involve in the computation.
Also, it is important to notice the results might look totally different on another system _(i.e. Operating System, Kernel Version, Framework version, etc.)_
Figure 8. sums up the best multi-instance configuration when targeting minimum latency by taking the minimum over the number of instances involved.
For instance, for `{batch = 8, sequence length = 128}` using 4 instances (each with `{batch = 2}` and 12 cores) gives the best latency measurements.
The Figure 9. reports all the setups minimizing latency for both PyTorch and TensorFlow for various problem-sizes.
_**Spoiler**: There are numerous other optimizations we will discuss in a follow-up blog post which will substantially impact this chart._
<br>
<figure class="image">
<img alt="Batch scaling experiment for PyTorch and Tensorflow" src="assets/21_bert_cpu_scaling_part_1/imgs/batch_scaling_exp.svg" style="width:100%"/>
<figcaption>Figure 8. Max latency evolution with respect to number of instances for a total batch size of 8</figcaption>
</figure>
<br>
<br>
<figure class="image">
<img alt="Optimal number of instance minimizing overall latency for a total batch size of 8" src="assets/21_bert_cpu_scaling_part_1/imgs/batch_size_scaling_latency_optimal_nb_instances.svg" style="width:100%"/>
<figcaption>Figure 9. Optimal number of instance minimizing overall latency for a total batch size of 8</figcaption>
</figure>
<br>
On a second hand, we observe the throughput as the sum of all the model instance executing in parallel.
It allows us to visualize the scalability of the system when adding more and more instances each of them with fewer resources but also proportional workload.
Here, the results show almost linear scalability and thus an optimal hardware usage.
<figure class="image">
<img alt="Batch scaling experiment for PyTorch and Tensorflow" src="assets/21_bert_cpu_scaling_part_1/imgs/batch_scaling_exp_throughput.svg" style="width:100%"/>
<figcaption>Figure 10. Sum throughput with respect to number of instances for a total batch size of 8</figcaption>
</figure>
<br>
## 9. Conclusion
Through this blog post, we covered out-of-box BERT inference performance one can expect for PyTorch and TensorFlow,
from a simple PyPi install and without further tuning.
It is important to highlight results provided here reflects out-of-the-box framework setup hence, they might not provide the absolute best performances.
We decided to not include optimizations as part of this blog post to focus on hardware and efficiency.
Optimizations will be discussed in the second part! 🚀
Then, we covered and detailed the impact, and the importance of setting the thread affinity along with the trade-off between the target problem size, and the number of cores required for achieving the task.
Also, it is important to define **which criteria** _(i.e. latency vs throughput)_ to use when optimizing your deployment as the resulting setups might be totally different.
On a more general note, small problem sizes (_short sequences and/or small batches_) might require much fewer cores to achieve the best possible latency than big problems (_very long sequences and/or big batches_).
It is interesting to cover all these aspects when thinking about the final deployment platform as it might cut the cost of the infrastructure drastically.
For instance, our 48 cores machine charges **4.848\$/h** whereas a smaller instances with only 8 cores lowers the cost to **0.808\$/h**, leading to a **6x cost reduction**.
Last but not least, many of the knobs discussed along this blog post can be automatically tuned through a [launcher script](https://github.com/huggingface/tune/blob/main/launcher.py)
highly inspired from the original script made by Intel and available [here](https://github.com/intel/intel-extension-for-pytorch/blob/master/intel_pytorch_extension_py/launch.py).
The launcher script is able to automatically starts your python process(es) with the correct thread affinity, effectively
splitting resources across instances along with many other performances tips! We will detail many of this tips in the second part 🧐.
In the follow-up blog post, more advanced settings and tuning techniques to decrease model latency even further will be involved, such as:
- Launcher script walk-through
- Tuning the memory allocation library
- Using Linux's Transparent Huge Pages mechanisms
- Using vendor-specific Math/Parallel libraries
Stay tuned! 🤗
## Acknowledgments
- [Omry Yadan](https://github.com/omry) (Facebook FAIR) - Author of [OmegaConf](https://github.com/omry/omegaconf) & [Hydra](https://github.com/facebookresearch/hydra) for all the tips setting up Hydra correctly.
- All Intel & Intel Labs' NLP colleagues - For the ongoing optimizations and research efforts they are putting into transformers and more generally in the NLP field.
- Hugging Face colleagues - For all the comments and improvements in the reviewing process.
## References
1. [Benchmarking Transformers: PyTorch and TensorFlow](https://medium.com/huggingface/benchmarking-transformers-pytorch-and-tensorflow-e2917fb891c2)
2. [HuggingFace's Transformers: State-of-the-art Natural Language Processing](https://arxiv.org/abs/1910.03771v2)
3. [HuggingFace's Model Hub](https://huggingface.co/models)
4. [BERT - Pre-training of Deep Bidirectional Transformers for Language Understanding (Devlin & al. 2018)](https://arxiv.org/abs/1810.04805v1)
5. [Illustrated Transformer blogpost from Jay Alammar](https://jalammar.github.io/illustrated-transformer/)
6. [PyTorch - TorchScript](https://pytorch.org/docs/stable/jit.html)
7. [Google Accelerated Linear Algebra (XLA)](https://www.tensorflow.org/xla)
8. [ONNX Runtime - Optimize and Accelerate Machine Learning Inferencing and Training](https://www.onnxruntime.ai/)
9. [Q8BERT - Quantized 8Bit BERT (Zafrir & al. 2019)](https://arxiv.org/abs/1910.06188)
10. [OpenMP](https://www.openmp.org/)
11. [Intel oneDNN](https://software.intel.com/content/www/us/en/develop/documentation/oneapi-programming-guide/top/api-based-programming/intel-oneapi-deep-neural-network-library-onednn.html)
12. [Intel® Hyper-Threading Technology - Technical User Guide](http://www.cslab.ece.ntua.gr/courses/advcomparch/2007/material/readings/Intel%20Hyper-Threading%20Technology.pdf)
13. [Introduction to Hyper-Threading Technology](https://software.intel.com/content/www/us/en/develop/articles/introduction-to-hyper-threading-technology.html)
14. [BLAS (Basic Linear Algebra Subprogram) - Wikipedia](https://en.wikipedia.org/wiki/Basic_Linear_Algebra_Subprograms#Level_3)
15. [Optimizing Applications for NUMA](https://software.intel.com/content/www/us/en/develop/articles/optimizing-applications-for-numa.html)
| huggingface/blog/blob/main/bert-cpu-scaling-part-1.md |
FrameworkSwitchCourse {fw} />
# Introduction[[introduction]]
<CourseFloatingBanner
chapter={3}
classNames="absolute z-10 right-0 top-0"
/>
In [Chapter 2](/course/chapter2) we explored how to use tokenizers and pretrained models to make predictions. But what if you want to fine-tune a pretrained model for your own dataset? That's the topic of this chapter! You will learn:
{#if fw === 'pt'}
* How to prepare a large dataset from the Hub
* How to use the high-level `Trainer` API to fine-tune a model
* How to use a custom training loop
* How to leverage the 🤗 Accelerate library to easily run that custom training loop on any distributed setup
{:else}
* How to prepare a large dataset from the Hub
* How to use Keras to fine-tune a model
* How to use Keras to get predictions
* How to use a custom metric
{/if}
In order to upload your trained checkpoints to the Hugging Face Hub, you will need a huggingface.co account: [create an account](https://huggingface.co/join) | huggingface/course/blob/main/chapters/en/chapter3/1.mdx |
PyTorch Image Models
- [What's New](#whats-new)
- [Introduction](#introduction)
- [Models](#models)
- [Features](#features)
- [Results](#results)
- [Getting Started (Documentation)](#getting-started-documentation)
- [Train, Validation, Inference Scripts](#train-validation-inference-scripts)
- [Awesome PyTorch Resources](#awesome-pytorch-resources)
- [Licenses](#licenses)
- [Citing](#citing)
## What's New
❗Updates after Oct 10, 2022 are available in version >= 0.9❗
* Many changes since the last 0.6.x stable releases. They were previewed in 0.8.x dev releases but not everyone transitioned.
* `timm.models.layers` moved to `timm.layers`:
* `from timm.models.layers import name` will still work via deprecation mapping (but please transition to `timm.layers`).
* `import timm.models.layers.module` or `from timm.models.layers.module import name` needs to be changed now.
* Builder, helper, non-model modules in `timm.models` have a `_` prefix added, ie `timm.models.helpers` -> `timm.models._helpers`, there are temporary deprecation mapping files but those will be removed.
* All models now support `architecture.pretrained_tag` naming (ex `resnet50.rsb_a1`).
* The pretrained_tag is the specific weight variant (different head) for the architecture.
* Using only `architecture` defaults to the first weights in the default_cfgs for that model architecture.
* In adding pretrained tags, many model names that existed to differentiate were renamed to use the tag (ex: `vit_base_patch16_224_in21k` -> `vit_base_patch16_224.augreg_in21k`). There are deprecation mappings for these.
* A number of models had their checkpoints remaped to match architecture changes needed to better support `features_only=True`, there are `checkpoint_filter_fn` methods in any model module that was remapped. These can be passed to `timm.models.load_checkpoint(..., filter_fn=timm.models.swin_transformer_v2.checkpoint_filter_fn)` to remap your existing checkpoint.
* The Hugging Face Hub (https://huggingface.co/timm) is now the primary source for `timm` weights. Model cards include link to papers, original source, license.
* Previous 0.6.x can be cloned from [0.6.x](https://github.com/rwightman/pytorch-image-models/tree/0.6.x) branch or installed via pip with version.
### Nov 23, 2023
* Added EfficientViT-Large models, thanks [SeeFun](https://github.com/seefun)
* Fix Python 3.7 compat, will be dropping support for it soon
* Other misc fixes
* Release 0.9.12
### Nov 20, 2023
* Added significant flexibility for Hugging Face Hub based timm models via `model_args` config entry. `model_args` will be passed as kwargs through to models on creation.
* See example at https://huggingface.co/gaunernst/vit_base_patch16_1024_128.audiomae_as2m_ft_as20k/blob/main/config.json
* Usage: https://github.com/huggingface/pytorch-image-models/discussions/2035
* Updated imagenet eval and test set csv files with latest models
* `vision_transformer.py` typing and doc cleanup by [Laureηt](https://github.com/Laurent2916)
* 0.9.11 release
### Nov 3, 2023
* [DFN (Data Filtering Networks)](https://huggingface.co/papers/2309.17425) and [MetaCLIP](https://huggingface.co/papers/2309.16671) ViT weights added
* DINOv2 'register' ViT model weights added (https://huggingface.co/papers/2309.16588, https://huggingface.co/papers/2304.07193)
* Add `quickgelu` ViT variants for OpenAI, DFN, MetaCLIP weights that use it (less efficient)
* Improved typing added to ResNet, MobileNet-v3 thanks to [Aryan](https://github.com/a-r-r-o-w)
* ImageNet-12k fine-tuned (from LAION-2B CLIP) `convnext_xxlarge`
* 0.9.9 release
### Oct 20, 2023
* [SigLIP](https://huggingface.co/papers/2303.15343) image tower weights supported in `vision_transformer.py`.
* Great potential for fine-tune and downstream feature use.
* Experimental 'register' support in vit models as per [Vision Transformers Need Registers](https://huggingface.co/papers/2309.16588)
* Updated RepViT with new weight release. Thanks [wangao](https://github.com/jameslahm)
* Add patch resizing support (on pretrained weight load) to Swin models
* 0.9.8 release pending
### Sep 1, 2023
* TinyViT added by [SeeFun](https://github.com/seefun)
* Fix EfficientViT (MIT) to use torch.autocast so it works back to PT 1.10
* 0.9.7 release
### Aug 28, 2023
* Add dynamic img size support to models in `vision_transformer.py`, `vision_transformer_hybrid.py`, `deit.py`, and `eva.py` w/o breaking backward compat.
* Add `dynamic_img_size=True` to args at model creation time to allow changing the grid size (interpolate abs and/or ROPE pos embed each forward pass).
* Add `dynamic_img_pad=True` to allow image sizes that aren't divisible by patch size (pad bottom right to patch size each forward pass).
* Enabling either dynamic mode will break FX tracing unless PatchEmbed module added as leaf.
* Existing method of resizing position embedding by passing different `img_size` (interpolate pretrained embed weights once) on creation still works.
* Existing method of changing `patch_size` (resize pretrained patch_embed weights once) on creation still works.
* Example validation cmd `python validate.py /imagenet --model vit_base_patch16_224 --amp --amp-dtype bfloat16 --img-size 255 --crop-pct 1.0 --model-kwargs dynamic_img_size=True dyamic_img_pad=True`
### Aug 25, 2023
* Many new models since last release
* FastViT - https://arxiv.org/abs/2303.14189
* MobileOne - https://arxiv.org/abs/2206.04040
* InceptionNeXt - https://arxiv.org/abs/2303.16900
* RepGhostNet - https://arxiv.org/abs/2211.06088 (thanks https://github.com/ChengpengChen)
* GhostNetV2 - https://arxiv.org/abs/2211.12905 (thanks https://github.com/yehuitang)
* EfficientViT (MSRA) - https://arxiv.org/abs/2305.07027 (thanks https://github.com/seefun)
* EfficientViT (MIT) - https://arxiv.org/abs/2205.14756 (thanks https://github.com/seefun)
* Add `--reparam` arg to `benchmark.py`, `onnx_export.py`, and `validate.py` to trigger layer reparameterization / fusion for models with any one of `reparameterize()`, `switch_to_deploy()` or `fuse()`
* Including FastViT, MobileOne, RepGhostNet, EfficientViT (MSRA), RepViT, RepVGG, and LeViT
* Preparing 0.9.6 'back to school' release
### Aug 11, 2023
* Swin, MaxViT, CoAtNet, and BEiT models support resizing of image/window size on creation with adaptation of pretrained weights
* Example validation cmd to test w/ non-square resize `python validate.py /imagenet --model swin_base_patch4_window7_224.ms_in22k_ft_in1k --amp --amp-dtype bfloat16 --input-size 3 256 320 --model-kwargs window_size=8,10 img_size=256,320`
### Aug 3, 2023
* Add GluonCV weights for HRNet w18_small and w18_small_v2. Converted by [SeeFun](https://github.com/seefun)
* Fix `selecsls*` model naming regression
* Patch and position embedding for ViT/EVA works for bfloat16/float16 weights on load (or activations for on-the-fly resize)
* v0.9.5 release prep
### July 27, 2023
* Added timm trained `seresnextaa201d_32x8d.sw_in12k_ft_in1k_384` weights (and `.sw_in12k` pretrain) with 87.3% top-1 on ImageNet-1k, best ImageNet ResNet family model I'm aware of.
* RepViT model and weights (https://arxiv.org/abs/2307.09283) added by [wangao](https://github.com/jameslahm)
* I-JEPA ViT feature weights (no classifier) added by [SeeFun](https://github.com/seefun)
* SAM-ViT (segment anything) feature weights (no classifier) added by [SeeFun](https://github.com/seefun)
* Add support for alternative feat extraction methods and -ve indices to EfficientNet
* Add NAdamW optimizer
* Misc fixes
### May 11, 2023
* `timm` 0.9 released, transition from 0.8.xdev releases
### May 10, 2023
* Hugging Face Hub downloading is now default, 1132 models on https://huggingface.co/timm, 1163 weights in `timm`
* DINOv2 vit feature backbone weights added thanks to [Leng Yue](https://github.com/leng-yue)
* FB MAE vit feature backbone weights added
* OpenCLIP DataComp-XL L/14 feat backbone weights added
* MetaFormer (poolformer-v2, caformer, convformer, updated poolformer (v1)) w/ weights added by [Fredo Guan](https://github.com/fffffgggg54)
* Experimental `get_intermediate_layers` function on vit/deit models for grabbing hidden states (inspired by DINO impl). This is WIP and may change significantly... feedback welcome.
* Model creation throws error if `pretrained=True` and no weights exist (instead of continuing with random initialization)
* Fix regression with inception / nasnet TF sourced weights with 1001 classes in original classifiers
* bitsandbytes (https://github.com/TimDettmers/bitsandbytes) optimizers added to factory, use `bnb` prefix, ie `bnbadam8bit`
* Misc cleanup and fixes
* Final testing before switching to a 0.9 and bringing `timm` out of pre-release state
### April 27, 2023
* 97% of `timm` models uploaded to HF Hub and almost all updated to support multi-weight pretrained configs
* Minor cleanup and refactoring of another batch of models as multi-weight added. More fused_attn (F.sdpa) and features_only support, and torchscript fixes.
### April 21, 2023
* Gradient accumulation support added to train script and tested (`--grad-accum-steps`), thanks [Taeksang Kim](https://github.com/voidbag)
* More weights on HF Hub (cspnet, cait, volo, xcit, tresnet, hardcorenas, densenet, dpn, vovnet, xception_aligned)
* Added `--head-init-scale` and `--head-init-bias` to train.py to scale classiifer head and set fixed bias for fine-tune
* Remove all InplaceABN (`inplace_abn`) use, replaced use in tresnet with standard BatchNorm (modified weights accordingly).
### April 12, 2023
* Add ONNX export script, validate script, helpers that I've had kicking around for along time. Tweak 'same' padding for better export w/ recent ONNX + pytorch.
* Refactor dropout args for vit and vit-like models, separate drop_rate into `drop_rate` (classifier dropout), `proj_drop_rate` (block mlp / out projections), `pos_drop_rate` (position embedding drop), `attn_drop_rate` (attention dropout). Also add patch dropout (FLIP) to vit and eva models.
* fused F.scaled_dot_product_attention support to more vit models, add env var (TIMM_FUSED_ATTN) to control, and config interface to enable/disable
* Add EVA-CLIP backbones w/ image tower weights, all the way up to 4B param 'enormous' model, and 336x336 OpenAI ViT mode that was missed.
### April 5, 2023
* ALL ResNet models pushed to Hugging Face Hub with multi-weight support
* All past `timm` trained weights added with recipe based tags to differentiate
* All ResNet strikes back A1/A2/A3 (seed 0) and R50 example B/C1/C2/D weights available
* Add torchvision v2 recipe weights to existing torchvision originals
* See comparison table in https://huggingface.co/timm/seresnextaa101d_32x8d.sw_in12k_ft_in1k_288#model-comparison
* New ImageNet-12k + ImageNet-1k fine-tunes available for a few anti-aliased ResNet models
* `resnetaa50d.sw_in12k_ft_in1k` - 81.7 @ 224, 82.6 @ 288
* `resnetaa101d.sw_in12k_ft_in1k` - 83.5 @ 224, 84.1 @ 288
* `seresnextaa101d_32x8d.sw_in12k_ft_in1k` - 86.0 @ 224, 86.5 @ 288
* `seresnextaa101d_32x8d.sw_in12k_ft_in1k_288` - 86.5 @ 288, 86.7 @ 320
### March 31, 2023
* Add first ConvNext-XXLarge CLIP -> IN-1k fine-tune and IN-12k intermediate fine-tunes for convnext-base/large CLIP models.
| model |top1 |top5 |img_size|param_count|gmacs |macts |
|----------------------------------------------------------------------------------------------------------------------|------|------|--------|-----------|------|------|
| [convnext_xxlarge.clip_laion2b_soup_ft_in1k](https://huggingface.co/timm/convnext_xxlarge.clip_laion2b_soup_ft_in1k) |88.612|98.704|256 |846.47 |198.09|124.45|
| convnext_large_mlp.clip_laion2b_soup_ft_in12k_in1k_384 |88.312|98.578|384 |200.13 |101.11|126.74|
| convnext_large_mlp.clip_laion2b_soup_ft_in12k_in1k_320 |87.968|98.47 |320 |200.13 |70.21 |88.02 |
| convnext_base.clip_laion2b_augreg_ft_in12k_in1k_384 |87.138|98.212|384 |88.59 |45.21 |84.49 |
| convnext_base.clip_laion2b_augreg_ft_in12k_in1k |86.344|97.97 |256 |88.59 |20.09 |37.55 |
* Add EVA-02 MIM pretrained and fine-tuned weights, push to HF hub and update model cards for all EVA models. First model over 90% top-1 (99% top-5)! Check out the original code & weights at https://github.com/baaivision/EVA for more details on their work blending MIM, CLIP w/ many model, dataset, and train recipe tweaks.
| model |top1 |top5 |param_count|img_size|
|----------------------------------------------------|------|------|-----------|--------|
| [eva02_large_patch14_448.mim_m38m_ft_in22k_in1k](https://huggingface.co/timm/eva02_large_patch14_448.mim_m38m_ft_in1k) |90.054|99.042|305.08 |448 |
| eva02_large_patch14_448.mim_in22k_ft_in22k_in1k |89.946|99.01 |305.08 |448 |
| eva_giant_patch14_560.m30m_ft_in22k_in1k |89.792|98.992|1014.45 |560 |
| eva02_large_patch14_448.mim_in22k_ft_in1k |89.626|98.954|305.08 |448 |
| eva02_large_patch14_448.mim_m38m_ft_in1k |89.57 |98.918|305.08 |448 |
| eva_giant_patch14_336.m30m_ft_in22k_in1k |89.56 |98.956|1013.01 |336 |
| eva_giant_patch14_336.clip_ft_in1k |89.466|98.82 |1013.01 |336 |
| eva_large_patch14_336.in22k_ft_in22k_in1k |89.214|98.854|304.53 |336 |
| eva_giant_patch14_224.clip_ft_in1k |88.882|98.678|1012.56 |224 |
| eva02_base_patch14_448.mim_in22k_ft_in22k_in1k |88.692|98.722|87.12 |448 |
| eva_large_patch14_336.in22k_ft_in1k |88.652|98.722|304.53 |336 |
| eva_large_patch14_196.in22k_ft_in22k_in1k |88.592|98.656|304.14 |196 |
| eva02_base_patch14_448.mim_in22k_ft_in1k |88.23 |98.564|87.12 |448 |
| eva_large_patch14_196.in22k_ft_in1k |87.934|98.504|304.14 |196 |
| eva02_small_patch14_336.mim_in22k_ft_in1k |85.74 |97.614|22.13 |336 |
| eva02_tiny_patch14_336.mim_in22k_ft_in1k |80.658|95.524|5.76 |336 |
* Multi-weight and HF hub for DeiT and MLP-Mixer based models
### March 22, 2023
* More weights pushed to HF hub along with multi-weight support, including: `regnet.py`, `rexnet.py`, `byobnet.py`, `resnetv2.py`, `swin_transformer.py`, `swin_transformer_v2.py`, `swin_transformer_v2_cr.py`
* Swin Transformer models support feature extraction (NCHW feat maps for `swinv2_cr_*`, and NHWC for all others) and spatial embedding outputs.
* FocalNet (from https://github.com/microsoft/FocalNet) models and weights added with significant refactoring, feature extraction, no fixed resolution / sizing constraint
* RegNet weights increased with HF hub push, SWAG, SEER, and torchvision v2 weights. SEER is pretty poor wrt to performance for model size, but possibly useful.
* More ImageNet-12k pretrained and 1k fine-tuned `timm` weights:
* `rexnetr_200.sw_in12k_ft_in1k` - 82.6 @ 224, 83.2 @ 288
* `rexnetr_300.sw_in12k_ft_in1k` - 84.0 @ 224, 84.5 @ 288
* `regnety_120.sw_in12k_ft_in1k` - 85.0 @ 224, 85.4 @ 288
* `regnety_160.lion_in12k_ft_in1k` - 85.6 @ 224, 86.0 @ 288
* `regnety_160.sw_in12k_ft_in1k` - 85.6 @ 224, 86.0 @ 288 (compare to SWAG PT + 1k FT this is same BUT much lower res, blows SEER FT away)
* Model name deprecation + remapping functionality added (a milestone for bringing 0.8.x out of pre-release). Mappings being added...
* Minor bug fixes and improvements.
### Feb 26, 2023
* Add ConvNeXt-XXLarge CLIP pretrained image tower weights for fine-tune & features (fine-tuning TBD) -- see [model card](https://huggingface.co/laion/CLIP-convnext_xxlarge-laion2B-s34B-b82K-augreg-soup)
* Update `convnext_xxlarge` default LayerNorm eps to 1e-5 (for CLIP weights, improved stability)
* 0.8.15dev0
### Feb 20, 2023
* Add 320x320 `convnext_large_mlp.clip_laion2b_ft_320` and `convnext_lage_mlp.clip_laion2b_ft_soup_320` CLIP image tower weights for features & fine-tune
* 0.8.13dev0 pypi release for latest changes w/ move to huggingface org
### Feb 16, 2023
* `safetensor` checkpoint support added
* Add ideas from 'Scaling Vision Transformers to 22 B. Params' (https://arxiv.org/abs/2302.05442) -- qk norm, RmsNorm, parallel block
* Add F.scaled_dot_product_attention support (PyTorch 2.0 only) to `vit_*`, `vit_relpos*`, `coatnet` / `maxxvit` (to start)
* Lion optimizer (w/ multi-tensor option) added (https://arxiv.org/abs/2302.06675)
* gradient checkpointing works with `features_only=True`
### Feb 7, 2023
* New inference benchmark numbers added in [results](results/) folder.
* Add convnext LAION CLIP trained weights and initial set of in1k fine-tunes
* `convnext_base.clip_laion2b_augreg_ft_in1k` - 86.2% @ 256x256
* `convnext_base.clip_laiona_augreg_ft_in1k_384` - 86.5% @ 384x384
* `convnext_large_mlp.clip_laion2b_augreg_ft_in1k` - 87.3% @ 256x256
* `convnext_large_mlp.clip_laion2b_augreg_ft_in1k_384` - 87.9% @ 384x384
* Add DaViT models. Supports `features_only=True`. Adapted from https://github.com/dingmyu/davit by [Fredo](https://github.com/fffffgggg54).
* Use a common NormMlpClassifierHead across MaxViT, ConvNeXt, DaViT
* Add EfficientFormer-V2 model, update EfficientFormer, and refactor LeViT (closely related architectures). Weights on HF hub.
* New EfficientFormer-V2 arch, significant refactor from original at (https://github.com/snap-research/EfficientFormer). Supports `features_only=True`.
* Minor updates to EfficientFormer.
* Refactor LeViT models to stages, add `features_only=True` support to new `conv` variants, weight remap required.
* Move ImageNet meta-data (synsets, indices) from `/results` to [`timm/data/_info`](timm/data/_info/).
* Add ImageNetInfo / DatasetInfo classes to provide labelling for various ImageNet classifier layouts in `timm`
* Update `inference.py` to use, try: `python inference.py /folder/to/images --model convnext_small.in12k --label-type detail --topk 5`
* Ready for 0.8.10 pypi pre-release (final testing).
### Jan 20, 2023
* Add two convnext 12k -> 1k fine-tunes at 384x384
* `convnext_tiny.in12k_ft_in1k_384` - 85.1 @ 384
* `convnext_small.in12k_ft_in1k_384` - 86.2 @ 384
* Push all MaxxViT weights to HF hub, and add new ImageNet-12k -> 1k fine-tunes for `rw` base MaxViT and CoAtNet 1/2 models
|model |top1 |top5 |samples / sec |Params (M) |GMAC |Act (M)|
|------------------------------------------------------------------------------------------------------------------------|----:|----:|--------------:|--------------:|-----:|------:|
|[maxvit_xlarge_tf_512.in21k_ft_in1k](https://huggingface.co/timm/maxvit_xlarge_tf_512.in21k_ft_in1k) |88.53|98.64| 21.76| 475.77|534.14|1413.22|
|[maxvit_xlarge_tf_384.in21k_ft_in1k](https://huggingface.co/timm/maxvit_xlarge_tf_384.in21k_ft_in1k) |88.32|98.54| 42.53| 475.32|292.78| 668.76|
|[maxvit_base_tf_512.in21k_ft_in1k](https://huggingface.co/timm/maxvit_base_tf_512.in21k_ft_in1k) |88.20|98.53| 50.87| 119.88|138.02| 703.99|
|[maxvit_large_tf_512.in21k_ft_in1k](https://huggingface.co/timm/maxvit_large_tf_512.in21k_ft_in1k) |88.04|98.40| 36.42| 212.33|244.75| 942.15|
|[maxvit_large_tf_384.in21k_ft_in1k](https://huggingface.co/timm/maxvit_large_tf_384.in21k_ft_in1k) |87.98|98.56| 71.75| 212.03|132.55| 445.84|
|[maxvit_base_tf_384.in21k_ft_in1k](https://huggingface.co/timm/maxvit_base_tf_384.in21k_ft_in1k) |87.92|98.54| 104.71| 119.65| 73.80| 332.90|
|[maxvit_rmlp_base_rw_384.sw_in12k_ft_in1k](https://huggingface.co/timm/maxvit_rmlp_base_rw_384.sw_in12k_ft_in1k) |87.81|98.37| 106.55| 116.14| 70.97| 318.95|
|[maxxvitv2_rmlp_base_rw_384.sw_in12k_ft_in1k](https://huggingface.co/timm/maxxvitv2_rmlp_base_rw_384.sw_in12k_ft_in1k) |87.47|98.37| 149.49| 116.09| 72.98| 213.74|
|[coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k](https://huggingface.co/timm/coatnet_rmlp_2_rw_384.sw_in12k_ft_in1k) |87.39|98.31| 160.80| 73.88| 47.69| 209.43|
|[maxvit_rmlp_base_rw_224.sw_in12k_ft_in1k](https://huggingface.co/timm/maxvit_rmlp_base_rw_224.sw_in12k_ft_in1k) |86.89|98.02| 375.86| 116.14| 23.15| 92.64|
|[maxxvitv2_rmlp_base_rw_224.sw_in12k_ft_in1k](https://huggingface.co/timm/maxxvitv2_rmlp_base_rw_224.sw_in12k_ft_in1k) |86.64|98.02| 501.03| 116.09| 24.20| 62.77|
|[maxvit_base_tf_512.in1k](https://huggingface.co/timm/maxvit_base_tf_512.in1k) |86.60|97.92| 50.75| 119.88|138.02| 703.99|
|[coatnet_2_rw_224.sw_in12k_ft_in1k](https://huggingface.co/timm/coatnet_2_rw_224.sw_in12k_ft_in1k) |86.57|97.89| 631.88| 73.87| 15.09| 49.22|
|[maxvit_large_tf_512.in1k](https://huggingface.co/timm/maxvit_large_tf_512.in1k) |86.52|97.88| 36.04| 212.33|244.75| 942.15|
|[coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k](https://huggingface.co/timm/coatnet_rmlp_2_rw_224.sw_in12k_ft_in1k) |86.49|97.90| 620.58| 73.88| 15.18| 54.78|
|[maxvit_base_tf_384.in1k](https://huggingface.co/timm/maxvit_base_tf_384.in1k) |86.29|97.80| 101.09| 119.65| 73.80| 332.90|
|[maxvit_large_tf_384.in1k](https://huggingface.co/timm/maxvit_large_tf_384.in1k) |86.23|97.69| 70.56| 212.03|132.55| 445.84|
|[maxvit_small_tf_512.in1k](https://huggingface.co/timm/maxvit_small_tf_512.in1k) |86.10|97.76| 88.63| 69.13| 67.26| 383.77|
|[maxvit_tiny_tf_512.in1k](https://huggingface.co/timm/maxvit_tiny_tf_512.in1k) |85.67|97.58| 144.25| 31.05| 33.49| 257.59|
|[maxvit_small_tf_384.in1k](https://huggingface.co/timm/maxvit_small_tf_384.in1k) |85.54|97.46| 188.35| 69.02| 35.87| 183.65|
|[maxvit_tiny_tf_384.in1k](https://huggingface.co/timm/maxvit_tiny_tf_384.in1k) |85.11|97.38| 293.46| 30.98| 17.53| 123.42|
|[maxvit_large_tf_224.in1k](https://huggingface.co/timm/maxvit_large_tf_224.in1k) |84.93|96.97| 247.71| 211.79| 43.68| 127.35|
|[coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k](https://huggingface.co/timm/coatnet_rmlp_1_rw2_224.sw_in12k_ft_in1k) |84.90|96.96| 1025.45| 41.72| 8.11| 40.13|
|[maxvit_base_tf_224.in1k](https://huggingface.co/timm/maxvit_base_tf_224.in1k) |84.85|96.99| 358.25| 119.47| 24.04| 95.01|
|[maxxvit_rmlp_small_rw_256.sw_in1k](https://huggingface.co/timm/maxxvit_rmlp_small_rw_256.sw_in1k) |84.63|97.06| 575.53| 66.01| 14.67| 58.38|
|[coatnet_rmlp_2_rw_224.sw_in1k](https://huggingface.co/timm/coatnet_rmlp_2_rw_224.sw_in1k) |84.61|96.74| 625.81| 73.88| 15.18| 54.78|
|[maxvit_rmlp_small_rw_224.sw_in1k](https://huggingface.co/timm/maxvit_rmlp_small_rw_224.sw_in1k) |84.49|96.76| 693.82| 64.90| 10.75| 49.30|
|[maxvit_small_tf_224.in1k](https://huggingface.co/timm/maxvit_small_tf_224.in1k) |84.43|96.83| 647.96| 68.93| 11.66| 53.17|
|[maxvit_rmlp_tiny_rw_256.sw_in1k](https://huggingface.co/timm/maxvit_rmlp_tiny_rw_256.sw_in1k) |84.23|96.78| 807.21| 29.15| 6.77| 46.92|
|[coatnet_1_rw_224.sw_in1k](https://huggingface.co/timm/coatnet_1_rw_224.sw_in1k) |83.62|96.38| 989.59| 41.72| 8.04| 34.60|
|[maxvit_tiny_rw_224.sw_in1k](https://huggingface.co/timm/maxvit_tiny_rw_224.sw_in1k) |83.50|96.50| 1100.53| 29.06| 5.11| 33.11|
|[maxvit_tiny_tf_224.in1k](https://huggingface.co/timm/maxvit_tiny_tf_224.in1k) |83.41|96.59| 1004.94| 30.92| 5.60| 35.78|
|[coatnet_rmlp_1_rw_224.sw_in1k](https://huggingface.co/timm/coatnet_rmlp_1_rw_224.sw_in1k) |83.36|96.45| 1093.03| 41.69| 7.85| 35.47|
|[maxxvitv2_nano_rw_256.sw_in1k](https://huggingface.co/timm/maxxvitv2_nano_rw_256.sw_in1k) |83.11|96.33| 1276.88| 23.70| 6.26| 23.05|
|[maxxvit_rmlp_nano_rw_256.sw_in1k](https://huggingface.co/timm/maxxvit_rmlp_nano_rw_256.sw_in1k) |83.03|96.34| 1341.24| 16.78| 4.37| 26.05|
|[maxvit_rmlp_nano_rw_256.sw_in1k](https://huggingface.co/timm/maxvit_rmlp_nano_rw_256.sw_in1k) |82.96|96.26| 1283.24| 15.50| 4.47| 31.92|
|[maxvit_nano_rw_256.sw_in1k](https://huggingface.co/timm/maxvit_nano_rw_256.sw_in1k) |82.93|96.23| 1218.17| 15.45| 4.46| 30.28|
|[coatnet_bn_0_rw_224.sw_in1k](https://huggingface.co/timm/coatnet_bn_0_rw_224.sw_in1k) |82.39|96.19| 1600.14| 27.44| 4.67| 22.04|
|[coatnet_0_rw_224.sw_in1k](https://huggingface.co/timm/coatnet_0_rw_224.sw_in1k) |82.39|95.84| 1831.21| 27.44| 4.43| 18.73|
|[coatnet_rmlp_nano_rw_224.sw_in1k](https://huggingface.co/timm/coatnet_rmlp_nano_rw_224.sw_in1k) |82.05|95.87| 2109.09| 15.15| 2.62| 20.34|
|[coatnext_nano_rw_224.sw_in1k](https://huggingface.co/timm/coatnext_nano_rw_224.sw_in1k) |81.95|95.92| 2525.52| 14.70| 2.47| 12.80|
|[coatnet_nano_rw_224.sw_in1k](https://huggingface.co/timm/coatnet_nano_rw_224.sw_in1k) |81.70|95.64| 2344.52| 15.14| 2.41| 15.41|
|[maxvit_rmlp_pico_rw_256.sw_in1k](https://huggingface.co/timm/maxvit_rmlp_pico_rw_256.sw_in1k) |80.53|95.21| 1594.71| 7.52| 1.85| 24.86|
### Jan 11, 2023
* Update ConvNeXt ImageNet-12k pretrain series w/ two new fine-tuned weights (and pre FT `.in12k` tags)
* `convnext_nano.in12k_ft_in1k` - 82.3 @ 224, 82.9 @ 288 (previously released)
* `convnext_tiny.in12k_ft_in1k` - 84.2 @ 224, 84.5 @ 288
* `convnext_small.in12k_ft_in1k` - 85.2 @ 224, 85.3 @ 288
### Jan 6, 2023
* Finally got around to adding `--model-kwargs` and `--opt-kwargs` to scripts to pass through rare args directly to model classes from cmd line
* `train.py /imagenet --model resnet50 --amp --model-kwargs output_stride=16 act_layer=silu`
* `train.py /imagenet --model vit_base_patch16_clip_224 --img-size 240 --amp --model-kwargs img_size=240 patch_size=12`
* Cleanup some popular models to better support arg passthrough / merge with model configs, more to go.
### Jan 5, 2023
* ConvNeXt-V2 models and weights added to existing `convnext.py`
* Paper: [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](http://arxiv.org/abs/2301.00808)
* Reference impl: https://github.com/facebookresearch/ConvNeXt-V2 (NOTE: weights currently CC-BY-NC)
### Dec 23, 2022 🎄☃
* Add FlexiViT models and weights from https://github.com/google-research/big_vision (check out paper at https://arxiv.org/abs/2212.08013)
* NOTE currently resizing is static on model creation, on-the-fly dynamic / train patch size sampling is a WIP
* Many more models updated to multi-weight and downloadable via HF hub now (convnext, efficientnet, mobilenet, vision_transformer*, beit)
* More model pretrained tag and adjustments, some model names changed (working on deprecation translations, consider main branch DEV branch right now, use 0.6.x for stable use)
* More ImageNet-12k (subset of 22k) pretrain models popping up:
* `efficientnet_b5.in12k_ft_in1k` - 85.9 @ 448x448
* `vit_medium_patch16_gap_384.in12k_ft_in1k` - 85.5 @ 384x384
* `vit_medium_patch16_gap_256.in12k_ft_in1k` - 84.5 @ 256x256
* `convnext_nano.in12k_ft_in1k` - 82.9 @ 288x288
### Dec 8, 2022
* Add 'EVA l' to `vision_transformer.py`, MAE style ViT-L/14 MIM pretrain w/ EVA-CLIP targets, FT on ImageNet-1k (w/ ImageNet-22k intermediate for some)
* original source: https://github.com/baaivision/EVA
| model | top1 | param_count | gmac | macts | hub |
|:------------------------------------------|-----:|------------:|------:|------:|:----------------------------------------|
| eva_large_patch14_336.in22k_ft_in22k_in1k | 89.2 | 304.5 | 191.1 | 270.2 | [link](https://huggingface.co/BAAI/EVA) |
| eva_large_patch14_336.in22k_ft_in1k | 88.7 | 304.5 | 191.1 | 270.2 | [link](https://huggingface.co/BAAI/EVA) |
| eva_large_patch14_196.in22k_ft_in22k_in1k | 88.6 | 304.1 | 61.6 | 63.5 | [link](https://huggingface.co/BAAI/EVA) |
| eva_large_patch14_196.in22k_ft_in1k | 87.9 | 304.1 | 61.6 | 63.5 | [link](https://huggingface.co/BAAI/EVA) |
### Dec 6, 2022
* Add 'EVA g', BEiT style ViT-g/14 model weights w/ both MIM pretrain and CLIP pretrain to `beit.py`.
* original source: https://github.com/baaivision/EVA
* paper: https://arxiv.org/abs/2211.07636
| model | top1 | param_count | gmac | macts | hub |
|:-----------------------------------------|-------:|--------------:|-------:|--------:|:----------------------------------------|
| eva_giant_patch14_560.m30m_ft_in22k_in1k | 89.8 | 1014.4 | 1906.8 | 2577.2 | [link](https://huggingface.co/BAAI/EVA) |
| eva_giant_patch14_336.m30m_ft_in22k_in1k | 89.6 | 1013 | 620.6 | 550.7 | [link](https://huggingface.co/BAAI/EVA) |
| eva_giant_patch14_336.clip_ft_in1k | 89.4 | 1013 | 620.6 | 550.7 | [link](https://huggingface.co/BAAI/EVA) |
| eva_giant_patch14_224.clip_ft_in1k | 89.1 | 1012.6 | 267.2 | 192.6 | [link](https://huggingface.co/BAAI/EVA) |
### Dec 5, 2022
* Pre-release (`0.8.0dev0`) of multi-weight support (`model_arch.pretrained_tag`). Install with `pip install --pre timm`
* vision_transformer, maxvit, convnext are the first three model impl w/ support
* model names are changing with this (previous _21k, etc. fn will merge), still sorting out deprecation handling
* bugs are likely, but I need feedback so please try it out
* if stability is needed, please use 0.6.x pypi releases or clone from [0.6.x branch](https://github.com/rwightman/pytorch-image-models/tree/0.6.x)
* Support for PyTorch 2.0 compile is added in train/validate/inference/benchmark, use `--torchcompile` argument
* Inference script allows more control over output, select k for top-class index + prob json, csv or parquet output
* Add a full set of fine-tuned CLIP image tower weights from both LAION-2B and original OpenAI CLIP models
| model | top1 | param_count | gmac | macts | hub |
|:-------------------------------------------------|-------:|--------------:|-------:|--------:|:-------------------------------------------------------------------------------------|
| vit_huge_patch14_clip_336.laion2b_ft_in12k_in1k | 88.6 | 632.5 | 391 | 407.5 | [link](https://huggingface.co/timm/vit_huge_patch14_clip_336.laion2b_ft_in12k_in1k) |
| vit_large_patch14_clip_336.openai_ft_in12k_in1k | 88.3 | 304.5 | 191.1 | 270.2 | [link](https://huggingface.co/timm/vit_large_patch14_clip_336.openai_ft_in12k_in1k) |
| vit_huge_patch14_clip_224.laion2b_ft_in12k_in1k | 88.2 | 632 | 167.4 | 139.4 | [link](https://huggingface.co/timm/vit_huge_patch14_clip_224.laion2b_ft_in12k_in1k) |
| vit_large_patch14_clip_336.laion2b_ft_in12k_in1k | 88.2 | 304.5 | 191.1 | 270.2 | [link](https://huggingface.co/timm/vit_large_patch14_clip_336.laion2b_ft_in12k_in1k) |
| vit_large_patch14_clip_224.openai_ft_in12k_in1k | 88.2 | 304.2 | 81.1 | 88.8 | [link](https://huggingface.co/timm/vit_large_patch14_clip_224.openai_ft_in12k_in1k) |
| vit_large_patch14_clip_224.laion2b_ft_in12k_in1k | 87.9 | 304.2 | 81.1 | 88.8 | [link](https://huggingface.co/timm/vit_large_patch14_clip_224.laion2b_ft_in12k_in1k) |
| vit_large_patch14_clip_224.openai_ft_in1k | 87.9 | 304.2 | 81.1 | 88.8 | [link](https://huggingface.co/timm/vit_large_patch14_clip_224.openai_ft_in1k) |
| vit_large_patch14_clip_336.laion2b_ft_in1k | 87.9 | 304.5 | 191.1 | 270.2 | [link](https://huggingface.co/timm/vit_large_patch14_clip_336.laion2b_ft_in1k) |
| vit_huge_patch14_clip_224.laion2b_ft_in1k | 87.6 | 632 | 167.4 | 139.4 | [link](https://huggingface.co/timm/vit_huge_patch14_clip_224.laion2b_ft_in1k) |
| vit_large_patch14_clip_224.laion2b_ft_in1k | 87.3 | 304.2 | 81.1 | 88.8 | [link](https://huggingface.co/timm/vit_large_patch14_clip_224.laion2b_ft_in1k) |
| vit_base_patch16_clip_384.laion2b_ft_in12k_in1k | 87.2 | 86.9 | 55.5 | 101.6 | [link](https://huggingface.co/timm/vit_base_patch16_clip_384.laion2b_ft_in12k_in1k) |
| vit_base_patch16_clip_384.openai_ft_in12k_in1k | 87 | 86.9 | 55.5 | 101.6 | [link](https://huggingface.co/timm/vit_base_patch16_clip_384.openai_ft_in12k_in1k) |
| vit_base_patch16_clip_384.laion2b_ft_in1k | 86.6 | 86.9 | 55.5 | 101.6 | [link](https://huggingface.co/timm/vit_base_patch16_clip_384.laion2b_ft_in1k) |
| vit_base_patch16_clip_384.openai_ft_in1k | 86.2 | 86.9 | 55.5 | 101.6 | [link](https://huggingface.co/timm/vit_base_patch16_clip_384.openai_ft_in1k) |
| vit_base_patch16_clip_224.laion2b_ft_in12k_in1k | 86.2 | 86.6 | 17.6 | 23.9 | [link](https://huggingface.co/timm/vit_base_patch16_clip_224.laion2b_ft_in12k_in1k) |
| vit_base_patch16_clip_224.openai_ft_in12k_in1k | 85.9 | 86.6 | 17.6 | 23.9 | [link](https://huggingface.co/timm/vit_base_patch16_clip_224.openai_ft_in12k_in1k) |
| vit_base_patch32_clip_448.laion2b_ft_in12k_in1k | 85.8 | 88.3 | 17.9 | 23.9 | [link](https://huggingface.co/timm/vit_base_patch32_clip_448.laion2b_ft_in12k_in1k) |
| vit_base_patch16_clip_224.laion2b_ft_in1k | 85.5 | 86.6 | 17.6 | 23.9 | [link](https://huggingface.co/timm/vit_base_patch16_clip_224.laion2b_ft_in1k) |
| vit_base_patch32_clip_384.laion2b_ft_in12k_in1k | 85.4 | 88.3 | 13.1 | 16.5 | [link](https://huggingface.co/timm/vit_base_patch32_clip_384.laion2b_ft_in12k_in1k) |
| vit_base_patch16_clip_224.openai_ft_in1k | 85.3 | 86.6 | 17.6 | 23.9 | [link](https://huggingface.co/timm/vit_base_patch16_clip_224.openai_ft_in1k) |
| vit_base_patch32_clip_384.openai_ft_in12k_in1k | 85.2 | 88.3 | 13.1 | 16.5 | [link](https://huggingface.co/timm/vit_base_patch32_clip_384.openai_ft_in12k_in1k) |
| vit_base_patch32_clip_224.laion2b_ft_in12k_in1k | 83.3 | 88.2 | 4.4 | 5 | [link](https://huggingface.co/timm/vit_base_patch32_clip_224.laion2b_ft_in12k_in1k) |
| vit_base_patch32_clip_224.laion2b_ft_in1k | 82.6 | 88.2 | 4.4 | 5 | [link](https://huggingface.co/timm/vit_base_patch32_clip_224.laion2b_ft_in1k) |
| vit_base_patch32_clip_224.openai_ft_in1k | 81.9 | 88.2 | 4.4 | 5 | [link](https://huggingface.co/timm/vit_base_patch32_clip_224.openai_ft_in1k) |
* Port of MaxViT Tensorflow Weights from official impl at https://github.com/google-research/maxvit
* There was larger than expected drops for the upscaled 384/512 in21k fine-tune weights, possible detail missing, but the 21k FT did seem sensitive to small preprocessing
| model | top1 | param_count | gmac | macts | hub |
|:-----------------------------------|-------:|--------------:|-------:|--------:|:-----------------------------------------------------------------------|
| maxvit_xlarge_tf_512.in21k_ft_in1k | 88.5 | 475.8 | 534.1 | 1413.2 | [link](https://huggingface.co/timm/maxvit_xlarge_tf_512.in21k_ft_in1k) |
| maxvit_xlarge_tf_384.in21k_ft_in1k | 88.3 | 475.3 | 292.8 | 668.8 | [link](https://huggingface.co/timm/maxvit_xlarge_tf_384.in21k_ft_in1k) |
| maxvit_base_tf_512.in21k_ft_in1k | 88.2 | 119.9 | 138 | 704 | [link](https://huggingface.co/timm/maxvit_base_tf_512.in21k_ft_in1k) |
| maxvit_large_tf_512.in21k_ft_in1k | 88 | 212.3 | 244.8 | 942.2 | [link](https://huggingface.co/timm/maxvit_large_tf_512.in21k_ft_in1k) |
| maxvit_large_tf_384.in21k_ft_in1k | 88 | 212 | 132.6 | 445.8 | [link](https://huggingface.co/timm/maxvit_large_tf_384.in21k_ft_in1k) |
| maxvit_base_tf_384.in21k_ft_in1k | 87.9 | 119.6 | 73.8 | 332.9 | [link](https://huggingface.co/timm/maxvit_base_tf_384.in21k_ft_in1k) |
| maxvit_base_tf_512.in1k | 86.6 | 119.9 | 138 | 704 | [link](https://huggingface.co/timm/maxvit_base_tf_512.in1k) |
| maxvit_large_tf_512.in1k | 86.5 | 212.3 | 244.8 | 942.2 | [link](https://huggingface.co/timm/maxvit_large_tf_512.in1k) |
| maxvit_base_tf_384.in1k | 86.3 | 119.6 | 73.8 | 332.9 | [link](https://huggingface.co/timm/maxvit_base_tf_384.in1k) |
| maxvit_large_tf_384.in1k | 86.2 | 212 | 132.6 | 445.8 | [link](https://huggingface.co/timm/maxvit_large_tf_384.in1k) |
| maxvit_small_tf_512.in1k | 86.1 | 69.1 | 67.3 | 383.8 | [link](https://huggingface.co/timm/maxvit_small_tf_512.in1k) |
| maxvit_tiny_tf_512.in1k | 85.7 | 31 | 33.5 | 257.6 | [link](https://huggingface.co/timm/maxvit_tiny_tf_512.in1k) |
| maxvit_small_tf_384.in1k | 85.5 | 69 | 35.9 | 183.6 | [link](https://huggingface.co/timm/maxvit_small_tf_384.in1k) |
| maxvit_tiny_tf_384.in1k | 85.1 | 31 | 17.5 | 123.4 | [link](https://huggingface.co/timm/maxvit_tiny_tf_384.in1k) |
| maxvit_large_tf_224.in1k | 84.9 | 211.8 | 43.7 | 127.4 | [link](https://huggingface.co/timm/maxvit_large_tf_224.in1k) |
| maxvit_base_tf_224.in1k | 84.9 | 119.5 | 24 | 95 | [link](https://huggingface.co/timm/maxvit_base_tf_224.in1k) |
| maxvit_small_tf_224.in1k | 84.4 | 68.9 | 11.7 | 53.2 | [link](https://huggingface.co/timm/maxvit_small_tf_224.in1k) |
| maxvit_tiny_tf_224.in1k | 83.4 | 30.9 | 5.6 | 35.8 | [link](https://huggingface.co/timm/maxvit_tiny_tf_224.in1k) |
### Oct 15, 2022
* Train and validation script enhancements
* Non-GPU (ie CPU) device support
* SLURM compatibility for train script
* HF datasets support (via ReaderHfds)
* TFDS/WDS dataloading improvements (sample padding/wrap for distributed use fixed wrt sample count estimate)
* in_chans !=3 support for scripts / loader
* Adan optimizer
* Can enable per-step LR scheduling via args
* Dataset 'parsers' renamed to 'readers', more descriptive of purpose
* AMP args changed, APEX via `--amp-impl apex`, bfloat16 supportedf via `--amp-dtype bfloat16`
* main branch switched to 0.7.x version, 0.6x forked for stable release of weight only adds
* master -> main branch rename
### Oct 10, 2022
* More weights in `maxxvit` series, incl first ConvNeXt block based `coatnext` and `maxxvit` experiments:
* `coatnext_nano_rw_224` - 82.0 @ 224 (G) -- (uses ConvNeXt conv block, no BatchNorm)
* `maxxvit_rmlp_nano_rw_256` - 83.0 @ 256, 83.7 @ 320 (G) (uses ConvNeXt conv block, no BN)
* `maxvit_rmlp_small_rw_224` - 84.5 @ 224, 85.1 @ 320 (G)
* `maxxvit_rmlp_small_rw_256` - 84.6 @ 256, 84.9 @ 288 (G) -- could be trained better, hparams need tuning (uses ConvNeXt block, no BN)
* `coatnet_rmlp_2_rw_224` - 84.6 @ 224, 85 @ 320 (T)
* NOTE: official MaxVit weights (in1k) have been released at https://github.com/google-research/maxvit -- some extra work is needed to port and adapt since my impl was created independently of theirs and has a few small differences + the whole TF same padding fun.
### Sept 23, 2022
* LAION-2B CLIP image towers supported as pretrained backbones for fine-tune or features (no classifier)
* vit_base_patch32_224_clip_laion2b
* vit_large_patch14_224_clip_laion2b
* vit_huge_patch14_224_clip_laion2b
* vit_giant_patch14_224_clip_laion2b
### Sept 7, 2022
* Hugging Face [`timm` docs](https://huggingface.co/docs/hub/timm) home now exists, look for more here in the future
* Add BEiT-v2 weights for base and large 224x224 models from https://github.com/microsoft/unilm/tree/master/beit2
* Add more weights in `maxxvit` series incl a `pico` (7.5M params, 1.9 GMACs), two `tiny` variants:
* `maxvit_rmlp_pico_rw_256` - 80.5 @ 256, 81.3 @ 320 (T)
* `maxvit_tiny_rw_224` - 83.5 @ 224 (G)
* `maxvit_rmlp_tiny_rw_256` - 84.2 @ 256, 84.8 @ 320 (T)
## Introduction
Py**T**orch **Im**age **M**odels (`timm`) is a collection of image models, layers, utilities, optimizers, schedulers, data-loaders / augmentations, and reference training / validation scripts that aim to pull together a wide variety of SOTA models with ability to reproduce ImageNet training results.
The work of many others is present here. I've tried to make sure all source material is acknowledged via links to github, arxiv papers, etc in the README, documentation, and code docstrings. Please let me know if I missed anything.
## Models
All model architecture families include variants with pretrained weights. There are specific model variants without any weights, it is NOT a bug. Help training new or better weights is always appreciated.
* Aggregating Nested Transformers - https://arxiv.org/abs/2105.12723
* BEiT - https://arxiv.org/abs/2106.08254
* Big Transfer ResNetV2 (BiT) - https://arxiv.org/abs/1912.11370
* Bottleneck Transformers - https://arxiv.org/abs/2101.11605
* CaiT (Class-Attention in Image Transformers) - https://arxiv.org/abs/2103.17239
* CoaT (Co-Scale Conv-Attentional Image Transformers) - https://arxiv.org/abs/2104.06399
* CoAtNet (Convolution and Attention) - https://arxiv.org/abs/2106.04803
* ConvNeXt - https://arxiv.org/abs/2201.03545
* ConvNeXt-V2 - http://arxiv.org/abs/2301.00808
* ConViT (Soft Convolutional Inductive Biases Vision Transformers)- https://arxiv.org/abs/2103.10697
* CspNet (Cross-Stage Partial Networks) - https://arxiv.org/abs/1911.11929
* DeiT - https://arxiv.org/abs/2012.12877
* DeiT-III - https://arxiv.org/pdf/2204.07118.pdf
* DenseNet - https://arxiv.org/abs/1608.06993
* DLA - https://arxiv.org/abs/1707.06484
* DPN (Dual-Path Network) - https://arxiv.org/abs/1707.01629
* EdgeNeXt - https://arxiv.org/abs/2206.10589
* EfficientFormer - https://arxiv.org/abs/2206.01191
* EfficientNet (MBConvNet Family)
* EfficientNet NoisyStudent (B0-B7, L2) - https://arxiv.org/abs/1911.04252
* EfficientNet AdvProp (B0-B8) - https://arxiv.org/abs/1911.09665
* EfficientNet (B0-B7) - https://arxiv.org/abs/1905.11946
* EfficientNet-EdgeTPU (S, M, L) - https://ai.googleblog.com/2019/08/efficientnet-edgetpu-creating.html
* EfficientNet V2 - https://arxiv.org/abs/2104.00298
* FBNet-C - https://arxiv.org/abs/1812.03443
* MixNet - https://arxiv.org/abs/1907.09595
* MNASNet B1, A1 (Squeeze-Excite), and Small - https://arxiv.org/abs/1807.11626
* MobileNet-V2 - https://arxiv.org/abs/1801.04381
* Single-Path NAS - https://arxiv.org/abs/1904.02877
* TinyNet - https://arxiv.org/abs/2010.14819
* EfficientViT (MIT) - https://arxiv.org/abs/2205.14756
* EfficientViT (MSRA) - https://arxiv.org/abs/2305.07027
* EVA - https://arxiv.org/abs/2211.07636
* EVA-02 - https://arxiv.org/abs/2303.11331
* FastViT - https://arxiv.org/abs/2303.14189
* FlexiViT - https://arxiv.org/abs/2212.08013
* FocalNet (Focal Modulation Networks) - https://arxiv.org/abs/2203.11926
* GCViT (Global Context Vision Transformer) - https://arxiv.org/abs/2206.09959
* GhostNet - https://arxiv.org/abs/1911.11907
* GhostNet-V2 - https://arxiv.org/abs/2211.12905
* gMLP - https://arxiv.org/abs/2105.08050
* GPU-Efficient Networks - https://arxiv.org/abs/2006.14090
* Halo Nets - https://arxiv.org/abs/2103.12731
* HRNet - https://arxiv.org/abs/1908.07919
* InceptionNeXt - https://arxiv.org/abs/2303.16900
* Inception-V3 - https://arxiv.org/abs/1512.00567
* Inception-ResNet-V2 and Inception-V4 - https://arxiv.org/abs/1602.07261
* Lambda Networks - https://arxiv.org/abs/2102.08602
* LeViT (Vision Transformer in ConvNet's Clothing) - https://arxiv.org/abs/2104.01136
* MaxViT (Multi-Axis Vision Transformer) - https://arxiv.org/abs/2204.01697
* MetaFormer (PoolFormer-v2, ConvFormer, CAFormer) - https://arxiv.org/abs/2210.13452
* MLP-Mixer - https://arxiv.org/abs/2105.01601
* MobileNet-V3 (MBConvNet w/ Efficient Head) - https://arxiv.org/abs/1905.02244
* FBNet-V3 - https://arxiv.org/abs/2006.02049
* HardCoRe-NAS - https://arxiv.org/abs/2102.11646
* LCNet - https://arxiv.org/abs/2109.15099
* MobileOne - https://arxiv.org/abs/2206.04040
* MobileViT - https://arxiv.org/abs/2110.02178
* MobileViT-V2 - https://arxiv.org/abs/2206.02680
* MViT-V2 (Improved Multiscale Vision Transformer) - https://arxiv.org/abs/2112.01526
* NASNet-A - https://arxiv.org/abs/1707.07012
* NesT - https://arxiv.org/abs/2105.12723
* NFNet-F - https://arxiv.org/abs/2102.06171
* NF-RegNet / NF-ResNet - https://arxiv.org/abs/2101.08692
* PNasNet - https://arxiv.org/abs/1712.00559
* PoolFormer (MetaFormer) - https://arxiv.org/abs/2111.11418
* Pooling-based Vision Transformer (PiT) - https://arxiv.org/abs/2103.16302
* PVT-V2 (Improved Pyramid Vision Transformer) - https://arxiv.org/abs/2106.13797
* RegNet - https://arxiv.org/abs/2003.13678
* RegNetZ - https://arxiv.org/abs/2103.06877
* RepVGG - https://arxiv.org/abs/2101.03697
* RepGhostNet - https://arxiv.org/abs/2211.06088
* RepViT - https://arxiv.org/abs/2307.09283
* ResMLP - https://arxiv.org/abs/2105.03404
* ResNet/ResNeXt
* ResNet (v1b/v1.5) - https://arxiv.org/abs/1512.03385
* ResNeXt - https://arxiv.org/abs/1611.05431
* 'Bag of Tricks' / Gluon C, D, E, S variations - https://arxiv.org/abs/1812.01187
* Weakly-supervised (WSL) Instagram pretrained / ImageNet tuned ResNeXt101 - https://arxiv.org/abs/1805.00932
* Semi-supervised (SSL) / Semi-weakly Supervised (SWSL) ResNet/ResNeXts - https://arxiv.org/abs/1905.00546
* ECA-Net (ECAResNet) - https://arxiv.org/abs/1910.03151v4
* Squeeze-and-Excitation Networks (SEResNet) - https://arxiv.org/abs/1709.01507
* ResNet-RS - https://arxiv.org/abs/2103.07579
* Res2Net - https://arxiv.org/abs/1904.01169
* ResNeSt - https://arxiv.org/abs/2004.08955
* ReXNet - https://arxiv.org/abs/2007.00992
* SelecSLS - https://arxiv.org/abs/1907.00837
* Selective Kernel Networks - https://arxiv.org/abs/1903.06586
* Sequencer2D - https://arxiv.org/abs/2205.01972
* Swin S3 (AutoFormerV2) - https://arxiv.org/abs/2111.14725
* Swin Transformer - https://arxiv.org/abs/2103.14030
* Swin Transformer V2 - https://arxiv.org/abs/2111.09883
* Transformer-iN-Transformer (TNT) - https://arxiv.org/abs/2103.00112
* TResNet - https://arxiv.org/abs/2003.13630
* Twins (Spatial Attention in Vision Transformers) - https://arxiv.org/pdf/2104.13840.pdf
* Visformer - https://arxiv.org/abs/2104.12533
* Vision Transformer - https://arxiv.org/abs/2010.11929
* VOLO (Vision Outlooker) - https://arxiv.org/abs/2106.13112
* VovNet V2 and V1 - https://arxiv.org/abs/1911.06667
* Xception - https://arxiv.org/abs/1610.02357
* Xception (Modified Aligned, Gluon) - https://arxiv.org/abs/1802.02611
* Xception (Modified Aligned, TF) - https://arxiv.org/abs/1802.02611
* XCiT (Cross-Covariance Image Transformers) - https://arxiv.org/abs/2106.09681
## Features
Several (less common) features that I often utilize in my projects are included. Many of their additions are the reason why I maintain my own set of models, instead of using others' via PIP:
* All models have a common default configuration interface and API for
* accessing/changing the classifier - `get_classifier` and `reset_classifier`
* doing a forward pass on just the features - `forward_features` (see [documentation](https://huggingface.co/docs/timm/feature_extraction))
* these makes it easy to write consistent network wrappers that work with any of the models
* All models support multi-scale feature map extraction (feature pyramids) via create_model (see [documentation](https://huggingface.co/docs/timm/feature_extraction))
* `create_model(name, features_only=True, out_indices=..., output_stride=...)`
* `out_indices` creation arg specifies which feature maps to return, these indices are 0 based and generally correspond to the `C(i + 1)` feature level.
* `output_stride` creation arg controls output stride of the network by using dilated convolutions. Most networks are stride 32 by default. Not all networks support this.
* feature map channel counts, reduction level (stride) can be queried AFTER model creation via the `.feature_info` member
* All models have a consistent pretrained weight loader that adapts last linear if necessary, and from 3 to 1 channel input if desired
* High performance [reference training, validation, and inference scripts](https://huggingface.co/docs/timm/training_script) that work in several process/GPU modes:
* NVIDIA DDP w/ a single GPU per process, multiple processes with APEX present (AMP mixed-precision optional)
* PyTorch DistributedDataParallel w/ multi-gpu, single process (AMP disabled as it crashes when enabled)
* PyTorch w/ single GPU single process (AMP optional)
* A dynamic global pool implementation that allows selecting from average pooling, max pooling, average + max, or concat([average, max]) at model creation. All global pooling is adaptive average by default and compatible with pretrained weights.
* A 'Test Time Pool' wrapper that can wrap any of the included models and usually provides improved performance doing inference with input images larger than the training size. Idea adapted from original DPN implementation when I ported (https://github.com/cypw/DPNs)
* Learning rate schedulers
* Ideas adopted from
* [AllenNLP schedulers](https://github.com/allenai/allennlp/tree/master/allennlp/training/learning_rate_schedulers)
* [FAIRseq lr_scheduler](https://github.com/pytorch/fairseq/tree/master/fairseq/optim/lr_scheduler)
* SGDR: Stochastic Gradient Descent with Warm Restarts (https://arxiv.org/abs/1608.03983)
* Schedulers include `step`, `cosine` w/ restarts, `tanh` w/ restarts, `plateau`
* Optimizers:
* `rmsprop_tf` adapted from PyTorch RMSProp by myself. Reproduces much improved Tensorflow RMSProp behaviour.
* `radam` by [Liyuan Liu](https://github.com/LiyuanLucasLiu/RAdam) (https://arxiv.org/abs/1908.03265)
* `novograd` by [Masashi Kimura](https://github.com/convergence-lab/novograd) (https://arxiv.org/abs/1905.11286)
* `lookahead` adapted from impl by [Liam](https://github.com/alphadl/lookahead.pytorch) (https://arxiv.org/abs/1907.08610)
* `fused<name>` optimizers by name with [NVIDIA Apex](https://github.com/NVIDIA/apex/tree/master/apex/optimizers) installed
* `adamp` and `sgdp` by [Naver ClovAI](https://github.com/clovaai) (https://arxiv.org/abs/2006.08217)
* `adafactor` adapted from [FAIRSeq impl](https://github.com/pytorch/fairseq/blob/master/fairseq/optim/adafactor.py) (https://arxiv.org/abs/1804.04235)
* `adahessian` by [David Samuel](https://github.com/davda54/ada-hessian) (https://arxiv.org/abs/2006.00719)
* Random Erasing from [Zhun Zhong](https://github.com/zhunzhong07/Random-Erasing/blob/master/transforms.py) (https://arxiv.org/abs/1708.04896)
* Mixup (https://arxiv.org/abs/1710.09412)
* CutMix (https://arxiv.org/abs/1905.04899)
* AutoAugment (https://arxiv.org/abs/1805.09501) and RandAugment (https://arxiv.org/abs/1909.13719) ImageNet configurations modeled after impl for EfficientNet training (https://github.com/tensorflow/tpu/blob/master/models/official/efficientnet/autoaugment.py)
* AugMix w/ JSD loss (https://arxiv.org/abs/1912.02781), JSD w/ clean + augmented mixing support works with AutoAugment and RandAugment as well
* SplitBachNorm - allows splitting batch norm layers between clean and augmented (auxiliary batch norm) data
* DropPath aka "Stochastic Depth" (https://arxiv.org/abs/1603.09382)
* DropBlock (https://arxiv.org/abs/1810.12890)
* Blur Pooling (https://arxiv.org/abs/1904.11486)
* Space-to-Depth by [mrT23](https://github.com/mrT23/TResNet/blob/master/src/models/tresnet/layers/space_to_depth.py) (https://arxiv.org/abs/1801.04590) -- original paper?
* Adaptive Gradient Clipping (https://arxiv.org/abs/2102.06171, https://github.com/deepmind/deepmind-research/tree/master/nfnets)
* An extensive selection of channel and/or spatial attention modules:
* Bottleneck Transformer - https://arxiv.org/abs/2101.11605
* CBAM - https://arxiv.org/abs/1807.06521
* Effective Squeeze-Excitation (ESE) - https://arxiv.org/abs/1911.06667
* Efficient Channel Attention (ECA) - https://arxiv.org/abs/1910.03151
* Gather-Excite (GE) - https://arxiv.org/abs/1810.12348
* Global Context (GC) - https://arxiv.org/abs/1904.11492
* Halo - https://arxiv.org/abs/2103.12731
* Involution - https://arxiv.org/abs/2103.06255
* Lambda Layer - https://arxiv.org/abs/2102.08602
* Non-Local (NL) - https://arxiv.org/abs/1711.07971
* Squeeze-and-Excitation (SE) - https://arxiv.org/abs/1709.01507
* Selective Kernel (SK) - (https://arxiv.org/abs/1903.06586
* Split (SPLAT) - https://arxiv.org/abs/2004.08955
* Shifted Window (SWIN) - https://arxiv.org/abs/2103.14030
## Results
Model validation results can be found in the [results tables](results/README.md)
## Getting Started (Documentation)
The official documentation can be found at https://huggingface.co/docs/hub/timm. Documentation contributions are welcome.
[Getting Started with PyTorch Image Models (timm): A Practitioner’s Guide](https://towardsdatascience.com/getting-started-with-pytorch-image-models-timm-a-practitioners-guide-4e77b4bf9055) by [Chris Hughes](https://github.com/Chris-hughes10) is an extensive blog post covering many aspects of `timm` in detail.
[timmdocs](http://timm.fast.ai/) is an alternate set of documentation for `timm`. A big thanks to [Aman Arora](https://github.com/amaarora) for his efforts creating timmdocs.
[paperswithcode](https://paperswithcode.com/lib/timm) is a good resource for browsing the models within `timm`.
## Train, Validation, Inference Scripts
The root folder of the repository contains reference train, validation, and inference scripts that work with the included models and other features of this repository. They are adaptable for other datasets and use cases with a little hacking. See [documentation](https://huggingface.co/docs/timm/training_script).
## Awesome PyTorch Resources
One of the greatest assets of PyTorch is the community and their contributions. A few of my favourite resources that pair well with the models and components here are listed below.
### Object Detection, Instance and Semantic Segmentation
* Detectron2 - https://github.com/facebookresearch/detectron2
* Segmentation Models (Semantic) - https://github.com/qubvel/segmentation_models.pytorch
* EfficientDet (Obj Det, Semantic soon) - https://github.com/rwightman/efficientdet-pytorch
### Computer Vision / Image Augmentation
* Albumentations - https://github.com/albumentations-team/albumentations
* Kornia - https://github.com/kornia/kornia
### Knowledge Distillation
* RepDistiller - https://github.com/HobbitLong/RepDistiller
* torchdistill - https://github.com/yoshitomo-matsubara/torchdistill
### Metric Learning
* PyTorch Metric Learning - https://github.com/KevinMusgrave/pytorch-metric-learning
### Training / Frameworks
* fastai - https://github.com/fastai/fastai
## Licenses
### Code
The code here is licensed Apache 2.0. I've taken care to make sure any third party code included or adapted has compatible (permissive) licenses such as MIT, BSD, etc. I've made an effort to avoid any GPL / LGPL conflicts. That said, it is your responsibility to ensure you comply with licenses here and conditions of any dependent licenses. Where applicable, I've linked the sources/references for various components in docstrings. If you think I've missed anything please create an issue.
### Pretrained Weights
So far all of the pretrained weights available here are pretrained on ImageNet with a select few that have some additional pretraining (see extra note below). ImageNet was released for non-commercial research purposes only (https://image-net.org/download). It's not clear what the implications of that are for the use of pretrained weights from that dataset. Any models I have trained with ImageNet are done for research purposes and one should assume that the original dataset license applies to the weights. It's best to seek legal advice if you intend to use the pretrained weights in a commercial product.
#### Pretrained on more than ImageNet
Several weights included or references here were pretrained with proprietary datasets that I do not have access to. These include the Facebook WSL, SSL, SWSL ResNe(Xt) and the Google Noisy Student EfficientNet models. The Facebook models have an explicit non-commercial license (CC-BY-NC 4.0, https://github.com/facebookresearch/semi-supervised-ImageNet1K-models, https://github.com/facebookresearch/WSL-Images). The Google models do not appear to have any restriction beyond the Apache 2.0 license (and ImageNet concerns). In either case, you should contact Facebook or Google with any questions.
## Citing
### BibTeX
```bibtex
@misc{rw2019timm,
author = {Ross Wightman},
title = {PyTorch Image Models},
year = {2019},
publisher = {GitHub},
journal = {GitHub repository},
doi = {10.5281/zenodo.4414861},
howpublished = {\url{https://github.com/rwightman/pytorch-image-models}}
}
```
### Latest DOI
[](https://zenodo.org/badge/latestdoi/168799526)
| huggingface/pytorch-image-models/blob/main/README.md |
!--Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# LUKE
## Overview
The LUKE model was proposed in [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda and Yuji Matsumoto.
It is based on RoBERTa and adds entity embeddings as well as an entity-aware self-attention mechanism, which helps
improve performance on various downstream tasks involving reasoning about entities such as named entity recognition,
extractive and cloze-style question answering, entity typing, and relation classification.
The abstract from the paper is the following:
*Entity representations are useful in natural language tasks involving entities. In this paper, we propose new
pretrained contextualized representations of words and entities based on the bidirectional transformer. The proposed
model treats words and entities in a given text as independent tokens, and outputs contextualized representations of
them. Our model is trained using a new pretraining task based on the masked language model of BERT. The task involves
predicting randomly masked words and entities in a large entity-annotated corpus retrieved from Wikipedia. We also
propose an entity-aware self-attention mechanism that is an extension of the self-attention mechanism of the
transformer, and considers the types of tokens (words or entities) when computing attention scores. The proposed model
achieves impressive empirical performance on a wide range of entity-related tasks. In particular, it obtains
state-of-the-art results on five well-known datasets: Open Entity (entity typing), TACRED (relation classification),
CoNLL-2003 (named entity recognition), ReCoRD (cloze-style question answering), and SQuAD 1.1 (extractive question
answering).*
This model was contributed by [ikuyamada](https://huggingface.co/ikuyamada) and [nielsr](https://huggingface.co/nielsr). The original code can be found [here](https://github.com/studio-ousia/luke).
## Usage tips
- This implementation is the same as [`RobertaModel`] with the addition of entity embeddings as well
as an entity-aware self-attention mechanism, which improves performance on tasks involving reasoning about entities.
- LUKE treats entities as input tokens; therefore, it takes `entity_ids`, `entity_attention_mask`,
`entity_token_type_ids` and `entity_position_ids` as extra input. You can obtain those using
[`LukeTokenizer`].
- [`LukeTokenizer`] takes `entities` and `entity_spans` (character-based start and end
positions of the entities in the input text) as extra input. `entities` typically consist of [MASK] entities or
Wikipedia entities. The brief description when inputting these entities are as follows:
- *Inputting [MASK] entities to compute entity representations*: The [MASK] entity is used to mask entities to be
predicted during pretraining. When LUKE receives the [MASK] entity, it tries to predict the original entity by
gathering the information about the entity from the input text. Therefore, the [MASK] entity can be used to address
downstream tasks requiring the information of entities in text such as entity typing, relation classification, and
named entity recognition.
- *Inputting Wikipedia entities to compute knowledge-enhanced token representations*: LUKE learns rich information
(or knowledge) about Wikipedia entities during pretraining and stores the information in its entity embedding. By
using Wikipedia entities as input tokens, LUKE outputs token representations enriched by the information stored in
the embeddings of these entities. This is particularly effective for tasks requiring real-world knowledge, such as
question answering.
- There are three head models for the former use case:
- [`LukeForEntityClassification`], for tasks to classify a single entity in an input text such as
entity typing, e.g. the [Open Entity dataset](https://www.cs.utexas.edu/~eunsol/html_pages/open_entity.html).
This model places a linear head on top of the output entity representation.
- [`LukeForEntityPairClassification`], for tasks to classify the relationship between two entities
such as relation classification, e.g. the [TACRED dataset](https://nlp.stanford.edu/projects/tacred/). This
model places a linear head on top of the concatenated output representation of the pair of given entities.
- [`LukeForEntitySpanClassification`], for tasks to classify the sequence of entity spans, such as
named entity recognition (NER). This model places a linear head on top of the output entity representations. You
can address NER using this model by inputting all possible entity spans in the text to the model.
[`LukeTokenizer`] has a `task` argument, which enables you to easily create an input to these
head models by specifying `task="entity_classification"`, `task="entity_pair_classification"`, or
`task="entity_span_classification"`. Please refer to the example code of each head models.
Usage example:
```python
>>> from transformers import LukeTokenizer, LukeModel, LukeForEntityPairClassification
>>> model = LukeModel.from_pretrained("studio-ousia/luke-base")
>>> tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-base")
# Example 1: Computing the contextualized entity representation corresponding to the entity mention "Beyoncé"
>>> text = "Beyoncé lives in Los Angeles."
>>> entity_spans = [(0, 7)] # character-based entity span corresponding to "Beyoncé"
>>> inputs = tokenizer(text, entity_spans=entity_spans, add_prefix_space=True, return_tensors="pt")
>>> outputs = model(**inputs)
>>> word_last_hidden_state = outputs.last_hidden_state
>>> entity_last_hidden_state = outputs.entity_last_hidden_state
# Example 2: Inputting Wikipedia entities to obtain enriched contextualized representations
>>> entities = [
... "Beyoncé",
... "Los Angeles",
... ] # Wikipedia entity titles corresponding to the entity mentions "Beyoncé" and "Los Angeles"
>>> entity_spans = [(0, 7), (17, 28)] # character-based entity spans corresponding to "Beyoncé" and "Los Angeles"
>>> inputs = tokenizer(text, entities=entities, entity_spans=entity_spans, add_prefix_space=True, return_tensors="pt")
>>> outputs = model(**inputs)
>>> word_last_hidden_state = outputs.last_hidden_state
>>> entity_last_hidden_state = outputs.entity_last_hidden_state
# Example 3: Classifying the relationship between two entities using LukeForEntityPairClassification head model
>>> model = LukeForEntityPairClassification.from_pretrained("studio-ousia/luke-large-finetuned-tacred")
>>> tokenizer = LukeTokenizer.from_pretrained("studio-ousia/luke-large-finetuned-tacred")
>>> entity_spans = [(0, 7), (17, 28)] # character-based entity spans corresponding to "Beyoncé" and "Los Angeles"
>>> inputs = tokenizer(text, entity_spans=entity_spans, return_tensors="pt")
>>> outputs = model(**inputs)
>>> logits = outputs.logits
>>> predicted_class_idx = int(logits[0].argmax())
>>> print("Predicted class:", model.config.id2label[predicted_class_idx])
```
## Resources
- [A demo notebook on how to fine-tune [`LukeForEntityPairClassification`] for relation classification](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/LUKE)
- [Notebooks showcasing how you to reproduce the results as reported in the paper with the HuggingFace implementation of LUKE](https://github.com/studio-ousia/luke/tree/master/notebooks)
- [Text classification task guide](../tasks/sequence_classification)
- [Token classification task guide](../tasks/token_classification)
- [Question answering task guide](../tasks/question_answering)
- [Masked language modeling task guide](../tasks/masked_language_modeling)
- [Multiple choice task guide](../tasks/multiple_choice)
## LukeConfig
[[autodoc]] LukeConfig
## LukeTokenizer
[[autodoc]] LukeTokenizer
- __call__
- save_vocabulary
## LukeModel
[[autodoc]] LukeModel
- forward
## LukeForMaskedLM
[[autodoc]] LukeForMaskedLM
- forward
## LukeForEntityClassification
[[autodoc]] LukeForEntityClassification
- forward
## LukeForEntityPairClassification
[[autodoc]] LukeForEntityPairClassification
- forward
## LukeForEntitySpanClassification
[[autodoc]] LukeForEntitySpanClassification
- forward
## LukeForSequenceClassification
[[autodoc]] LukeForSequenceClassification
- forward
## LukeForMultipleChoice
[[autodoc]] LukeForMultipleChoice
- forward
## LukeForTokenClassification
[[autodoc]] LukeForTokenClassification
- forward
## LukeForQuestionAnswering
[[autodoc]] LukeForQuestionAnswering
- forward
| huggingface/transformers/blob/main/docs/source/en/model_doc/luke.md |
Gradio Demo: image_selections
```
!pip install -q gradio
```
```
import gradio as gr
import numpy as np
with gr.Blocks() as demo:
tolerance = gr.Slider(label="Tolerance", info="How different colors can be in a segment.", minimum=0, maximum=256*3, value=50)
with gr.Row():
input_img = gr.Image(label="Input")
output_img = gr.Image(label="Selected Segment")
def get_select_coords(img, tolerance, evt: gr.SelectData):
visited_pixels = set()
pixels_in_queue = set()
pixels_in_segment = set()
start_pixel = img[evt.index[1], evt.index[0]]
pixels_in_queue.add((evt.index[1], evt.index[0]))
while len(pixels_in_queue) > 0:
pixel = pixels_in_queue.pop()
visited_pixels.add(pixel)
neighbors = []
if pixel[0] > 0:
neighbors.append((pixel[0] - 1, pixel[1]))
if pixel[0] < img.shape[0] - 1:
neighbors.append((pixel[0] + 1, pixel[1]))
if pixel[1] > 0:
neighbors.append((pixel[0], pixel[1] - 1))
if pixel[1] < img.shape[1] - 1:
neighbors.append((pixel[0], pixel[1] + 1))
for neighbor in neighbors:
if neighbor in visited_pixels:
continue
neighbor_pixel = img[neighbor[0], neighbor[1]]
if np.abs(neighbor_pixel - start_pixel).sum() < tolerance:
pixels_in_queue.add(neighbor)
pixels_in_segment.add(neighbor)
out = img.copy() * 0.2
out = out.astype(np.uint8)
for pixel in pixels_in_segment:
out[pixel[0], pixel[1]] = img[pixel[0], pixel[1]]
return out
input_img.select(get_select_coords, [input_img, tolerance], output_img)
if __name__ == "__main__":
demo.launch()
```
| gradio-app/gradio/blob/main/demo/image_selections/run.ipynb |
!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Image classification
[[open-in-colab]]
<Youtube id="tjAIM7BOYhw"/>
Image classification assigns a label or class to an image. Unlike text or audio classification, the inputs are the
pixel values that comprise an image. There are many applications for image classification, such as detecting damage
after a natural disaster, monitoring crop health, or helping screen medical images for signs of disease.
This guide illustrates how to:
1. Fine-tune [ViT](model_doc/vit) on the [Food-101](https://huggingface.co/datasets/food101) dataset to classify a food item in an image.
2. Use your fine-tuned model for inference.
<Tip>
The task illustrated in this tutorial is supported by the following model architectures:
<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
[BEiT](../model_doc/beit), [BiT](../model_doc/bit), [ConvNeXT](../model_doc/convnext), [ConvNeXTV2](../model_doc/convnextv2), [CvT](../model_doc/cvt), [Data2VecVision](../model_doc/data2vec-vision), [DeiT](../model_doc/deit), [DiNAT](../model_doc/dinat), [DINOv2](../model_doc/dinov2), [EfficientFormer](../model_doc/efficientformer), [EfficientNet](../model_doc/efficientnet), [FocalNet](../model_doc/focalnet), [ImageGPT](../model_doc/imagegpt), [LeViT](../model_doc/levit), [MobileNetV1](../model_doc/mobilenet_v1), [MobileNetV2](../model_doc/mobilenet_v2), [MobileViT](../model_doc/mobilevit), [MobileViTV2](../model_doc/mobilevitv2), [NAT](../model_doc/nat), [Perceiver](../model_doc/perceiver), [PoolFormer](../model_doc/poolformer), [PVT](../model_doc/pvt), [RegNet](../model_doc/regnet), [ResNet](../model_doc/resnet), [SegFormer](../model_doc/segformer), [SwiftFormer](../model_doc/swiftformer), [Swin Transformer](../model_doc/swin), [Swin Transformer V2](../model_doc/swinv2), [VAN](../model_doc/van), [ViT](../model_doc/vit), [ViT Hybrid](../model_doc/vit_hybrid), [ViTMSN](../model_doc/vit_msn)
<!--End of the generated tip-->
</Tip>
Before you begin, make sure you have all the necessary libraries installed:
```bash
pip install transformers datasets evaluate
```
We encourage you to log in to your Hugging Face account to upload and share your model with the community. When prompted, enter your token to log in:
```py
>>> from huggingface_hub import notebook_login
>>> notebook_login()
```
## Load Food-101 dataset
Start by loading a smaller subset of the Food-101 dataset from the 🤗 Datasets library. This will give you a chance to
experiment and make sure everything works before spending more time training on the full dataset.
```py
>>> from datasets import load_dataset
>>> food = load_dataset("food101", split="train[:5000]")
```
Split the dataset's `train` split into a train and test set with the [`~datasets.Dataset.train_test_split`] method:
```py
>>> food = food.train_test_split(test_size=0.2)
```
Then take a look at an example:
```py
>>> food["train"][0]
{'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=512x512 at 0x7F52AFC8AC50>,
'label': 79}
```
Each example in the dataset has two fields:
- `image`: a PIL image of the food item
- `label`: the label class of the food item
To make it easier for the model to get the label name from the label id, create a dictionary that maps the label name
to an integer and vice versa:
```py
>>> labels = food["train"].features["label"].names
>>> label2id, id2label = dict(), dict()
>>> for i, label in enumerate(labels):
... label2id[label] = str(i)
... id2label[str(i)] = label
```
Now you can convert the label id to a label name:
```py
>>> id2label[str(79)]
'prime_rib'
```
## Preprocess
The next step is to load a ViT image processor to process the image into a tensor:
```py
>>> from transformers import AutoImageProcessor
>>> checkpoint = "google/vit-base-patch16-224-in21k"
>>> image_processor = AutoImageProcessor.from_pretrained(checkpoint)
```
<frameworkcontent>
<pt>
Apply some image transformations to the images to make the model more robust against overfitting. Here you'll use torchvision's [`transforms`](https://pytorch.org/vision/stable/transforms.html) module, but you can also use any image library you like.
Crop a random part of the image, resize it, and normalize it with the image mean and standard deviation:
```py
>>> from torchvision.transforms import RandomResizedCrop, Compose, Normalize, ToTensor
>>> normalize = Normalize(mean=image_processor.image_mean, std=image_processor.image_std)
>>> size = (
... image_processor.size["shortest_edge"]
... if "shortest_edge" in image_processor.size
... else (image_processor.size["height"], image_processor.size["width"])
... )
>>> _transforms = Compose([RandomResizedCrop(size), ToTensor(), normalize])
```
Then create a preprocessing function to apply the transforms and return the `pixel_values` - the inputs to the model - of the image:
```py
>>> def transforms(examples):
... examples["pixel_values"] = [_transforms(img.convert("RGB")) for img in examples["image"]]
... del examples["image"]
... return examples
```
To apply the preprocessing function over the entire dataset, use 🤗 Datasets [`~datasets.Dataset.with_transform`] method. The transforms are applied on the fly when you load an element of the dataset:
```py
>>> food = food.with_transform(transforms)
```
Now create a batch of examples using [`DefaultDataCollator`]. Unlike other data collators in 🤗 Transformers, the `DefaultDataCollator` does not apply additional preprocessing such as padding.
```py
>>> from transformers import DefaultDataCollator
>>> data_collator = DefaultDataCollator()
```
</pt>
</frameworkcontent>
<frameworkcontent>
<tf>
To avoid overfitting and to make the model more robust, add some data augmentation to the training part of the dataset.
Here we use Keras preprocessing layers to define the transformations for the training data (includes data augmentation),
and transformations for the validation data (only center cropping, resizing and normalizing). You can use `tf.image`or
any other library you prefer.
```py
>>> from tensorflow import keras
>>> from tensorflow.keras import layers
>>> size = (image_processor.size["height"], image_processor.size["width"])
>>> train_data_augmentation = keras.Sequential(
... [
... layers.RandomCrop(size[0], size[1]),
... layers.Rescaling(scale=1.0 / 127.5, offset=-1),
... layers.RandomFlip("horizontal"),
... layers.RandomRotation(factor=0.02),
... layers.RandomZoom(height_factor=0.2, width_factor=0.2),
... ],
... name="train_data_augmentation",
... )
>>> val_data_augmentation = keras.Sequential(
... [
... layers.CenterCrop(size[0], size[1]),
... layers.Rescaling(scale=1.0 / 127.5, offset=-1),
... ],
... name="val_data_augmentation",
... )
```
Next, create functions to apply appropriate transformations to a batch of images, instead of one image at a time.
```py
>>> import numpy as np
>>> import tensorflow as tf
>>> from PIL import Image
>>> def convert_to_tf_tensor(image: Image):
... np_image = np.array(image)
... tf_image = tf.convert_to_tensor(np_image)
... # `expand_dims()` is used to add a batch dimension since
... # the TF augmentation layers operates on batched inputs.
... return tf.expand_dims(tf_image, 0)
>>> def preprocess_train(example_batch):
... """Apply train_transforms across a batch."""
... images = [
... train_data_augmentation(convert_to_tf_tensor(image.convert("RGB"))) for image in example_batch["image"]
... ]
... example_batch["pixel_values"] = [tf.transpose(tf.squeeze(image)) for image in images]
... return example_batch
... def preprocess_val(example_batch):
... """Apply val_transforms across a batch."""
... images = [
... val_data_augmentation(convert_to_tf_tensor(image.convert("RGB"))) for image in example_batch["image"]
... ]
... example_batch["pixel_values"] = [tf.transpose(tf.squeeze(image)) for image in images]
... return example_batch
```
Use 🤗 Datasets [`~datasets.Dataset.set_transform`] to apply the transformations on the fly:
```py
food["train"].set_transform(preprocess_train)
food["test"].set_transform(preprocess_val)
```
As a final preprocessing step, create a batch of examples using `DefaultDataCollator`. Unlike other data collators in 🤗 Transformers, the
`DefaultDataCollator` does not apply additional preprocessing, such as padding.
```py
>>> from transformers import DefaultDataCollator
>>> data_collator = DefaultDataCollator(return_tensors="tf")
```
</tf>
</frameworkcontent>
## Evaluate
Including a metric during training is often helpful for evaluating your model's performance. You can quickly load an
evaluation method with the 🤗 [Evaluate](https://huggingface.co/docs/evaluate/index) library. For this task, load
the [accuracy](https://huggingface.co/spaces/evaluate-metric/accuracy) metric (see the 🤗 Evaluate [quick tour](https://huggingface.co/docs/evaluate/a_quick_tour) to learn more about how to load and compute a metric):
```py
>>> import evaluate
>>> accuracy = evaluate.load("accuracy")
```
Then create a function that passes your predictions and labels to [`~evaluate.EvaluationModule.compute`] to calculate the accuracy:
```py
>>> import numpy as np
>>> def compute_metrics(eval_pred):
... predictions, labels = eval_pred
... predictions = np.argmax(predictions, axis=1)
... return accuracy.compute(predictions=predictions, references=labels)
```
Your `compute_metrics` function is ready to go now, and you'll return to it when you set up your training.
## Train
<frameworkcontent>
<pt>
<Tip>
If you aren't familiar with finetuning a model with the [`Trainer`], take a look at the basic tutorial [here](../training#train-with-pytorch-trainer)!
</Tip>
You're ready to start training your model now! Load ViT with [`AutoModelForImageClassification`]. Specify the number of labels along with the number of expected labels, and the label mappings:
```py
>>> from transformers import AutoModelForImageClassification, TrainingArguments, Trainer
>>> model = AutoModelForImageClassification.from_pretrained(
... checkpoint,
... num_labels=len(labels),
... id2label=id2label,
... label2id=label2id,
... )
```
At this point, only three steps remain:
1. Define your training hyperparameters in [`TrainingArguments`]. It is important you don't remove unused columns because that'll drop the `image` column. Without the `image` column, you can't create `pixel_values`. Set `remove_unused_columns=False` to prevent this behavior! The only other required parameter is `output_dir` which specifies where to save your model. You'll push this model to the Hub by setting `push_to_hub=True` (you need to be signed in to Hugging Face to upload your model). At the end of each epoch, the [`Trainer`] will evaluate the accuracy and save the training checkpoint.
2. Pass the training arguments to [`Trainer`] along with the model, dataset, tokenizer, data collator, and `compute_metrics` function.
3. Call [`~Trainer.train`] to finetune your model.
```py
>>> training_args = TrainingArguments(
... output_dir="my_awesome_food_model",
... remove_unused_columns=False,
... evaluation_strategy="epoch",
... save_strategy="epoch",
... learning_rate=5e-5,
... per_device_train_batch_size=16,
... gradient_accumulation_steps=4,
... per_device_eval_batch_size=16,
... num_train_epochs=3,
... warmup_ratio=0.1,
... logging_steps=10,
... load_best_model_at_end=True,
... metric_for_best_model="accuracy",
... push_to_hub=True,
... )
>>> trainer = Trainer(
... model=model,
... args=training_args,
... data_collator=data_collator,
... train_dataset=food["train"],
... eval_dataset=food["test"],
... tokenizer=image_processor,
... compute_metrics=compute_metrics,
... )
>>> trainer.train()
```
Once training is completed, share your model to the Hub with the [`~transformers.Trainer.push_to_hub`] method so everyone can use your model:
```py
>>> trainer.push_to_hub()
```
</pt>
</frameworkcontent>
<frameworkcontent>
<tf>
<Tip>
If you are unfamiliar with fine-tuning a model with Keras, check out the [basic tutorial](./training#train-a-tensorflow-model-with-keras) first!
</Tip>
To fine-tune a model in TensorFlow, follow these steps:
1. Define the training hyperparameters, and set up an optimizer and a learning rate schedule.
2. Instantiate a pre-trained model.
3. Convert a 🤗 Dataset to a `tf.data.Dataset`.
4. Compile your model.
5. Add callbacks and use the `fit()` method to run the training.
6. Upload your model to 🤗 Hub to share with the community.
Start by defining the hyperparameters, optimizer and learning rate schedule:
```py
>>> from transformers import create_optimizer
>>> batch_size = 16
>>> num_epochs = 5
>>> num_train_steps = len(food["train"]) * num_epochs
>>> learning_rate = 3e-5
>>> weight_decay_rate = 0.01
>>> optimizer, lr_schedule = create_optimizer(
... init_lr=learning_rate,
... num_train_steps=num_train_steps,
... weight_decay_rate=weight_decay_rate,
... num_warmup_steps=0,
... )
```
Then, load ViT with [`TFAutoModelForImageClassification`] along with the label mappings:
```py
>>> from transformers import TFAutoModelForImageClassification
>>> model = TFAutoModelForImageClassification.from_pretrained(
... checkpoint,
... id2label=id2label,
... label2id=label2id,
... )
```
Convert your datasets to the `tf.data.Dataset` format using the [`~datasets.Dataset.to_tf_dataset`] and your `data_collator`:
```py
>>> # converting our train dataset to tf.data.Dataset
>>> tf_train_dataset = food["train"].to_tf_dataset(
... columns="pixel_values", label_cols="label", shuffle=True, batch_size=batch_size, collate_fn=data_collator
... )
>>> # converting our test dataset to tf.data.Dataset
>>> tf_eval_dataset = food["test"].to_tf_dataset(
... columns="pixel_values", label_cols="label", shuffle=True, batch_size=batch_size, collate_fn=data_collator
... )
```
Configure the model for training with `compile()`:
```py
>>> from tensorflow.keras.losses import SparseCategoricalCrossentropy
>>> loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
>>> model.compile(optimizer=optimizer, loss=loss)
```
To compute the accuracy from the predictions and push your model to the 🤗 Hub, use [Keras callbacks](../main_classes/keras_callbacks).
Pass your `compute_metrics` function to [KerasMetricCallback](../main_classes/keras_callbacks#transformers.KerasMetricCallback),
and use the [PushToHubCallback](../main_classes/keras_callbacks#transformers.PushToHubCallback) to upload the model:
```py
>>> from transformers.keras_callbacks import KerasMetricCallback, PushToHubCallback
>>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_eval_dataset)
>>> push_to_hub_callback = PushToHubCallback(
... output_dir="food_classifier",
... tokenizer=image_processor,
... save_strategy="no",
... )
>>> callbacks = [metric_callback, push_to_hub_callback]
```
Finally, you are ready to train your model! Call `fit()` with your training and validation datasets, the number of epochs,
and your callbacks to fine-tune the model:
```py
>>> model.fit(tf_train_dataset, validation_data=tf_eval_dataset, epochs=num_epochs, callbacks=callbacks)
Epoch 1/5
250/250 [==============================] - 313s 1s/step - loss: 2.5623 - val_loss: 1.4161 - accuracy: 0.9290
Epoch 2/5
250/250 [==============================] - 265s 1s/step - loss: 0.9181 - val_loss: 0.6808 - accuracy: 0.9690
Epoch 3/5
250/250 [==============================] - 252s 1s/step - loss: 0.3910 - val_loss: 0.4303 - accuracy: 0.9820
Epoch 4/5
250/250 [==============================] - 251s 1s/step - loss: 0.2028 - val_loss: 0.3191 - accuracy: 0.9900
Epoch 5/5
250/250 [==============================] - 238s 949ms/step - loss: 0.1232 - val_loss: 0.3259 - accuracy: 0.9890
```
Congratulations! You have fine-tuned your model and shared it on the 🤗 Hub. You can now use it for inference!
</tf>
</frameworkcontent>
<Tip>
For a more in-depth example of how to finetune a model for image classification, take a look at the corresponding [PyTorch notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb).
</Tip>
## Inference
Great, now that you've fine-tuned a model, you can use it for inference!
Load an image you'd like to run inference on:
```py
>>> ds = load_dataset("food101", split="validation[:10]")
>>> image = ds["image"][0]
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png" alt="image of beignets"/>
</div>
The simplest way to try out your finetuned model for inference is to use it in a [`pipeline`]. Instantiate a `pipeline` for image classification with your model, and pass your image to it:
```py
>>> from transformers import pipeline
>>> classifier = pipeline("image-classification", model="my_awesome_food_model")
>>> classifier(image)
[{'score': 0.31856709718704224, 'label': 'beignets'},
{'score': 0.015232225880026817, 'label': 'bruschetta'},
{'score': 0.01519392803311348, 'label': 'chicken_wings'},
{'score': 0.013022331520915031, 'label': 'pork_chop'},
{'score': 0.012728818692266941, 'label': 'prime_rib'}]
```
You can also manually replicate the results of the `pipeline` if you'd like:
<frameworkcontent>
<pt>
Load an image processor to preprocess the image and return the `input` as PyTorch tensors:
```py
>>> from transformers import AutoImageProcessor
>>> import torch
>>> image_processor = AutoImageProcessor.from_pretrained("my_awesome_food_model")
>>> inputs = image_processor(image, return_tensors="pt")
```
Pass your inputs to the model and return the logits:
```py
>>> from transformers import AutoModelForImageClassification
>>> model = AutoModelForImageClassification.from_pretrained("my_awesome_food_model")
>>> with torch.no_grad():
... logits = model(**inputs).logits
```
Get the predicted label with the highest probability, and use the model's `id2label` mapping to convert it to a label:
```py
>>> predicted_label = logits.argmax(-1).item()
>>> model.config.id2label[predicted_label]
'beignets'
```
</pt>
</frameworkcontent>
<frameworkcontent>
<tf>
Load an image processor to preprocess the image and return the `input` as TensorFlow tensors:
```py
>>> from transformers import AutoImageProcessor
>>> image_processor = AutoImageProcessor.from_pretrained("MariaK/food_classifier")
>>> inputs = image_processor(image, return_tensors="tf")
```
Pass your inputs to the model and return the logits:
```py
>>> from transformers import TFAutoModelForImageClassification
>>> model = TFAutoModelForImageClassification.from_pretrained("MariaK/food_classifier")
>>> logits = model(**inputs).logits
```
Get the predicted label with the highest probability, and use the model's `id2label` mapping to convert it to a label:
```py
>>> predicted_class_id = int(tf.math.argmax(logits, axis=-1)[0])
>>> model.config.id2label[predicted_class_id]
'beignets'
```
</tf>
</frameworkcontent>
| huggingface/transformers/blob/main/docs/source/en/tasks/image_classification.md |
[DreamBooth](https://github.com/huggingface/diffusers/tree/main/examples/dreambooth) by [colossalai](https://github.com/hpcaitech/ColossalAI.git)
[DreamBooth](https://arxiv.org/abs/2208.12242) is a method to personalize text2image models like stable diffusion given just a few(3~5) images of a subject.
The `train_dreambooth_colossalai.py` script shows how to implement the training procedure and adapt it for stable diffusion.
By accommodating model data in CPU and GPU and moving the data to the computing device when necessary, [Gemini](https://www.colossalai.org/docs/advanced_tutorials/meet_gemini), the Heterogeneous Memory Manager of [Colossal-AI](https://github.com/hpcaitech/ColossalAI) can breakthrough the GPU memory wall by using GPU and CPU memory (composed of CPU DRAM or nvme SSD memory) together at the same time. Moreover, the model scale can be further improved by combining heterogeneous training with the other parallel approaches, such as data parallel, tensor parallel and pipeline parallel.
## Installing the dependencies
Before running the scripts, make sure to install the library's training dependencies:
```bash
pip install -r requirements.txt
```
## Install [ColossalAI](https://github.com/hpcaitech/ColossalAI.git)
**From PyPI**
```bash
pip install colossalai
```
**From source**
```bash
git clone https://github.com/hpcaitech/ColossalAI.git
cd ColossalAI
# install colossalai
pip install .
```
## Dataset for Teyvat BLIP captions
Dataset used to train [Teyvat characters text to image model](https://github.com/hpcaitech/ColossalAI/tree/main/examples/images/diffusion).
BLIP generated captions for characters images from [genshin-impact fandom wiki](https://genshin-impact.fandom.com/wiki/Character#Playable_Characters)and [biligame wiki for genshin impact](https://wiki.biligame.com/ys/%E8%A7%92%E8%89%B2).
For each row the dataset contains `image` and `text` keys. `image` is a varying size PIL png, and `text` is the accompanying text caption. Only a train split is provided.
The `text` include the tag `Teyvat`, `Name`,`Element`, `Weapon`, `Region`, `Model type`, and `Description`, the `Description` is captioned with the [pre-trained BLIP model](https://github.com/salesforce/BLIP).
## Training
The argument `placement` can be `cpu`, `auto`, `cuda`, with `cpu` the GPU RAM required can be minimized to 4GB but will deceleration, with `cuda` you can also reduce GPU memory by half but accelerated training， with `auto` a more balanced solution for speed and memory can be obtained。
**___Note: Change the `resolution` to 768 if you are using the [stable-diffusion-2](https://huggingface.co/stabilityai/stable-diffusion-2) 768x768 model.___**
```bash
export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="path-to-instance-images"
export OUTPUT_DIR="path-to-save-model"
torchrun --nproc_per_node 2 train_dreambooth_colossalai.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--output_dir=$OUTPUT_DIR \
--instance_prompt="a photo of sks dog" \
--resolution=512 \
--train_batch_size=1 \
--learning_rate=5e-6 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=400 \
--placement="cuda"
```
### Training with prior-preservation loss
Prior-preservation is used to avoid overfitting and language-drift. Refer to the paper to learn more about it. For prior-preservation we first generate images using the model with a class prompt and then use those during training along with our data.
According to the paper, it's recommended to generate `num_epochs * num_samples` images for prior-preservation. 200-300 works well for most cases. The `num_class_images` flag sets the number of images to generate with the class prompt. You can place existing images in `class_data_dir`, and the training script will generate any additional images so that `num_class_images` are present in `class_data_dir` during training time.
```bash
export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="path-to-instance-images"
export CLASS_DIR="path-to-class-images"
export OUTPUT_DIR="path-to-save-model"
torchrun --nproc_per_node 2 train_dreambooth_colossalai.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--class_data_dir=$CLASS_DIR \
--output_dir=$OUTPUT_DIR \
--with_prior_preservation --prior_loss_weight=1.0 \
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog" \
--resolution=512 \
--train_batch_size=1 \
--learning_rate=5e-6 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=800 \
--placement="cuda"
```
## Inference
Once you have trained a model using above command, the inference can be done simply using the `StableDiffusionPipeline`. Make sure to include the `identifier`(e.g. sks in above example) in your prompt.
```python
from diffusers import StableDiffusionPipeline
import torch
model_id = "path-to-save-model"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
prompt = "A photo of sks dog in a bucket"
image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
image.save("dog-bucket.png")
```
| huggingface/diffusers/blob/main/examples/research_projects/colossalai/README.md |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# DiffEdit
[[open-in-colab]]
Image editing typically requires providing a mask of the area to be edited. DiffEdit automatically generates the mask for you based on a text query, making it easier overall to create a mask without image editing software. The DiffEdit algorithm works in three steps:
1. the diffusion model denoises an image conditioned on some query text and reference text which produces different noise estimates for different areas of the image; the difference is used to infer a mask to identify which area of the image needs to be changed to match the query text
2. the input image is encoded into latent space with DDIM
3. the latents are decoded with the diffusion model conditioned on the text query, using the mask as a guide such that pixels outside the mask remain the same as in the input image
This guide will show you how to use DiffEdit to edit images without manually creating a mask.
Before you begin, make sure you have the following libraries installed:
```py
# uncomment to install the necessary libraries in Colab
#!pip install -q diffusers transformers accelerate
```
The [`StableDiffusionDiffEditPipeline`] requires an image mask and a set of partially inverted latents. The image mask is generated from the [`~StableDiffusionDiffEditPipeline.generate_mask`] function, and includes two parameters, `source_prompt` and `target_prompt`. These parameters determine what to edit in the image. For example, if you want to change a bowl of *fruits* to a bowl of *pears*, then:
```py
source_prompt = "a bowl of fruits"
target_prompt = "a bowl of pears"
```
The partially inverted latents are generated from the [`~StableDiffusionDiffEditPipeline.invert`] function, and it is generally a good idea to include a `prompt` or *caption* describing the image to help guide the inverse latent sampling process. The caption can often be your `source_prompt`, but feel free to experiment with other text descriptions!
Let's load the pipeline, scheduler, inverse scheduler, and enable some optimizations to reduce memory usage:
```py
import torch
from diffusers import DDIMScheduler, DDIMInverseScheduler, StableDiffusionDiffEditPipeline
pipeline = StableDiffusionDiffEditPipeline.from_pretrained(
"stabilityai/stable-diffusion-2-1",
torch_dtype=torch.float16,
safety_checker=None,
use_safetensors=True,
)
pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
pipeline.enable_model_cpu_offload()
pipeline.enable_vae_slicing()
```
Load the image to edit:
```py
from diffusers.utils import load_image, make_image_grid
img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
raw_image = load_image(img_url).resize((768, 768))
raw_image
```
Use the [`~StableDiffusionDiffEditPipeline.generate_mask`] function to generate the image mask. You'll need to pass it the `source_prompt` and `target_prompt` to specify what to edit in the image:
```py
from PIL import Image
source_prompt = "a bowl of fruits"
target_prompt = "a basket of pears"
mask_image = pipeline.generate_mask(
image=raw_image,
source_prompt=source_prompt,
target_prompt=target_prompt,
)
Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L").resize((768, 768))
```
Next, create the inverted latents and pass it a caption describing the image:
```py
inv_latents = pipeline.invert(prompt=source_prompt, image=raw_image).latents
```
Finally, pass the image mask and inverted latents to the pipeline. The `target_prompt` becomes the `prompt` now, and the `source_prompt` is used as the `negative_prompt`:
```py
output_image = pipeline(
prompt=target_prompt,
mask_image=mask_image,
image_latents=inv_latents,
negative_prompt=source_prompt,
).images[0]
mask_image = Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L").resize((768, 768))
make_image_grid([raw_image, mask_image, output_image], rows=1, cols=3)
```
<div class="flex gap-4">
<div>
<img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">original image</figcaption>
</div>
<div>
<img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/blob/main/assets/target.png?raw=true"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">edited image</figcaption>
</div>
</div>
## Generate source and target embeddings
The source and target embeddings can be automatically generated with the [Flan-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5) model instead of creating them manually.
Load the Flan-T5 model and tokenizer from the 🤗 Transformers library:
```py
import torch
from transformers import AutoTokenizer, T5ForConditionalGeneration
tokenizer = AutoTokenizer.from_pretrained("google/flan-t5-large")
model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-large", device_map="auto", torch_dtype=torch.float16)
```
Provide some initial text to prompt the model to generate the source and target prompts.
```py
source_concept = "bowl"
target_concept = "basket"
source_text = f"Provide a caption for images containing a {source_concept}. "
"The captions should be in English and should be no longer than 150 characters."
target_text = f"Provide a caption for images containing a {target_concept}. "
"The captions should be in English and should be no longer than 150 characters."
```
Next, create a utility function to generate the prompts:
```py
@torch.no_grad()
def generate_prompts(input_prompt):
input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids.to("cuda")
outputs = model.generate(
input_ids, temperature=0.8, num_return_sequences=16, do_sample=True, max_new_tokens=128, top_k=10
)
return tokenizer.batch_decode(outputs, skip_special_tokens=True)
source_prompts = generate_prompts(source_text)
target_prompts = generate_prompts(target_text)
print(source_prompts)
print(target_prompts)
```
<Tip>
Check out the [generation strategy](https://huggingface.co/docs/transformers/main/en/generation_strategies) guide if you're interested in learning more about strategies for generating different quality text.
</Tip>
Load the text encoder model used by the [`StableDiffusionDiffEditPipeline`] to encode the text. You'll use the text encoder to compute the text embeddings:
```py
import torch
from diffusers import StableDiffusionDiffEditPipeline
pipeline = StableDiffusionDiffEditPipeline.from_pretrained(
"stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16, use_safetensors=True
)
pipeline.enable_model_cpu_offload()
pipeline.enable_vae_slicing()
@torch.no_grad()
def embed_prompts(sentences, tokenizer, text_encoder, device="cuda"):
embeddings = []
for sent in sentences:
text_inputs = tokenizer(
sent,
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
prompt_embeds = text_encoder(text_input_ids.to(device), attention_mask=None)[0]
embeddings.append(prompt_embeds)
return torch.concatenate(embeddings, dim=0).mean(dim=0).unsqueeze(0)
source_embeds = embed_prompts(source_prompts, pipeline.tokenizer, pipeline.text_encoder)
target_embeds = embed_prompts(target_prompts, pipeline.tokenizer, pipeline.text_encoder)
```
Finally, pass the embeddings to the [`~StableDiffusionDiffEditPipeline.generate_mask`] and [`~StableDiffusionDiffEditPipeline.invert`] functions, and pipeline to generate the image:
```diff
from diffusers import DDIMInverseScheduler, DDIMScheduler
from diffusers.utils import load_image, make_image_grid
from PIL import Image
pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
raw_image = load_image(img_url).resize((768, 768))
mask_image = pipeline.generate_mask(
image=raw_image,
- source_prompt=source_prompt,
- target_prompt=target_prompt,
+ source_prompt_embeds=source_embeds,
+ target_prompt_embeds=target_embeds,
)
inv_latents = pipeline.invert(
- prompt=source_prompt,
+ prompt_embeds=source_embeds,
image=raw_image,
).latents
output_image = pipeline(
mask_image=mask_image,
image_latents=inv_latents,
- prompt=target_prompt,
- negative_prompt=source_prompt,
+ prompt_embeds=target_embeds,
+ negative_prompt_embeds=source_embeds,
).images[0]
mask_image = Image.fromarray((mask_image.squeeze()*255).astype("uint8"), "L")
make_image_grid([raw_image, mask_image, output_image], rows=1, cols=3)
```
## Generate a caption for inversion
While you can use the `source_prompt` as a caption to help generate the partially inverted latents, you can also use the [BLIP](https://huggingface.co/docs/transformers/model_doc/blip) model to automatically generate a caption.
Load the BLIP model and processor from the 🤗 Transformers library:
```py
import torch
from transformers import BlipForConditionalGeneration, BlipProcessor
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base", torch_dtype=torch.float16, low_cpu_mem_usage=True)
```
Create a utility function to generate a caption from the input image:
```py
@torch.no_grad()
def generate_caption(images, caption_generator, caption_processor):
text = "a photograph of"
inputs = caption_processor(images, text, return_tensors="pt").to(device="cuda", dtype=caption_generator.dtype)
caption_generator.to("cuda")
outputs = caption_generator.generate(**inputs, max_new_tokens=128)
# offload caption generator
caption_generator.to("cpu")
caption = caption_processor.batch_decode(outputs, skip_special_tokens=True)[0]
return caption
```
Load an input image and generate a caption for it using the `generate_caption` function:
```py
from diffusers.utils import load_image
img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
raw_image = load_image(img_url).resize((768, 768))
caption = generate_caption(raw_image, model, processor)
```
<div class="flex justify-center">
<figure>
<img class="rounded-xl" src="https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"/>
<figcaption class="text-center">generated caption: "a photograph of a bowl of fruit on a table"</figcaption>
</figure>
</div>
Now you can drop the caption into the [`~StableDiffusionDiffEditPipeline.invert`] function to generate the partially inverted latents!
| huggingface/diffusers/blob/main/docs/source/en/using-diffusers/diffedit.md |
!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Whisper
## Overview
The Whisper model was proposed in [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf) by Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever.
The abstract from the paper is the following:
*We study the capabilities of speech processing systems trained simply to predict large amounts of transcripts of audio on the internet. When scaled to 680,000 hours of multilingual and multitask supervision, the resulting models generalize well to standard benchmarks and are often competitive with prior fully supervised results but in a zeroshot transfer setting without the need for any finetuning. When compared to humans, the models approach their accuracy and robustness. We are releasing models and inference code to serve as a foundation for further work on robust speech processing.*
This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ). The Tensorflow version of this model was contributed by [amyeroberts](https://huggingface.co/amyeroberts).
The original code can be found [here](https://github.com/openai/whisper).
## Usage tips
- The model usually performs well without requiring any finetuning.
- The architecture follows a classic encoder-decoder architecture, which means that it relies on the [`~generation.GenerationMixin.generate`] function for inference.
- Inference is currently only implemented for short-form i.e. audio is pre-segmented into <=30s segments. Long-form (including timestamps) will be implemented in a future release.
- One can use [`WhisperProcessor`] to prepare audio for the model, and decode the predicted ID's back into text.
- To convert the model and the processor, we recommend using the following:
```bash
python src/transformers/models/whisper/convert_openai_to_hf.py --checkpoint_path "" --pytorch_dump_folder_path "Arthur/whisper-3" --convert_preprocessor True
```
The script will automatically determine all necessary parameters from the OpenAI checkpoint. A `tiktoken` library needs to be installed
to perform the conversion of the OpenAI tokenizer to the `tokenizers` version.
## Inference
Here is a step-by-step guide to transcribing an audio sample using a pre-trained Whisper model:
```python
>>> from datasets import load_dataset
>>> from transformers import WhisperProcessor, WhisperForConditionalGeneration
>>> # Select an audio file and read it:
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> audio_sample = ds[0]["audio"]
>>> waveform = audio_sample["array"]
>>> sampling_rate = audio_sample["sampling_rate"]
>>> # Load the Whisper model in Hugging Face format:
>>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
>>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
>>> # Use the model and processor to transcribe the audio:
>>> input_features = processor(
... waveform, sampling_rate=sampling_rate, return_tensors="pt"
... ).input_features
>>> # Generate token ids
>>> predicted_ids = model.generate(input_features)
>>> # Decode token ids to text
>>> transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
>>> transcription[0]
' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
```
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with Whisper. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
- A fork with a script to [convert a Whisper model in Hugging Face format to OpenAI format](https://github.com/zuazo-forks/transformers/blob/convert_hf_to_openai/src/transformers/models/whisper/convert_hf_to_openai.py). 🌎
Usage example:
```bash
pip install -U openai-whisper
python convert_hf_to_openai.py \
--checkpoint openai/whisper-tiny \
--whisper_dump_path whisper-tiny-openai.pt
```
## WhisperConfig
[[autodoc]] WhisperConfig
## WhisperTokenizer
[[autodoc]] WhisperTokenizer
- set_prefix_tokens
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
- batch_decode
- decode
## WhisperTokenizerFast
[[autodoc]] WhisperTokenizerFast
- set_prefix_tokens
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
- batch_decode
- decode
## WhisperFeatureExtractor
[[autodoc]] WhisperFeatureExtractor
- __call__
## WhisperProcessor
[[autodoc]] WhisperProcessor
- __call__
- from_pretrained
- save_pretrained
- batch_decode
- decode
<frameworkcontent>
<pt>
## WhisperModel
[[autodoc]] WhisperModel
- forward
- _mask_input_features
## WhisperForConditionalGeneration
[[autodoc]] WhisperForConditionalGeneration
- forward
- generate
## WhisperForCausalLM
[[autodoc]] WhisperForCausalLM
- forward
## WhisperForAudioClassification
[[autodoc]] WhisperForAudioClassification
- forward
</pt>
<tf>
## TFWhisperModel
[[autodoc]] TFWhisperModel
- call
## TFWhisperForConditionalGeneration
[[autodoc]] TFWhisperForConditionalGeneration
- call
</tf>
<jax>
## FlaxWhisperModel
[[autodoc]] FlaxWhisperModel
- __call__
## FlaxWhisperForConditionalGeneration
[[autodoc]] FlaxWhisperForConditionalGeneration
- __call__
## FlaxWhisperForAudioClassification
[[autodoc]] FlaxWhisperForAudioClassification
- __call__
</jax>
</frameworkcontent>
| huggingface/transformers/blob/main/docs/source/en/model_doc/whisper.md |
Spaces
[Hugging Face Spaces](https://huggingface.co/spaces) offer a simple way to host ML demo apps directly on your profile or your organization's profile. This allows you to create your ML portfolio, showcase your projects at conferences or to stakeholders, and work collaboratively with other people in the ML ecosystem.
We have built-in support for two awesome SDKs that let you build cool apps in Python in a matter of minutes: **[Streamlit](https://streamlit.io/)** and **[Gradio](https://gradio.app/)**, but you can also unlock the whole power of Docker and host an arbitrary Dockerfile. Finally, you can create static Spaces using JavaScript and HTML.
You'll also be able to upgrade your Space to run [on a GPU or other accelerated hardware](./spaces-gpus). ⚡️
## Contents
- [Spaces Overview](./spaces-overview)
- [Handling Spaces Dependencies](./spaces-dependencies)
- [Spaces Settings](./spaces-settings)
- [Using OpenCV in Spaces](./spaces-using-opencv)
- [Using Spaces for Organization Cards](./spaces-organization-cards)
- [More ways to create Spaces](./spaces-more-ways-to-create)
- [Managing Spaces with Github Actions](./spaces-github-actions)
- [How to Add a Space to ArXiv](./spaces-add-to-arxiv)
- [Spaces GPU Upgrades](./spaces-gpus)
- [Spaces Persistent Storage](./spaces-storage)
- [Gradio Spaces](./spaces-sdks-gradio)
- [Streamlit Spaces](./spaces-sdks-streamlit)
- [Docker Spaces](./spaces-sdks-docker)
- [Static HTML Spaces](./spaces-sdks-static)
- [Custom Python Spaces](./spaces-sdks-python)
- [Embed your Space](./spaces-embed)
- [Run your Space with Docker](./spaces-run-with-docker)
- [Reference](./spaces-config-reference)
- [Changelog](./spaces-changelog)
## Contact
Feel free to ask questions on the [forum](https://discuss.huggingface.co/c/spaces/24) if you need help with making a Space, or if you run into any other issues on the Hub.
If you're interested in infra challenges, custom demos, advanced GPUs, or something else, please reach out to us by sending an email to **website at huggingface.co**.
You can also tag us [on Twitter](https://twitter.com/huggingface)! 🤗
| huggingface/hub-docs/blob/main/docs/hub/spaces.md |
--
title: "Welcome PaddlePaddle to the Hugging Face Hub"
thumbnail: /blog/assets/126_paddlepaddle/thumbnail.jpg
authors:
- user: PaddlePaddle
guest: true
---
# Welcome PaddlePaddle to the Hugging Face Hub
We are happy to share an open source collaboration between Hugging Face and [PaddlePaddle](https://www.paddlepaddle.org.cn/en) on a shared mission to advance and democratize AI through open source!
First open sourced by Baidu in 2016, PaddlePaddle enables developers of all skill levels to adopt and implement Deep Learning at scale. As of Q4 2022, PaddlePaddle is being used by more than 5.35 million developers and 200,000 enterprises, ranking first in terms of market share among Deep Learning platforms in China. PaddlePaddle features popular open source repositories such as the [Paddle](https://github.com/PaddlePaddle/Paddle) Deep Learning Framework, model libraries across different modalities (e.g. [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR), [PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection), [PaddleNLP](https://github.com/PaddlePaddle/PaddleNLP), [PaddleSpeech](https://github.com/PaddlePaddle/PaddleSpeech)), [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim) for model compression, [FastDeploy](https://github.com/PaddlePaddle/FastDeploy) for model deployment and many more.
**With [PaddleNLP](https://huggingface.co/docs/hub/paddlenlp) leading the way, PaddlePaddle will gradually integrate its libraries with the Hugging Face Hub.** You will soon be able to play with the full suite of awesome pre-trained PaddlePaddle models across text, image, audio, video and multi-modalities on the Hub!
## Find PaddlePaddle Models
You can find all PaddlePaddle models on the Model Hub by filtering with the [PaddlePaddle library tag](https://huggingface.co/models?library=paddlepaddle).
<p align="center">
<img src="assets/126_paddlepaddle/paddle_tag.png" alt="PaddlePaddle Tag"/>
</p>
There are already over 75 PaddlePaddle models on the Hub. As an example, you can find our multi-task Information Extraction model series [UIE](https://huggingface.co/PaddlePaddle/uie-base), State-of-the-Art Chinese Language Model [ERNIE 3.0 model series](https://huggingface.co/PaddlePaddle/ernie-3.0-nano-zh), novel document pre-training model [Ernie-Layout](PaddlePaddle/ernie-layoutx-base-uncased) with layout knowledge enhancement in the whole workflow and so on.
You are also welcome to check out the [PaddlePaddle](https://huggingface.co/PaddlePaddle) org on the HuggingFace Hub. In additional to the above-mentioned models, you can also explore our Spaces, including our text-to-image [Ernie-ViLG](https://huggingface.co/spaces/PaddlePaddle/ERNIE-ViLG), cross-modal Information Extraction engine [UIE-X](https://huggingface.co/spaces/PaddlePaddle/UIE-X) and awesome multilingual OCR toolkit [PaddleOCR](https://huggingface.co/spaces/PaddlePaddle/PaddleOCR).
## Inference API and Widgets
PaddlePaddle models are available through the [Inference API](https://huggingface.co/docs/hub/models-inference), which you can access through HTTP with cURL, Python’s requests library, or your preferred method for making network requests.
Models that support a [task](https://huggingface.co/tasks) are equipped with an interactive widget that allows you to play with the model directly in the browser.
## Use Existing Models
If you want to see how to load a specific model, you can click `Use in paddlenlp` (or other PaddlePaddle libraries in the future) and you will be given a working snippet that to load it!
## Share Models
Depending on the PaddlePaddle library, you may be able to share your models by pushing to the Hub. For example, you can share PaddleNLP models by using the `save_to_hf_hub` method.
```python
from paddlenlp.transformers import AutoTokenizer, AutoModelForMaskedLM
tokenizer = AutoTokenizer.from_pretrained("PaddlePaddle/ernie-3.0-base-zh", from_hf_hub=True)
model = AutoModelForMaskedLM.from_pretrained("PaddlePaddle/ernie-3.0-base-zh", from_hf_hub=True)
tokenizer.save_to_hf_hub(repo_id="<my_org_name>/<my_repo_name>")
model.save_to_hf_hub(repo_id="<my_org_name>/<my_repo_name>")
```
## Conclusion
PaddlePaddle is an open source Deep Learning platform that originated from industrial practice and has been open-sourcing innovative and industry-grade projects since 2016. We are excited to join the Hub to share our work with the HuggingFace community and you can expect more fun and State-of-the-Art projects from us soon! To stay up to date with the latest news, you can follow us on Twitter at [@PaddlePaddle](https://twitter.com/PaddlePaddle).
| huggingface/blog/blob/main/paddlepaddle.md |
Self-Play: a classic technique to train competitive agents in adversarial games
Now that we've studied the basics of multi-agents, we're ready to go deeper. As mentioned in the introduction, we're going **to train agents in an adversarial game with SoccerTwos, a 2vs2 game**.
<figure>
<img src="https://huggingface.co/datasets/huggingface-deep-rl-course/course-images/resolve/main/en/unit10/soccertwos.gif" alt="SoccerTwos"/>
<figcaption>This environment was made by the <a href="https://github.com/Unity-Technologies/ml-agents">Unity MLAgents Team</a></figcaption>
</figure>
## What is Self-Play?
Training agents correctly in an adversarial game can be **quite complex**.
On the one hand, we need to find how to get a well-trained opponent to play against your training agent. And on the other hand, if you find a very good trained opponent, how will your agent improve its policy when the opponent is too strong?
Think of a child that just started to learn soccer. Playing against a very good soccer player will be useless since it will be too hard to win or at least get the ball from time to time. So the child will continuously lose without having time to learn a good policy.
The best solution would be **to have an opponent that is on the same level as the agent and will upgrade its level as the agent upgrades its own**. Because if the opponent is too strong, we’ll learn nothing; if it is too weak, we’ll overlearn useless behavior against a stronger opponent then.
This solution is called *self-play*. In self-play, **the agent uses former copies of itself (of its policy) as an opponent**. This way, the agent will play against an agent of the same level (challenging but not too much), have opportunities to gradually improve its policy, and then update its opponent as it becomes better. It’s a way to bootstrap an opponent and progressively increase the opponent's complexity.
It’s the same way humans learn in competition:
- We start to train against an opponent of similar level
- Then we learn from it, and when we acquire some skills, we can move further with stronger opponents.
We do the same with self-play:
- We **start with a copy of our agent as an opponent** this way, this opponent is on a similar level.
- We **learn from it** and, when we acquire some skills, we **update our opponent with a more recent copy of our training policy**.
The theory behind self-play is not something new. It was already used by Arthur Samuel’s checker player system in the fifties and by Gerald Tesauro’s TD-Gammon in 1995. If you want to learn more about the history of self-play [check out this very good blogpost by Andrew Cohen](https://blog.unity.com/technology/training-intelligent-adversaries-using-self-play-with-ml-agents)
## Self-Play in MLAgents
Self-Play is integrated into the MLAgents library and is managed by multiple hyperparameters that we’re going to study. But the main focus, as explained in the documentation, is the **tradeoff between the skill level and generality of the final policy and the stability of learning**.
Training against a set of slowly changing or unchanging adversaries with low diversity **results in more stable training. But a risk to overfit if the change is too slow.**
So we need to control:
- How **often we change opponents** with the `swap_steps` and `team_change` parameters.
- The **number of opponents saved** with the `window` parameter. A larger value of `window`
means that an agent's pool of opponents will contain a larger diversity of behaviors since it will contain policies from earlier in the training run.
- The **probability of playing against the current self vs opponent** sampled from the pool with `play_against_latest_model_ratio`. A larger value of `play_against_latest_model_ratio`
indicates that an agent will be playing against the current opponent more often.
- The **number of training steps before saving a new opponent** with `save_steps` parameters. A larger value of `save_steps`
will yield a set of opponents that cover a wider range of skill levels and possibly play styles since the policy receives more training.
To get more details about these hyperparameters, you definitely need [to check out this part of the documentation](https://github.com/Unity-Technologies/ml-agents/blob/develop/docs/Training-Configuration-File.md#self-play)
## The ELO Score to evaluate our agent
### What is ELO Score?
In adversarial games, tracking the **cumulative reward is not always a meaningful metric to track the learning progress:** because this metric is **dependent only on the skill of the opponent.**
Instead, we’re using an ***ELO rating system*** (named after Arpad Elo) that calculates the **relative skill level** between 2 players from a given population in a zero-sum game.
In a zero-sum game: one agent wins, and the other agent loses. It’s a mathematical representation of a situation in which each participant’s gain or loss of utility **is exactly balanced by the gain or loss of the utility of the other participants.** We talk about zero-sum games because the sum of utility is equal to zero.
This ELO (starting at a specific score: frequently 1200) can decrease initially but should increase progressively during the training.
The Elo system is **inferred from the losses and draws against other players.** It means that player ratings depend **on the ratings of their opponents and the results scored against them.**
Elo defines an Elo score that is the relative skills of a player in a zero-sum game. **We say relative because it depends on the performance of opponents.**
The central idea is to think of the performance of a player **as a random variable that is normally distributed.**
The difference in rating between 2 players serves as **the predictor of the outcomes of a match.** If the player wins, but the probability of winning is high, it will only win a few points from its opponent since it means that it is much stronger than it.
After every game:
- The winning player takes **points from the losing one.**
- The number of points is determined **by the difference in the 2 players ratings (hence relative).**
- If the higher-rated player wins → few points will be taken from the lower-rated player.
- If the lower-rated player wins → a lot of points will be taken from the high-rated player.
- If it’s a draw → the lower-rated player gains a few points from the higher.
So if A and B have rating Ra, and Rb, then the **expected scores are** given by:
<img src="https://huggingface.co/datasets/huggingface-deep-rl-course/course-images/resolve/main/en/unit10/elo1.png" alt="ELO Score"/>
Then, at the end of the game, we need to update the player’s actual Elo score. We use a linear adjustment **proportional to the amount by which the player over-performed or under-performed.**
We also define a maximum adjustment rating per game: K-factor.
- K=16 for master.
- K=32 for weaker players.
If Player A has Ea points but scored Sa points, then the player’s rating is updated using the formula:
<img src="https://huggingface.co/datasets/huggingface-deep-rl-course/course-images/resolve/main/en/unit10/elo2.png" alt="ELO Score"/>
### Example
If we take an example:
Player A has a rating of 2600
Player B has a rating of 2300
- We first calculate the expected score:
\\(E_{A} = \frac{1}{1+10^{(2300-2600)/400}} = 0.849 \\)
\\(E_{B} = \frac{1}{1+10^{(2600-2300)/400}} = 0.151 \\)
- If the organizers determined that K=16 and A wins, the new rating would be:
\\(ELO_A = 2600 + 16*(1-0.849) = 2602 \\)
\\(ELO_B = 2300 + 16*(0-0.151) = 2298 \\)
- If the organizers determined that K=16 and B wins, the new rating would be:
\\(ELO_A = 2600 + 16*(0-0.849) = 2586 \\)
\\(ELO_B = 2300 + 16 *(1-0.151) = 2314 \\)
### The Advantages
Using the ELO score has multiple advantages:
- Points are **always balanced** (more points are exchanged when there is an unexpected outcome, but the sum is always the same).
- It is a **self-corrected system** since if a player wins against a weak player, they will only win a few points.
- It **works with team games**: we calculate the average for each team and use it in Elo.
### The Disadvantages
- ELO **does not take into account the individual contribution** of each people in the team.
- Rating deflation: **a good rating requires skill over time to keep the same rating**.
- **Can’t compare rating in history**.
| huggingface/deep-rl-class/blob/main/units/en/unit7/self-play.mdx |
--
title: Image Similarity with Hugging Face Datasets and Transformers
thumbnail: /blog/assets/image_similarity/thumbnail.png
authors:
- user: sayakpaul
---
# Image Similarity with Hugging Face Datasets and Transformers
<a target="_blank" href="https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_similarity.ipynb">
<img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
</a>
In this post, you'll learn to build an image similarity system with 🤗 Transformers. Finding out the similarity between a query image and potential candidates is an important use case for information retrieval systems, such as reverse image search, for example. All the system is trying to answer is that, given a _query_ image and a set of _candidate_ images, which images are the most similar to the query image.
We'll leverage the [🤗 `datasets` library](https://huggingface.co/docs/datasets/) as it seamlessly supports parallel processing which will come in handy when building this system.
Although the post uses a ViT-based model ([`nateraw/vit-base-beans`](https://huggingface.co/nateraw/vit-base-beans)) and a particular dataset ([Beans](https://huggingface.co/datasets/beans)), it can be extended to use other models supporting vision modality and other image datasets. Some notable models you could try:
* [Swin Transformer](https://huggingface.co/docs/transformers/model_doc/swin)
* [ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)
* [RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)
Also, the approach presented in the post can potentially be extended to other modalities as well.
To study the fully working image-similarity system, you can refer to the Colab Notebook linked at the beginning.
## How do we define similarity?
To build this system, we first need to define how we want to compute the similarity between two images. One widely popular practice is to compute dense representations (embeddings) of the given images and then use the [cosine similarity metric](https://en.wikipedia.org/wiki/Cosine_similarity) to determine how similar the two images are.
For this post, we'll use “embeddings” to represent images in vector space. This gives us a nice way to meaningfully compress the high-dimensional pixel space of images (224 x 224 x 3, for example) to something much lower dimensional (768, for example). The primary advantage of doing this is the reduced computation time in the subsequent steps.
<div align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/image_similarity/embeddings.png" width=700/>
</div>
## Computing embeddings
To compute the embeddings from the images, we'll use a vision model that has some understanding of how to represent the input images in the vector space. This type of model is also commonly referred to as image encoder.
For loading the model, we leverage the [`AutoModel` class](https://huggingface.co/docs/transformers/model_doc/auto#transformers.AutoModel). It provides an interface for us to load any compatible model checkpoint from the Hugging Face Hub. Alongside the model, we also load the processor associated with the model for data preprocessing.
```py
from transformers import AutoImageProcessor, AutoModel
model_ckpt = "nateraw/vit-base-beans"
processor = AutoImageProcessor.from_pretrained(model_ckpt)
model = AutoModel.from_pretrained(model_ckpt)
```
In this case, the checkpoint was obtained by fine-tuning a [Vision Transformer based model](https://huggingface.co/google/vit-base-patch16-224-in21k) on the [`beans` dataset](https://huggingface.co/datasets/beans).
Some questions that might arise here:
**Q1**: Why did we not use `AutoModelForImageClassification`?
This is because we want to obtain dense representations of the images and not discrete categories, which are what `AutoModelForImageClassification` would have provided.
**Q2**: Why this checkpoint in particular?
As mentioned earlier, we're using a specific dataset to build the system. So, instead of using a generalist model (like the [ones trained on the ImageNet-1k dataset](https://huggingface.co/models?dataset=dataset:imagenet-1k&sort=downloads), for example), it's better to use a model that has been fine-tuned on the dataset being used. That way, the underlying model better understands the input images.
**Note** that you can also use a checkpoint that was obtained through self-supervised pre-training. The checkpoint doesn't necessarily have to come from supervised learning. In fact, if pre-trained well, self-supervised models can [yield](https://ai.facebook.com/blog/dino-paws-computer-vision-with-self-supervised-transformers-and-10x-more-efficient-training/) impressive retrieval
performance.
Now that we have a model for computing the embeddings, we need some candidate images to query against.
## Loading a dataset for candidate images
In some time, we'll be building hash tables mapping the candidate images to hashes. During the query time, we'll use these hash tables. We'll talk more about hash tables in the respective section but for now, to have a set of candidate images, we will use the `train` split of the [`beans` dataset](https://huggingface.co/datasets/beans).
```py
from datasets import load_dataset
dataset = load_dataset("beans")
```
This is how a single sample from the training split looks like:
<div align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/image_similarity/beans.png" width=600/>
</div>
The dataset has three features:
```py
dataset["train"].features
>>> {'image_file_path': Value(dtype='string', id=None),
'image': Image(decode=True, id=None),
'labels': ClassLabel(names=['angular_leaf_spot', 'bean_rust', 'healthy'], id=None)}
```
To demonstrate the image similarity system, we'll use 100 samples from the candidate image dataset to keep
the overall runtime short.
```py
num_samples = 100
seed = 42
candidate_subset = dataset["train"].shuffle(seed=seed).select(range(num_samples))
```
## The process of finding similar images
Below, you can find a pictorial overview of the process underlying fetching similar images.
<div align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/image_similarity/fetch-similar-process.png">
</div>
Breaking down the above figure a bit, we have:
1. Extract the embeddings from the candidate images (`candidate_subset`), storing them in a matrix.
2. Take a query image and extract its embeddings.
3. Iterate over the embedding matrix (computed in step 1) and compute the similarity score between the query embedding and the current candidate embeddings. We usually maintain a dictionary-like mapping maintaining a correspondence between some identifier of the candidate image and the similarity scores.
4. Sort the mapping structure w.r.t the similarity scores and return the underlying identifiers. We use these identifiers to fetch the candidate samples.
We can write a simple utility and `map()` it to our dataset of candidate images to compute the embeddings efficiently.
```py
import torch
def extract_embeddings(model: torch.nn.Module):
"""Utility to compute embeddings."""
device = model.device
def pp(batch):
images = batch["image"]
# `transformation_chain` is a compostion of preprocessing
# transformations we apply to the input images to prepare them
# for the model. For more details, check out the accompanying Colab Notebook.
image_batch_transformed = torch.stack(
[transformation_chain(image) for image in images]
)
new_batch = {"pixel_values": image_batch_transformed.to(device)}
with torch.no_grad():
embeddings = model(**new_batch).last_hidden_state[:, 0].cpu()
return {"embeddings": embeddings}
return pp
```
And we can map `extract_embeddings()` like so:
```py
device = "cuda" if torch.cuda.is_available() else "cpu"
extract_fn = extract_embeddings(model.to(device))
candidate_subset_emb = candidate_subset.map(extract_fn, batched=True, batch_size=batch_size)
```
Next, for convenience, we create a list containing the identifiers of the candidate images.
```py
candidate_ids = []
for id in tqdm(range(len(candidate_subset_emb))):
label = candidate_subset_emb[id]["labels"]
# Create a unique indentifier.
entry = str(id) + "_" + str(label)
candidate_ids.append(entry)
```
We'll use the matrix of the embeddings of all the candidate images for computing the similarity scores with a query image. We have already computed the candidate image embeddings. In the next cell, we just gather them together in a matrix.
```py
all_candidate_embeddings = np.array(candidate_subset_emb["embeddings"])
all_candidate_embeddings = torch.from_numpy(all_candidate_embeddings)
```
We'll use [cosine similarity](https://en.wikipedia.org/wiki/Cosine_similarity) to compute the similarity score in between two embedding vectors. We'll then use it to fetch similar candidate samples given a query sample.
```py
def compute_scores(emb_one, emb_two):
"""Computes cosine similarity between two vectors."""
scores = torch.nn.functional.cosine_similarity(emb_one, emb_two)
return scores.numpy().tolist()
def fetch_similar(image, top_k=5):
"""Fetches the `top_k` similar images with `image` as the query."""
# Prepare the input query image for embedding computation.
image_transformed = transformation_chain(image).unsqueeze(0)
new_batch = {"pixel_values": image_transformed.to(device)}
# Comute the embedding.
with torch.no_grad():
query_embeddings = model(**new_batch).last_hidden_state[:, 0].cpu()
# Compute similarity scores with all the candidate images at one go.
# We also create a mapping between the candidate image identifiers
# and their similarity scores with the query image.
sim_scores = compute_scores(all_candidate_embeddings, query_embeddings)
similarity_mapping = dict(zip(candidate_ids, sim_scores))
# Sort the mapping dictionary and return `top_k` candidates.
similarity_mapping_sorted = dict(
sorted(similarity_mapping.items(), key=lambda x: x[1], reverse=True)
)
id_entries = list(similarity_mapping_sorted.keys())[:top_k]
ids = list(map(lambda x: int(x.split("_")[0]), id_entries))
labels = list(map(lambda x: int(x.split("_")[-1]), id_entries))
return ids, labels
```
## Perform a query
Given all the utilities, we're equipped to do a similarity search. Let's have a query image from the `test` split of
the `beans` dataset:
```py
test_idx = np.random.choice(len(dataset["test"]))
test_sample = dataset["test"][test_idx]["image"]
test_label = dataset["test"][test_idx]["labels"]
sim_ids, sim_labels = fetch_similar(test_sample)
print(f"Query label: {test_label}")
print(f"Top 5 candidate labels: {sim_labels}")
```
Leads to:
```
Query label: 0
Top 5 candidate labels: [0, 0, 0, 0, 0]
```
Seems like our system got the right set of similar images. When visualized, we'd get:
<div align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/image_similarity/results_one.png">
</div>
## Further extensions and conclusions
We now have a working image similarity system. But in reality, you'll be dealing with a lot more candidate images. Taking that into consideration, our current procedure has got multiple drawbacks:
* If we store the embeddings as is, the memory requirements can shoot up quickly, especially when dealing with millions of candidate images. The embeddings are 768-d in our case, which can still be relatively high in the large-scale regime.
* Having high-dimensional embeddings have a direct effect on the subsequent computations involved in the retrieval part.
If we can somehow reduce the dimensionality of the embeddings without disturbing their meaning, we can still maintain a good trade-off between speed and retrieval quality. The [accompanying Colab Notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_similarity.ipynb) of this post implements and demonstrates utilities for achieving this with random projection and locality-sensitive hashing.
🤗 Datasets offers direct integrations with [FAISS](https://github.com/facebookresearch/faiss) which further simplifies the process of building similarity systems. Let's say you've already extracted the embeddings of the candidate images (the `beans` dataset) and stored them
inside a feature called `embeddings`. You can now easily use the [`add_faiss_index()`](https://huggingface.co/docs/datasets/v2.7.1/en/package_reference/main_classes#datasets.Dataset.add_faiss_index) of the dataset to build a dense index:
```py
dataset_with_embeddings.add_faiss_index(column="embeddings")
```
Once the index is built, `dataset_with_embeddings` can be used to retrieve the nearest examples given query embeddings with [`get_nearest_examples()`](https://huggingface.co/docs/datasets/v2.7.1/en/package_reference/main_classes#datasets.Dataset.get_nearest_examples):
```py
scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples(
"embeddings", qi_embedding, k=top_k
)
```
The method returns scores and corresponding candidate examples. To know more, you can check out the [official documentation](https://huggingface.co/docs/datasets/faiss_es) and [this notebook](https://colab.research.google.com/gist/sayakpaul/5b5b5a9deabd3c5d8cb5ef8c7b4bb536/image_similarity_faiss.ipynb).
Finally, you can try out the following Space that builds a mini image similarity application:
<script type="module" src="https://gradio.s3-us-west-2.amazonaws.com/3.12.0/gradio.js"></script>
<gradio-app theme_mode="light" space="sayakpaul/fetch-similar-images"></gradio-app>
In this post, we ran through a quickstart for building image similarity systems. If you found this post interesting, we highly recommend building on top of the concepts we discussed here so you can get more comfortable with the inner workings.
Still looking to learn more? Here are some additional resources that might be useful for you:
* [Faiss: A library for efficient similarity search](https://engineering.fb.com/2017/03/29/data-infrastructure/faiss-a-library-for-efficient-similarity-search/)
* [ScaNN: Efficient Vector Similarity Search](http://ai.googleblog.com/2020/07/announcing-scann-efficient-vector.html)
* [Integrating Image Searchers within Mobile Applications](https://www.tensorflow.org/lite/inference_with_metadata/task_library/image_searcher) | huggingface/blog/blob/main/image-similarity.md |
--
title: "Japanese Stable Diffusion"
thumbnail: /blog/assets/106_japanese_stable_diffusion/jsd_thumbnail.png
authors:
- user: mshing
guest: true
- user: keisawada
guest: true
---
# Japanese Stable Diffusion
<a target="_blank" href="https://huggingface.co/spaces/rinna/japanese-stable-diffusion" target="_parent"><img src="https://img.shields.io/badge/🤗 Hugging Face-Spaces-blue" alt="Open In Hugging Face Spaces"/></a>
<a target="_blank" href="https://colab.research.google.com/github/rinnakk/japanese-stable-diffusion/blob/master/scripts/txt2img.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/></a>
Stable Diffusion, developed by [CompVis](https://github.com/CompVis), [Stability AI](https://stability.ai/), and [LAION](https://laion.ai/), has generated a great deal of interest due to its ability to generate highly accurate images by simply entering text prompts. Stable Diffusion mainly uses the English subset [LAION2B-en](https://huggingface.co/datasets/laion/laion2B-en) of the [LAION-5B](https://laion.ai/blog/laion-5b/) dataset for its training data and, as a result, requires English text prompts to be entered producing images that tend to be more oriented towards Western culture.
[rinna Co., Ltd](https://rinna.co.jp/). has developed a Japanese-specific text-to-image model named "Japanese Stable Diffusion" by fine-tuning Stable Diffusion on Japanese-captioned images. Japanese Stable Diffusion accepts Japanese text prompts and generates images that reflect the culture of the Japanese-speaking world which may be difficult to express through translation.
In this blog, we will discuss the background of the development of Japanese Stable Diffusion and its learning methodology.
Japanese Stable Diffusion is available on Hugging Face and GitHub. The code is based on [🧨 Diffusers](https://huggingface.co/docs/diffusers/index).
- Hugging Face model card: https://huggingface.co/rinna/japanese-stable-diffusion
- Hugging Face Spaces: https://huggingface.co/spaces/rinna/japanese-stable-diffusion
- GitHub: https://github.com/rinnakk/japanese-stable-diffusion
## Stable Diffusion
Recently diffusion models have been reported to be very effective in artificial synthesis, even more so than GANs (Generative Adversarial Networks) for images. Hugging Face explains how diffusion models work in the following articles:
- [The Annotated Diffusion Model](https://huggingface.co/blog/annotated-diffusion)
- [Getting started with 🧨 Diffusers](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/diffusers_intro.ipynb)
Generally, a text-to-image model consists of a text encoder that interprets text and a generative model that generates an image from its output.
Stable Diffusion uses CLIP, the language-image pre-training model from OpenAI, as its text encoder and a latent diffusion model, which is an improved version of the diffusion model, as the generative model. Stable Diffusion was trained mainly on the English subset of LAION-5B and can generate high-performance images simply by entering text prompts. In addition to its high performance, Stable Diffusion is also easy to use with inference running at a computing cost of about 10GB VRAM GPU.
<p align="center">
<img src="https://raw.githubusercontent.com/patrickvonplaten/scientific_images/master/stable_diffusion.png" alt="sd-pipeline" width="300"/>
</p>
*from [Stable Diffusion with 🧨 Diffusers](https://huggingface.co/blog/stable_diffusion)*
## Japanese Stable Diffusion
### Why do we need Japanese Stable Diffusion?
Stable Diffusion is a very powerful text-to-image model not only in terms of quality but also in terms of computational cost. Because Stable Diffusion was trained on an English dataset, it is required to translate non-English prompts to English first. Surprisingly, Stable Diffusion can sometimes generate proper images even when using non-English prompts.
So, why do we need a language-specific Stable Diffusion? The answer is because we want a text-to-image model that can understand Japanese culture, identity, and unique expressions including slang. For example, one of the more common Japanese terms re-interpreted from the English word businessman is "salary man" which we most often imagine as a man wearing a suit. Stable Diffusion cannot understand such Japanese unique words correctly because Japanese is not their target.
<p align="center">
<img src="assets/106_japanese_stable_diffusion/sd.jpeg" alt="salary man of stable diffusion" title="salary man of stable diffusion">
</p>
*"salary man, oil painting" from the original Stable Diffusion*
So, this is why we made a language-specific version of Stable Diffusion. Japanese Stable Diffusion can achieve the following points compared to the original Stable Diffusion.
- Generate Japanese-style images
- Understand Japanese words adapted from English
- Understand Japanese unique onomatope
- Understand Japanese proper noun
### Training Data
We used approximately 100 million images with Japanese captions, including the Japanese subset of [LAION-5B](https://laion.ai/blog/laion-5b/). In addition, to remove low quality samples, we used [japanese-cloob-vit-b-16](https://huggingface.co/rinna/japanese-cloob-vit-b-16) published by rinna Co., Ltd. as a preprocessing step to remove samples whose scores were lower than a certain threshold.
### Training Details
The biggest challenge in making a Japanese-specific text-to-image model is the size of the dataset. Non-English datasets are much smaller than English datasets, and this causes performance degradation in deep learning-based models. The dataset used to train Japanese Stable Diffusion is 1/20th the size of the dataset on which Stable Diffusion is trained. To make a good model with such a small dataset, we fine-tuned the powerful [Stable Diffusion](https://huggingface.co/CompVis/stable-diffusion-v1-4) trained on the English dataset, rather than training a text-to-image model from scratch.
To make a good language-specific text-to-image model, we did not simply fine-tune but applied 2 training stages following the idea of [PITI](https://arxiv.org/abs/2205.12952).
#### 1st stage: Train a Japanese-specific text encoder
In the 1st stage, the latent diffusion model is fixed and we replaced the English text encoder with a Japanese-specific text encoder, which is trained. At this time, our Japanese sentencepiece tokenizer is used as the tokenizer. If the CLIP tokenizer is used as it is, Japanese texts are tokenized bytes, which makes it difficult to learn the token dependency, and the number of tokens becomes unnecessarily large. For example, if we tokenize "サラリーマン 油絵", we get `['ãĤ', 'µ', 'ãĥ©', 'ãĥª', 'ãĥ¼ãĥ', 'ŀ', 'ãĥ³</w>', 'æ', '²', '¹', 'çµ', 'µ</w>']` which are uninterpretable tokens.
```python
from transformers import CLIPTokenizer
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
text = "サラリーマン 油絵"
tokens = tokenizer(text, add_special_tokens=False)['input_ids']
print("tokens:", tokenizer.convert_ids_to_tokens(tokens))
# tokens: ['ãĤ', 'µ', 'ãĥ©', 'ãĥª', 'ãĥ¼ãĥ', 'ŀ', 'ãĥ³</w>', 'æ', '²', '¹', 'çµ', 'µ</w>']
print("decoded text:", tokenizer.decode(tokens))
# decoded text: サラリーマン 油絵
```
On the other hand, by using our Japanese tokenizer, the prompt is split into interpretable tokens and the number of tokens is reduced. For example, "サラリーマン 油絵" can be tokenized as `['▁', 'サラリーマン', '▁', '油', '絵']`, which is correctly tokenized in Japanese.
```python
from transformers import T5Tokenizer
tokenizer = T5Tokenizer.from_pretrained("rinna/japanese-stable-diffusion", subfolder="tokenizer", use_auth_token=True)
tokenizer.do_lower_case = True
tokens = tokenizer(text, add_special_tokens=False)['input_ids']
print("tokens:", tokenizer.convert_ids_to_tokens(tokens))
# tokens: ['▁', 'サラリーマン', '▁', '油', '絵']
print("decoded text:", tokenizer.decode(tokens))
# decoded text: サラリーマン 油絵
```
This stage enables the model to understand Japanese prompts but does not still output Japanese-style images because the latent diffusion model has not been changed at all. In other words, the Japanese word "salary man" can be interpreted as the English word "businessman," but the generated result is a businessman with a Western face, as shown below.
<p align="center">
<img src="assets/106_japanese_stable_diffusion/jsd-stage1.jpeg" alt="salary man of japanese stable diffusion at stage 1" title="salary man of japanese stable diffusion at stage 1">
</p>
*"サラリーマン 油絵", which means exactly "salary man, oil painting", from the 1st-stage Japanese Stable Diffusion*
Therefore, in the 2nd stage, we train to output more Japanese-style images.
#### 2nd stage: Fine-tune the text encoder and the latent diffusion model jointly
In the 2nd stage, we will train both the text encoder and the latent diffusion model to generate Japanese-style images. This stage is essential to make the model become a more language-specific model. After this, the model can finally generate a businessman with a Japanese face, as shown in the image below.
<p align="center">
<img src="assets/106_japanese_stable_diffusion/jsd-stage2.jpeg" alt="salary man of japanese stable diffusion" title="salary man of japanese stable diffusion">
</p>
*"サラリーマン 油絵", which means exactly "salary man, oil painting", from the 2nd-stage Japanese Stable Diffusion*
## rinna’s Open Strategy
Numerous research institutes are releasing their research results based on the idea of democratization of AI, aiming for a world where anyone can easily use AI. In particular, recently, pre-trained models with a large number of parameters based on large-scale training data have become the mainstream, and there are concerns about a monopoly of high-performance AI by research institutes with computational resources. Still, fortunately, many pre-trained models have been released and are contributing to the development of AI technology. However, pre-trained models on text often target English, the world's most popular language. For a world in which anyone can easily use AI, we believe that it is desirable to be able to use state-of-the-art AI in languages other than English.
Therefore, rinna Co., Ltd. has released [GPT](https://huggingface.co/rinna/japanese-gpt-1b), [BERT](https://huggingface.co/rinna/japanese-roberta-base), and [CLIP](https://huggingface.co/rinna/japanese-clip-vit-b-16), which are specialized for Japanese, and now have also released [Japanese Stable Diffusion](https://huggingface.co/rinna/japanese-stable-diffusion). By releasing a pre-trained model specialized for Japanese, we hope to make AI that is not biased toward the cultures of the English-speaking world but also incorporates the culture of the Japanese-speaking world. Making it available to everyone will help to democratize an AI that guarantees Japanese cultural identity.
## What’s Next?
Compared to Stable Diffusion, Japanese Stable Diffusion is not as versatile and still has some accuracy issues. However, through the development and release of Japanese Stable Diffusion, we hope to communicate to the research community the importance and potential of language-specific model development.
rinna Co., Ltd. has released GPT and BERT models for Japanese text, and CLIP, CLOOB, and Japanese Stable Diffusion models for Japanese text and images. We will continue to improve these models and next we will consider releasing models based on self-supervised learning specialized for Japanese speech.
| huggingface/blog/blob/main/japanese-stable-diffusion.md |
使用 Gradio Python 客户端入门
Tags: CLIENT, API, SPACES
Gradio Python 客户端使得将任何 Gradio 应用程序作为 API 使用变得非常容易。例如，考虑一下从麦克风录制的[Whisper 音频文件](https://huggingface.co/spaces/abidlabs/whisper)的转录。
使用 `gradio_client` 库，我们可以轻松地将 Gradio 用作 API，以编程方式转录音频文件。
下面是完成此操作的整个代码：
```python
from gradio_client import Client
client = Client("abidlabs/whisper")
client.predict("audio_sample.wav")
>> "这是Whisper语音识别模型的测试。"
```
Gradio 客户端适用于任何托管在 Hugging Face Spaces 上的 Gradio 应用程序，无论是图像生成器、文本摘要生成器、有状态聊天机器人、税金计算器还是其他任何应用程序！Gradio 客户端主要用于托管在[Hugging Face Spaces](https://hf.space)上的应用程序，但你的应用程序可以托管在任何地方，比如你自己的服务器。
**先决条件**：要使用 Gradio 客户端，你不需要详细了解 `gradio` 库。但是，了解 Gradio 的输入和输出组件的概念会有所帮助。
## 安装
如果你已经安装了最新版本的 `gradio`，那么 `gradio_client` 就作为依赖项包含在其中。
否则，可以使用 pip（或 pip3）安装轻量级的 `gradio_client` 包，并且已经测试可以在 Python 3.9 或更高版本上运行：
```bash
$ pip install gradio_client
```
## 连接到运行中的 Gradio 应用程序
首先创建一个 `Client` 对象，并将其连接到运行在 Hugging Face Spaces 上或其他任何地方的 Gradio 应用程序。
## 连接到 Hugging Face 空间
```python
from gradio_client import Client
client = Client("abidlabs/en2fr") # 一个将英文翻译为法文的Space
```
你还可以通过在 `hf_token` 参数中传递你的 HF 令牌来连接到私有空间。你可以在这里获取你的 HF 令牌：https://huggingface.co/settings/tokens
```python
from gradio_client import Client
client = Client("abidlabs/my-private-space", hf_token="...")
```
## 复制空间以供私人使用
虽然你可以将任何公共空间用作 API，但如果你发出太多请求，你可能会受到 Hugging Face 的频率限制。要无限制地使用一个空间，只需将其复制以创建一个私有空间，然后可以根据需要进行多个请求！
`gradio_client` 包括一个类方法：`Client.duplicate()`，使这个过程变得简单（你需要传递你的[Hugging Face 令牌](https://huggingface.co/settings/tokens)或使用 Hugging Face CLI 登录）：
```python
import os
from gradio_client import Client
HF_TOKEN = os.environ.get("HF_TOKEN")
client = Client.duplicate("abidlabs/whisper", hf_token=HF_TOKEN)
client.predict("audio_sample.wav")
>> "This is a test of the whisper speech recognition model."
```
> > " 这是 Whisper 语音识别模型的测试。"
如果之前已复制了一个空间，重新运行 `duplicate()` 将*不会*创建一个新的空间。相反，客户端将连接到之前创建的空间。因此，多次运行 `Client.duplicate()` 方法是安全的。
**注意：** 如果原始空间使用了 GPU，你的私有空间也将使用 GPU，并且你的 Hugging Face 账户将根据 GPU 的价格计费。为了降低费用，在 1 小时没有活动后，你的空间将自动休眠。你还可以使用 `duplicate()` 的 `hardware` 参数来设置硬件。
## 连接到通用 Gradio 应用程序
如果你的应用程序运行在其他地方，只需提供完整的 URL，包括 "http://" 或 "https://"。下面是一个在共享 URL 上运行的 Gradio 应用程序进行预测的示例：
```python
from gradio_client import Client
client = Client("https://bec81a83-5b5c-471e.gradio.live")
```
## 检查 API 端点
一旦连接到 Gradio 应用程序，可以通过调用 `Client.view_api()` 方法查看可用的 API 端点。对于 Whisper 空间，我们可以看到以下信息：
```bash
Client.predict() Usage Info
---------------------------
Named API endpoints: 1
- predict(input_audio, api_name="/predict") -> value_0
Parameters:
- [Audio] input_audio: str (filepath or URL)
Returns:
- [Textbox] value_0: str (value)
```
这显示了在此空间中有 1 个 API 端点，并显示了如何使用 API 端点进行预测：我们应该调用 `.predict()` 方法（我们将在下面探讨），提供类型为 `str` 的参数 `input_audio`，它是一个`文件路径或 URL`。
我们还应该提供 `api_name='/predict'` 参数给 `predict()` 方法。虽然如果一个 Gradio 应用程序只有一个命名的端点，这不是必需的，但它允许我们在单个应用程序中调用不同的端点（如果它们可用）。如果一个应用程序有无名的 API 端点，可以通过运行 `.view_api(all_endpoints=True)` 来显示它们。
## 进行预测
进行预测的最简单方法是只需使用相应的参数调用 `.predict()` 函数：
```python
from gradio_client import Client
client = Client("abidlabs/en2fr", api_name='/predict')
client.predict("Hello")
>> Bonjour
```
如果有多个参数，那么你应该将它们作为单独的参数传递给 `.predict()`，就像这样：
````python
from gradio_client import Client
client = Client("gradio/calculator")
client.predict(4, "add", 5)
>> 9.0
对于某些输入，例如图像，你应该传递文件的文件路径或URL。同样，对应的输出类型，你将获得一个文件路径或URL。
```python
from gradio_client import Client
client = Client("abidlabs/whisper")
client.predict("https://audio-samples.github.io/samples/mp3/blizzard_unconditional/sample-0.mp3")
>> "My thought I have nobody by a beauty and will as you poured. Mr. Rochester is serve in that so don't find simpus, and devoted abode, to at might in a r—"
````
## 异步运行任务（Running jobs asynchronously）
应注意`.predict()`是一个*阻塞*操作，因为它在返回预测之前等待操作完成。
在许多情况下，直到你需要预测结果之前，你最好让作业在后台运行。你可以通过使用`.submit()`方法创建一个`Job`实例，然后稍后调用`.result()`在作业上获取结果。例如：
```python
from gradio_client import Client
client = Client(space="abidlabs/en2fr")
job = client.submit("Hello", api_name="/predict") # 这不是阻塞的
# 做其他事情
job.result() # 这是阻塞的
>> Bonjour
```
## 添加回调 （Adding callbacks）
或者，可以添加一个或多个回调来在作业完成后执行操作，像这样：
```python
from gradio_client import Client
def print_result(x):
print(" 翻译的结果是：{x}")
client = Client(space="abidlabs/en2fr")
job = client.submit("Hello", api_name="/predict", result_callbacks=[print_result])
# 做其他事情
>> 翻译的结果是：Bonjour
```
## 状态 （Status）
`Job`对象还允许您通过调用`.status()`方法获取运行作业的状态。这将返回一个`StatusUpdate`对象，具有以下属性：`code`（状态代码，其中之一表示状态的一组定义的字符串。参见`utils.Status`类）、`rank`（此作业在队列中的当前位置）、`queue_size`（总队列大小）、`eta`（此作业将完成的预计时间）、`success`（表示作业是否成功完成的布尔值）和`time`（生成状态的时间）。
```py
from gradio_client import Client
client = Client(src="gradio/calculator")
job = client.submit(5, "add", 4, api_name="/predict")
job.status()
>> <Status.STARTING: 'STARTING'>
```
_注意_：`Job`类还有一个`.done()`实例方法，返回一个布尔值，指示作业是否已完成。
## 取消作业 （Cancelling Jobs）
`Job`类还有一个`.cancel()`实例方法，取消已排队但尚未开始的作业。例如，如果你运行：
```py
client = Client("abidlabs/whisper")
job1 = client.submit("audio_sample1.wav")
job2 = client.submit("audio_sample2.wav")
job1.cancel() # 将返回 False，假设作业已开始
job2.cancel() # 将返回 True，表示作业已取消
```
如果第一个作业已开始处理，则它将不会被取消。如果第二个作业尚未开始，则它将成功取消并从队列中删除。
## 生成器端点 （Generator Endpoints）
某些Gradio API端点不返回单个值，而是返回一系列值。你可以随时从这样的生成器端点获取返回的一系列值，方法是运行`job.outputs()`：
```py
from gradio_client import Client
client = Client(src="gradio/count_generator")
job = client.submit(3, api_name="/count")
while not job.done():
time.sleep(0.1)
job.outputs()
>> ['0', '1', '2']
```
请注意，在生成器端点上运行`job.result()`只会获得端点返回的*第一个*值。
`Job`对象还是可迭代的，这意味着您可以使用它按照从端点返回的结果逐个显示生成器函数的结果。以下是使用`Job`作为生成器的等效示例：
```py
from gradio_client import Client
client = Client(src="gradio/count_generator")
job = client.submit(3, api_name="/count")
for o in job:
print(o)
>> 0
>> 1
>> 2
```
你还可以取消具有迭代输出的作业，在这种情况下，作业将在当前迭代完成运行后完成。
```py
from gradio_client import Client
import time
client = Client("abidlabs/test-yield")
job = client.submit("abcdef")
time.sleep(3)
job.cancel() # 作业在运行 2 个迭代后取消
```
| gradio-app/gradio/blob/main/guides/cn/06_client-libraries/01_getting-started-with-the-python-client.md |
Gradio Demo: markdown_component
```
!pip install -q gradio
```
```
import gradio as gr
with gr.Blocks() as demo:
gr.Markdown(value="This _example_ was **written** in [Markdown](https://en.wikipedia.org/wiki/Markdown)\n")
demo.launch()
```
| gradio-app/gradio/blob/main/demo/markdown_component/run.ipynb |
Using Adapter Transformers at Hugging Face
`adapter-transformers` is a library that extends 🤗 `transformers` by allowing to integrate, train and use Adapters and other efficient fine-tuning methods. The library is fully compatible with 🤗 `transformers`. Adapters are small learnt layers inserted within each layer of a pre-trained model. You can learn more about this in the [original paper](https://arxiv.org/abs/2007.07779).
## Exploring adapter-transformers in the Hub
You can find over a hundred `adapter-transformer` models by filtering at the left of the [models page](https://huggingface.co/models?library=adapter-transformers&sort=downloads). Some adapter models can be found in the Adapter Hub [repository](https://github.com/adapter-hub/hub). Models from both sources are then aggregated in the [AdapterHub](https://adapterhub.ml/explore/).
## Using existing models
For a full guide on loading pre-trained adapters, we recommend checking out the [official guide](https://docs.adapterhub.ml/loading.html).
As a brief summary, once you load a model with the usual `*Model` classes from 🤗`transformers`, you can use the `load_adapter` method to load and activate the Adapter (remember `adapter-transformers` extends 🤗`transformers`.).
```py
from transformers import AutoModelWithHeads
model = AutoModelWithHeads.from_pretrained("bert-base-uncased")
adapter_name = model.load_adapter("AdapterHub/bert-base-uncased-pf-imdb", source="hf")
model.active_adapters = adapter_name
```
You can also use `list_adapters` to find all Adapter Models programmatically
```py
from transformers import list_adapters
# source can be "ah" (AdapterHub), "hf" (hf.co) or None (for both, default)
adapter_infos = list_adapters(source="hf", model_name="bert-base-uncased")
```
If you want to see how to load a specific model, you can click `Use in Adapter Transformers` and you will be given a working snippet that you can load it!
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/libraries-adapter_transformers_snippet1.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/libraries-adapter_transformers-snippet1-dark.png"/>
</div>
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/libraries-adapter_transformers_snippet2.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/libraries-adapter_transformers-snippet2-dark.png"/>
</div>
## Sharing your models
For a full guide on sharing models with `adapter-transformers`, we recommend checking out the [official guide](https://docs.adapterhub.ml/huggingface_hub.html#uploading-to-the-hub).
You can share your Adapter by using the `push_adapter_to_hub` method from a model that already contains an adapter.
```py
model.push_adapter_to_hub(
"my-awesome-adapter",
"awesome_adapter",
adapterhub_tag="sentiment/imdb",
datasets_tag="imdb"
)
```
This command creates a repository with an automatically generated model card and all necessary metadata.
## Additional resources
* Adapter Transformers [library](https://github.com/adapter-hub/adapter-transformers).
* Adapter Transformers [docs](https://docs.adapterhub.ml/index.html).
* Integration with Hub [docs](https://docs.adapterhub.ml/huggingface_hub.html).
| huggingface/hub-docs/blob/main/docs/hub/adapter-transformers.md |
--
title: "Fine-Tune Wav2Vec2 for English ASR in Hugging Face with 🤗 Transformers"
thumbnail: /blog/assets/15_fine_tune_wav2vec2/wav2vec2.png
authors:
- user: patrickvonplaten
---
# Fine-Tune Wav2Vec2 for English ASR with 🤗 Transformers
<a target="_blank" href="https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Fine_tuning_Wav2Vec2_for_English_ASR.ipynb">
<img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
</a>
Wav2Vec2 is a pretrained model for Automatic Speech Recognition (ASR)
and was released in [September
2020](https://ai.facebook.com/blog/wav2vec-20-learning-the-structure-of-speech-from-raw-audio/)
by Alexei Baevski, Michael Auli, and Alex Conneau.
Using a novel contrastive pretraining objective, Wav2Vec2 learns
powerful speech representations from more than 50.000 hours of unlabeled
speech. Similar, to [BERT\'s masked language
modeling](http://jalammar.github.io/illustrated-bert/), the model learns
contextualized speech representations by randomly masking feature
vectors before passing them to a transformer network.
For the first time, it has been shown that pretraining, followed by
fine-tuning on very little labeled speech data achieves competitive
results to state-of-the-art ASR systems. Using as little as 10 minutes
of labeled data, Wav2Vec2 yields a word error rate (WER) of less than 5%
on the clean test set of
[LibriSpeech](https://huggingface.co/datasets/librispeech_asr) - *cf.*
with Table 9 of the [paper](https://arxiv.org/pdf/2006.11477.pdf).
In this notebook, we will give an in-detail explanation of how
Wav2Vec2\'s pretrained checkpoints can be fine-tuned on any English ASR
dataset. Note that in this notebook, we will fine-tune Wav2Vec2 without
making use of a language model. It is much simpler to use Wav2Vec2
without a language model as an end-to-end ASR system and it has been
shown that a standalone Wav2Vec2 acoustic model achieves impressive
results. For demonstration purposes, we fine-tune the \"base\"-sized
[pretrained checkpoint](https://huggingface.co/facebook/wav2vec2-base)
on the rather small [Timit](https://huggingface.co/datasets/timit_asr)
dataset that contains just 5h of training data.
Wav2Vec2 is fine-tuned using Connectionist Temporal Classification
(CTC), which is an algorithm that is used to train neural networks for
sequence-to-sequence problems and mainly in Automatic Speech Recognition
and handwriting recognition.
I highly recommend reading the blog post [Sequence Modeling with CTC
(2017)](https://distill.pub/2017/ctc/) very well-written blog post by
Awni Hannun.
Before we start, let\'s install both `datasets` and `transformers` from
master. Also, we need the `soundfile` package to load audio files and
the `jiwer` to evaluate our fine-tuned model using the [word error rate
(WER)](https://huggingface.co/metrics/wer) metric \\({}^1\\).
```bash
!pip install datasets>=1.18.3
!pip install transformers==4.11.3
!pip install librosa
!pip install jiwer
```
Next we strongly suggest to upload your training checkpoints directly to the [Hugging Face Hub](https://huggingface.co/) while training. The Hub has integrated version control so you can be sure that no model checkpoint is getting lost during training.
To do so you have to store your authentication token from the Hugging Face website (sign up [here](https://huggingface.co/join) if you haven't already!)
```python
from huggingface_hub import notebook_login
notebook_login()
```
**Print Output:**
```bash
Login successful
Your token has been saved to /root/.huggingface/token
Authenticated through git-crendential store but this isn't the helper defined on your machine.
You will have to re-authenticate when pushing to the Hugging Face Hub. Run the following command in your terminal to set it as the default
git config --global credential.helper store
```
Then you need to install Git-LFS to upload your model checkpoints:
```python
!apt install git-lfs
```
------------------------------------------------------------------------
\\({}^1\\) Timit is usually evaluated using the phoneme error rate (PER),
but by far the most common metric in ASR is the word error rate (WER).
To keep this notebook as general as possible we decided to evaluate the
model using WER.
Prepare Data, Tokenizer, Feature Extractor
------------------------------------------
ASR models transcribe speech to text, which means that we both need a
feature extractor that processes the speech signal to the model\'s input
format, *e.g.* a feature vector, and a tokenizer that processes the
model\'s output format to text.
In 🤗 Transformers, the Wav2Vec2 model is thus accompanied by both a
tokenizer, called
[Wav2Vec2CTCTokenizer](https://huggingface.co/transformers/master/model_doc/wav2vec2.html#wav2vec2ctctokenizer),
and a feature extractor, called
[Wav2Vec2FeatureExtractor](https://huggingface.co/transformers/master/model_doc/wav2vec2.html#wav2vec2featureextractor).
Let\'s start by creating the tokenizer responsible for decoding the
model\'s predictions.
### Create Wav2Vec2CTCTokenizer
The [pretrained Wav2Vec2 checkpoint](https://huggingface.co/facebook/wav2vec2-base) maps the speech signal to a
sequence of context representations as illustrated in the figure above.
A fine-tuned Wav2Vec2 checkpoint needs to map this sequence of context
representations to its corresponding transcription so that a linear
layer has to be added on top of the transformer block (shown in yellow).
This linear layer is used to classifies each context representation to a
token class analogous how, *e.g.*, after pretraining a linear layer is
added on top of BERT\'s embeddings for further classification - *cf.*
with *\"BERT\"* section of this [blog post](https://huggingface.co/blog/warm-starting-encoder-decoder).
The output size of this layer corresponds to the number of tokens in the
vocabulary, which does **not** depend on Wav2Vec2\'s pretraining task,
but only on the labeled dataset used for fine-tuning. So in the first
step, we will take a look at Timit and define a vocabulary based on the
dataset\'s transcriptions.
Let\'s start by loading the dataset and taking a look at its structure.
```python
from datasets import load_dataset, load_metric
timit = load_dataset("timit_asr")
print(timit)
```
**Print Output:**
```bash
DatasetDict({
train: Dataset({
features: ['file', 'audio', 'text', 'phonetic_detail', 'word_detail', 'dialect_region', 'sentence_type', 'speaker_id', 'id'],
num_rows: 4620
})
test: Dataset({
features: ['file', 'audio', 'text', 'phonetic_detail', 'word_detail', 'dialect_region', 'sentence_type', 'speaker_id', 'id'],
num_rows: 1680
})
})
```
Many ASR datasets only provide the target text, `'text'` for each audio
file `'file'`. Timit actually provides much more information about each
audio file, such as the `'phonetic_detail'`, etc., which is why many
researchers choose to evaluate their models on phoneme classification
instead of speech recognition when working with Timit. However, we want
to keep the notebook as general as possible, so that we will only
consider the transcribed text for fine-tuning.
```python
timit = timit.remove_columns(["phonetic_detail", "word_detail", "dialect_region", "id", "sentence_type", "speaker_id"])
```
Let\'s write a short function to display some random samples of the
dataset and run it a couple of times to get a feeling for the
transcriptions.
```python
from datasets import ClassLabel
import random
import pandas as pd
from IPython.display import display, HTML
def show_random_elements(dataset, num_examples=10):
assert num_examples <= len(dataset), "Can't pick more elements than there are in the dataset."
picks = []
for _ in range(num_examples):
pick = random.randint(0, len(dataset)-1)
while pick in picks:
pick = random.randint(0, len(dataset)-1)
picks.append(pick)
df = pd.DataFrame(dataset[picks])
display(HTML(df.to_html()))
show_random_elements(timit["train"].remove_columns(["file", "audio"]))
```
**Print Output:**
| Idx | Transcription |
|----------|:-------------:|
| 1 | Who took the kayak down the bayou? |
| 2 | As such it acts as an anchor for the people. |
| 3 | She had your dark suit in greasy wash water all year. |
| 4 | We're not drunkards, she said. |
| 5 | The most recent geological survey found seismic activity. |
| 6 | Alimony harms a divorced man's wealth. |
| 7 | Our entire economy will have a terrific uplift. |
| 8 | Don't ask me to carry an oily rag like that. |
| 9 | The gorgeous butterfly ate a lot of nectar. |
| 10 | Where're you takin' me? |
Alright! The transcriptions look very clean and the language seems to
correspond more to written text than dialogue. This makes sense taking
into account that [Timit](https://huggingface.co/datasets/timit_asr) is
a read speech corpus.
We can see that the transcriptions contain some special characters, such
as `,.?!;:`. Without a language model, it is much harder to classify
speech chunks to such special characters because they don\'t really
correspond to a characteristic sound unit. *E.g.*, the letter `"s"` has
a more or less clear sound, whereas the special character `"."` does
not. Also in order to understand the meaning of a speech signal, it is
usually not necessary to include special characters in the
transcription.
In addition, we normalize the text to only have lower case letters.
```python
import re
chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"]'
def remove_special_characters(batch):
batch["text"] = re.sub(chars_to_ignore_regex, '', batch["text"]).lower()
return batch
timit = timit.map(remove_special_characters)
```
Let's take a look at the preprocessed transcriptions.
```python
show_random_elements(timit["train"].remove_columns(["file", "audio"]))
```
**Print Output:**
| Idx | Transcription |
|----------|:-------------:|
| 1 | anyhow it was high time the boy was salted |
| 2 | their basis seems deeper than mere authority |
| 3 | only the best players enjoy popularity |
| 4 | tornados often destroy acres of farm land |
| 5 | where're you takin' me |
| 6 | soak up local color |
| 7 | satellites sputniks rockets balloons what next |
| 8 | i gave them several choices and let them set the priorities |
| 9 | reading in poor light gives you eyestrain |
| 10 | that dog chases cats mercilessly |
Good! This looks better. We have removed most special characters from
transcriptions and normalized them to lower-case only.
In CTC, it is common to classify speech chunks into letters, so we will
do the same here. Let\'s extract all distinct letters of the training
and test data and build our vocabulary from this set of letters.
We write a mapping function that concatenates all transcriptions into
one long transcription and then transforms the string into a set of
chars. It is important to pass the argument `batched=True` to the
`map(...)` function so that the mapping function has access to all
transcriptions at once.
```python
def extract_all_chars(batch):
all_text = " ".join(batch["text"])
vocab = list(set(all_text))
return {"vocab": [vocab], "all_text": [all_text]}
vocabs = timit.map(extract_all_chars, batched=True, batch_size=-1, keep_in_memory=True, remove_columns=timit.column_names["train"])
```
Now, we create the union of all distinct letters in the training dataset
and test dataset and convert the resulting list into an enumerated
dictionary.
```python
vocab_list = list(set(vocabs["train"]["vocab"][0]) | set(vocabs["test"]["vocab"][0]))
vocab_dict = {v: k for k, v in enumerate(vocab_list)}
vocab_dict
```
**Print Output:**
```bash
{
' ': 21,
"'": 13,
'a': 24,
'b': 17,
'c': 25,
'd': 2,
'e': 9,
'f': 14,
'g': 22,
'h': 8,
'i': 4,
'j': 18,
'k': 5,
'l': 16,
'm': 6,
'n': 7,
'o': 10,
'p': 19,
'q': 3,
'r': 20,
's': 11,
't': 0,
'u': 26,
'v': 27,
'w': 1,
'x': 23,
'y': 15,
'z': 12
}
```
Cool, we see that all letters of the alphabet occur in the dataset
(which is not really surprising) and we also extracted the special
characters `" "` and `'`. Note that we did not exclude those special
characters because:
- The model has to learn to predict when a word finished or else the
model prediction would always be a sequence of chars which would
make it impossible to separate words from each other.
- In English, we need to keep the `'` character to differentiate
between words, *e.g.*, `"it's"` and `"its"` which have very
different meanings.
To make it clearer that `" "` has its own token class, we give it a more
visible character `|`. In addition, we also add an \"unknown\" token so
that the model can later deal with characters not encountered in
Timit\'s training set.
```python
vocab_dict["|"] = vocab_dict[" "]
del vocab_dict[" "]
```
Finally, we also add a padding token that corresponds to CTC\'s \"*blank
token*\". The \"blank token\" is a core component of the CTC algorithm.
For more information, please take a look at the \"Alignment\" section
[here](https://distill.pub/2017/ctc/).
```python
vocab_dict["[UNK]"] = len(vocab_dict)
vocab_dict["[PAD]"] = len(vocab_dict)
print(len(vocab_dict))
```
**Print Output:**
```bash
30
```
Cool, now our vocabulary is complete and consists of 30 tokens, which
means that the linear layer that we will add on top of the pretrained
Wav2Vec2 checkpoint will have an output dimension of 30.
Let\'s now save the vocabulary as a json file.
```python
import json
with open('vocab.json', 'w') as vocab_file:
json.dump(vocab_dict, vocab_file)
```
In a final step, we use the json file to instantiate an object of the
`Wav2Vec2CTCTokenizer` class.
```python
from transformers import Wav2Vec2CTCTokenizer
tokenizer = Wav2Vec2CTCTokenizer("./vocab.json", unk_token="[UNK]", pad_token="[PAD]", word_delimiter_token="|")
```
If one wants to re-use the just created tokenizer with the fine-tuned model of this notebook, it is strongly advised to upload the `tokenizer` to the [🤗 Hub](https://huggingface.co/). Let's call the repo to which we will upload the files
`"wav2vec2-large-xlsr-turkish-demo-colab"`:
```python
repo_name = "wav2vec2-base-timit-demo-colab"
```
and upload the tokenizer to the [🤗 Hub](https://huggingface.co/).
```python
tokenizer.push_to_hub(repo_name)
```
Great, you can see the just created repository under `https://huggingface.co/<your-username>/wav2vec2-base-timit-demo-colab`
### Create Wav2Vec2 Feature Extractor
Speech is a continuous signal and to be treated by computers, it first
has to be discretized, which is usually called **sampling**. The
sampling rate hereby plays an important role in that it defines how many
data points of the speech signal are measured per second. Therefore,
sampling with a higher sampling rate results in a better approximation
of the *real* speech signal but also necessitates more values per
second.
A pretrained checkpoint expects its input data to have been sampled more
or less from the same distribution as the data it was trained on. The
same speech signals sampled at two different rates have a very different
distribution, *e.g.*, doubling the sampling rate results in data points
being twice as long. Thus, before fine-tuning a pretrained checkpoint of
an ASR model, it is crucial to verify that the sampling rate of the data
that was used to pretrain the model matches the sampling rate of the
dataset used to fine-tune the model.
Wav2Vec2 was pretrained on the audio data of
[LibriSpeech](https://huggingface.co/datasets/librispeech_asr) and
LibriVox which both were sampling with 16kHz. Our fine-tuning dataset,
[Timit](hhtps://huggingface.co/datasets/timit_asr), was luckily also
sampled with 16kHz. If the fine-tuning dataset would have been sampled
with a rate lower or higher than 16kHz, we first would have had to up or
downsample the speech signal to match the sampling rate of the data used
for pretraining.
A Wav2Vec2 feature extractor object requires the following parameters to
be instantiated:
- `feature_size`: Speech models take a sequence of feature vectors as
an input. While the length of this sequence obviously varies, the
feature size should not. In the case of Wav2Vec2, the feature size
is 1 because the model was trained on the raw speech signal \\({}^2\\) .
- `sampling_rate`: The sampling rate at which the model is trained on.
- `padding_value`: For batched inference, shorter inputs need to be
padded with a specific value
- `do_normalize`: Whether the input should be
*zero-mean-unit-variance* normalized or not. Usually, speech models
perform better when normalizing the input
- `return_attention_mask`: Whether the model should make use of an
`attention_mask` for batched inference. In general, models should
**always** make use of the `attention_mask` to mask padded tokens.
However, due to a very specific design choice of `Wav2Vec2`\'s
\"base\" checkpoint, better results are achieved when using no
`attention_mask`. This is **not** recommended for other speech
models. For more information, one can take a look at
[this](https://github.com/pytorch/fairseq/issues/3227) issue.
**Important** If you want to use this notebook to fine-tune
[large-lv60](https://huggingface.co/facebook/wav2vec2-large-lv60),
this parameter should be set to `True`.
```python
from transformers import Wav2Vec2FeatureExtractor
feature_extractor = Wav2Vec2FeatureExtractor(feature_size=1, sampling_rate=16000, padding_value=0.0, do_normalize=True, return_attention_mask=False)
```
Great, Wav2Vec2\'s feature extraction pipeline is thereby fully defined!
To make the usage of Wav2Vec2 as user-friendly as possible, the feature
extractor and tokenizer are *wrapped* into a single `Wav2Vec2Processor`
class so that one only needs a `model` and `processor` object.
```python
from transformers import Wav2Vec2Processor
processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)
```
### Preprocess Data
So far, we have not looked at the actual values of the speech signal but just the transcription. In addition to sentence, our datasets include two more column names path and audio. path states the absolute path of the audio file. Let's take a look.
```python
print(timit[0]["path"])
```
**Print Output:**
```bash
'/root/.cache/huggingface/datasets/downloads/extracted/404950a46da14eac65eb4e2a8317b1372fb3971d980d91d5d5b221275b1fd7e0/data/TRAIN/DR4/MMDM0/SI681.WAV'
```
**`Wav2Vec2`** expects the input in the format of a 1-dimensional array of 16 kHz. This means that the audio file has to be loaded and resampled.
Thankfully, datasets does this automatically by calling the other column audio. Let try it out.
```python
common_voice_train[0]["audio"]
```
**Print Output:**
```bash
{'array': array([-2.1362305e-04, 6.1035156e-05, 3.0517578e-05, ...,
-3.0517578e-05, -9.1552734e-05, -6.1035156e-05], dtype=float32),
'path': '/root/.cache/huggingface/datasets/downloads/extracted/404950a46da14eac65eb4e2a8317b1372fb3971d980d91d5d5b221275b1fd7e0/data/TRAIN/DR4/MMDM0/SI681.WAV',
'sampling_rate': 16000}
```
We can see that the audio file has automatically been loaded. This is thanks to the new [`"Audio" feature`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=audio#datasets.Audio) introduced in datasets == 4.13.3, which loads and resamples audio files on-the-fly upon calling.
The sampling rate is set to 16kHz which is what `Wav2Vec2` expects as an input.
Great, let's listen to a couple of audio files to better understand the dataset and verify that the audio was correctly loaded.
```python
import IPython.display as ipd
import numpy as np
import random
rand_int = random.randint(0, len(timit["train"]))
print(timit["train"][rand_int]["text"])
ipd.Audio(data=np.asarray(timit["train"][rand_int]["audio"]["array"]), autoplay=True, rate=16000)
```
It can be heard, that the speakers change along with their speaking rate, accent, etc. Overall, the recordings sound relatively clear though, which is to be expected from a read speech corpus.
Let's do a final check that the data is correctly prepared, by printing the shape of the speech input, its transcription, and the corresponding sampling rate.
```python
rand_int = random.randint(0, len(timit["train"]))
print("Target text:", timit["train"][rand_int]["text"])
print("Input array shape:", np.asarray(timit["train"][rand_int]["audio"]["array"]).shape)
print("Sampling rate:", timit["train"][rand_int]["audio"]["sampling_rate"])
```
**Print Output:**
```bash
Target text: she had your dark suit in greasy wash water all year
Input array shape: (52941,)
Sampling rate: 16000
```
Good! Everything looks fine - the data is a 1-dimensional array, the
sampling rate always corresponds to 16kHz, and the target text is
normalized.
Finally, we can process the dataset to the format expected by the model for training. We will make use of the `map(...)` function.
First, we load and resample the audio data, simply by calling `batch["audio"]`.
Second, we extract the `input_values` from the loaded audio file. In our case, the `Wav2Vec2Processor` only normalizes the data. For other speech models, however, this step can include more complex feature extraction, such as [Log-Mel feature extraction](https://en.wikipedia.org/wiki/Mel-frequency_cepstrum).
Third, we encode the transcriptions to label ids.
**Note**: This mapping function is a good example of how the `Wav2Vec2Processor` class should be used. In "normal" context, calling `processor(...)` is redirected to `Wav2Vec2FeatureExtractor`'s call method. When wrapping the processor into the `as_target_processor` context, however, the same method is redirected to `Wav2Vec2CTCTokenizer`'s call method.
For more information please check the [docs](https://huggingface.co/transformers/master/model_doc/wav2vec2.html#transformers.Wav2Vec2Processor.__call__).
```python
def prepare_dataset(batch):
audio = batch["audio"]
# batched output is "un-batched" to ensure mapping is correct
batch["input_values"] = processor(audio["array"], sampling_rate=audio["sampling_rate"]).input_values[0]
with processor.as_target_processor():
batch["labels"] = processor(batch["text"]).input_ids
return batch
```
Let's apply the data preparation function to all examples.
```python
timit = timit.map(prepare_dataset, remove_columns=timit.column_names["train"], num_proc=4)
```
**Note**: Currently `datasets` make use of [`torchaudio`](https://pytorch.org/audio/stable/index.html) and [`librosa`](https://librosa.org/doc/latest/index.html) for audio loading and resampling. If you wish to implement your own costumized data loading/sampling, feel free to just make use of the `"path"` column instead and disregard the `"audio"` column.
Training & Evaluation
---------------------
The data is processed so that we are ready to start setting up the
training pipeline. We will make use of 🤗\'s
[Trainer](https://huggingface.co/transformers/master/main_classes/trainer.html?highlight=trainer)
for which we essentially need to do the following:
- Define a data collator. In contrast to most NLP models, Wav2Vec2 has
a much larger input length than output length. *E.g.*, a sample of
input length 50000 has an output length of no more than 100. Given
the large input sizes, it is much more efficient to pad the training
batches dynamically meaning that all training samples should only be
padded to the longest sample in their batch and not the overall
longest sample. Therefore, fine-tuning Wav2Vec2 requires a special
padding data collator, which we will define below
- Evaluation metric. During training, the model should be evaluated on
the word error rate. We should define a `compute_metrics` function
accordingly
- Load a pretrained checkpoint. We need to load a pretrained
checkpoint and configure it correctly for training.
- Define the training configuration.
After having fine-tuned the model, we will correctly evaluate it on the
test data and verify that it has indeed learned to correctly transcribe
speech.
### Set-up Trainer
Let\'s start by defining the data collator. The code for the data
collator was copied from [this
example](https://github.com/huggingface/transformers/blob/7e61d56a45c19284cfda0cee8995fb552f6b1f4e/examples/pytorch/speech-recognition/run_speech_recognition_ctc.py#L219).
Without going into too many details, in contrast to the common data
collators, this data collator treats the `input_values` and `labels`
differently and thus applies to separate padding functions on them
(again making use of Wav2Vec2\'s context manager). This is necessary
because in speech input and output are of different modalities meaning
that they should not be treated by the same padding function. Analogous
to the common data collators, the padding tokens in the labels with
`-100` so that those tokens are **not** taken into account when
computing the loss.
```python
import torch
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Union
@dataclass
class DataCollatorCTCWithPadding:
"""
Data collator that will dynamically pad the inputs received.
Args:
processor (:class:`~transformers.Wav2Vec2Processor`)
The processor used for proccessing the data.
padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
among:
* :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
sequence if provided).
* :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
maximum acceptable input length for the model if that argument is not provided.
* :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
different lengths).
max_length (:obj:`int`, `optional`):
Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
max_length_labels (:obj:`int`, `optional`):
Maximum length of the ``labels`` returned list and optionally padding length (see above).
pad_to_multiple_of (:obj:`int`, `optional`):
If set will pad the sequence to a multiple of the provided value.
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
7.5 (Volta).
"""
processor: Wav2Vec2Processor
padding: Union[bool, str] = True
max_length: Optional[int] = None
max_length_labels: Optional[int] = None
pad_to_multiple_of: Optional[int] = None
pad_to_multiple_of_labels: Optional[int] = None
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
# split inputs and labels since they have to be of different lengths and need
# different padding methods
input_features = [{"input_values": feature["input_values"]} for feature in features]
label_features = [{"input_ids": feature["labels"]} for feature in features]
batch = self.processor.pad(
input_features,
padding=self.padding,
max_length=self.max_length,
pad_to_multiple_of=self.pad_to_multiple_of,
return_tensors="pt",
)
with self.processor.as_target_processor():
labels_batch = self.processor.pad(
label_features,
padding=self.padding,
max_length=self.max_length_labels,
pad_to_multiple_of=self.pad_to_multiple_of_labels,
return_tensors="pt",
)
# replace padding with -100 to ignore loss correctly
labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
batch["labels"] = labels
return batch
```
Let's initialize the data collator.
```python
data_collator = DataCollatorCTCWithPadding(processor=processor, padding=True)
```
Next, the evaluation metric is defined. As mentioned earlier, the
predominant metric in ASR is the word error rate (WER), hence we will
use it in this notebook as well.
```python
wer_metric = load_metric("wer")
```
The model will return a sequence of logit vectors:
$$ \mathbf{y}_1, \ldots, \mathbf{y}_m $$,
with \\(\mathbf{y}_1 = f_{\theta}(x_1, \ldots, x_n)[0]\\) and \\(n >> m\\).
A logit vector \\( \mathbf{y}_1 \\) contains the log-odds for each word in the
vocabulary we defined earlier, thus \\(\text{len}(\mathbf{y}_i) =\\)
`config.vocab_size`. We are interested in the most likely prediction of
the model and thus take the `argmax(...)` of the logits. Also, we
transform the encoded labels back to the original string by replacing
`-100` with the `pad_token_id` and decoding the ids while making sure
that consecutive tokens are **not** grouped to the same token in CTC
style \\({}^1\\).
```python
def compute_metrics(pred):
pred_logits = pred.predictions
pred_ids = np.argmax(pred_logits, axis=-1)
pred.label_ids[pred.label_ids == -100] = processor.tokenizer.pad_token_id
pred_str = processor.batch_decode(pred_ids)
# we do not want to group tokens when computing the metrics
label_str = processor.batch_decode(pred.label_ids, group_tokens=False)
wer = wer_metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
```
Now, we can load the pretrained `Wav2Vec2` checkpoint. The tokenizer\'s
`pad_token_id` must be to define the model\'s `pad_token_id` or in the
case of `Wav2Vec2ForCTC` also CTC\'s *blank token* \\({}^2\\). To save GPU
memory, we enable PyTorch\'s [gradient
checkpointing](https://pytorch.org/docs/stable/checkpoint.html) and also
set the loss reduction to \"*mean*\".
```python
from transformers import Wav2Vec2ForCTC
model = Wav2Vec2ForCTC.from_pretrained(
"facebook/wav2vec2-base",
ctc_loss_reduction="mean",
pad_token_id=processor.tokenizer.pad_token_id,
)
```
**Print Output:**
```bash
Some weights of Wav2Vec2ForCTC were not initialized from the model checkpoint at facebook/wav2vec2-base and are newly initialized: ['lm_head.weight', 'lm_head.bias']
You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
```
The first component of Wav2Vec2 consists of a stack of CNN layers that
are used to extract acoustically meaningful - but contextually
independent - features from the raw speech signal. This part of the
model has already been sufficiently trained during pretrainind and as
stated in the [paper](https://arxiv.org/abs/2006.11477) does not need to
be fine-tuned anymore. Thus, we can set the `requires_grad` to `False`
for all parameters of the *feature extraction* part.
```python
model.freeze_feature_extractor()
```
In a final step, we define all parameters related to training. To give
more explanation on some of the parameters:
- `group_by_length` makes training more efficient by grouping training
samples of similar input length into one batch. This can
significantly speed up training time by heavily reducing the overall
number of useless padding tokens that are passed through the model
- `learning_rate` and `weight_decay` were heuristically tuned until
fine-tuning has become stable. Note that those parameters strongly
depend on the Timit dataset and might be suboptimal for other speech
datasets.
For more explanations on other parameters, one can take a look at the
[docs](https://huggingface.co/transformers/master/main_classes/trainer.html?highlight=trainer#trainingarguments).
During training, a checkpoint will be uploaded asynchronously to the hub every 400 training steps. It allows you to also play around with the demo widget even while your model is still training.
**Note**: If one does not want to upload the model checkpoints to the hub, simply set `push_to_hub=False`.
```python
from transformers import TrainingArguments
training_args = TrainingArguments(
output_dir=repo_name,
group_by_length=True,
per_device_train_batch_size=32,
evaluation_strategy="steps",
num_train_epochs=30,
fp16=True,
gradient_checkpointing=True,
save_steps=500,
eval_steps=500,
logging_steps=500,
learning_rate=1e-4,
weight_decay=0.005,
warmup_steps=1000,
save_total_limit=2,
)
```
Now, all instances can be passed to Trainer and we are ready to start
training!
```python
from transformers import Trainer
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=timit_prepared["train"],
eval_dataset=timit_prepared["test"],
tokenizer=processor.feature_extractor,
)
```
------------------------------------------------------------------------
\\({}^1\\) To allow models to become independent of the speaker rate, in
CTC, consecutive tokens that are identical are simply grouped as a
single token. However, the encoded labels should not be grouped when
decoding since they don\'t correspond to the predicted tokens of the
model, which is why the `group_tokens=False` parameter has to be passed.
If we wouldn\'t pass this parameter a word like `"hello"` would
incorrectly be encoded, and decoded as `"helo"`.
\\({}^2\\) The blank token allows the model to predict a word, such as
`"hello"` by forcing it to insert the blank token between the two l\'s.
A CTC-conform prediction of `"hello"` of our model would be
`[PAD] [PAD] "h" "e" "e" "l" "l" [PAD] "l" "o" "o" [PAD]`.
### Training
Training will take between 90 and 180 minutes depending on the GPU
allocated to the google colab attached to this notebook. While the trained model yields satisfying
results on *Timit*\'s test data, it is by no means an optimally
fine-tuned model. The purpose of this notebook is to demonstrate how
Wav2Vec2\'s [base](https://huggingface.co/facebook/wav2vec2-base),
[large](https://huggingface.co/facebook/wav2vec2-large), and
[large-lv60](https://huggingface.co/facebook/wav2vec2-large-lv60)
checkpoints can be fine-tuned on any English dataset.
In case you want to use this google colab to fine-tune your model, you
should make sure that your training doesn\'t stop due to inactivity. A
simple hack to prevent this is to paste the following code into the
console of this tab (*right mouse click -\> inspect -\> Console tab and
insert code*).
```javascript
function ConnectButton(){
console.log("Connect pushed");
document.querySelector("#top-toolbar > colab-connect-button").shadowRoot.querySelector("#connect").click()
}
setInterval(ConnectButton,60000);
```
```python
trainer.train()
```
Depending on your GPU, it might be possible that you are seeing an `"out-of-memory"` error here. In this case, it's probably best to reduce `per_device_train_batch_size` to 16 or even less and eventually make use of [`gradient_accumulation`](https://huggingface.co/transformers/master/main_classes/trainer.html#trainingarguments).
**Print Output:**
| Step | Training Loss | Validation Loss | WER | Runtime | Samples per Second |
|---|---|---|---|---|---|
| 500 | 3.758100 | 1.686157 | 0.945214 | 97.299000 | 17.266000 |
| 1000 | 0.691400 | 0.476487 | 0.391427 | 98.283300 | 17.093000 |
| 1500 | 0.202400 | 0.403425 | 0.330715 | 99.078100 | 16.956000 |
| 2000 | 0.115200 | 0.405025 | 0.307353 | 98.116500 | 17.122000 |
| 2500 | 0.075000 | 0.428119 | 0.294053 | 98.496500 | 17.056000 |
| 3000 | 0.058200 | 0.442629 | 0.287299 | 98.871300 | 16.992000 |
| 3500 | 0.047600 | 0.442619 | 0.285783 | 99.477500 | 16.888000 |
| 4000 | 0.034500 | 0.456989 | 0.282200 | 99.419100 | 16.898000 |
The final WER should be below 0.3 which is reasonable given that
state-of-the-art phoneme error rates (PER) are just below 0.1 (see
[leaderboard](https://paperswithcode.com/sota/speech-recognition-on-timit))
and that WER is usually worse than PER.
You can now upload the result of the training to the Hub, just execute this instruction:
```python
trainer.push_to_hub()
```
You can now share this model with all your friends, family, favorite pets: they can all load it with the identifier "your-username/the-name-you-picked" so for instance:
```python
from transformers import AutoModelForCTC, Wav2Vec2Processor
model = AutoModelForCTC.from_pretrained("patrickvonplaten/wav2vec2-base-timit-demo-colab")
processor = Wav2Vec2Processor.from_pretrained("patrickvonplaten/wav2vec2-base-timit-demo-colab")
```
### Evaluation
In the final part, we evaluate our fine-tuned model on the test set and
play around with it a bit.
Let\'s load the `processor` and `model`.
```python
processor = Wav2Vec2Processor.from_pretrained(repo_name)
model = Wav2Vec2ForCTC.from_pretrained(repo_name)
```
Now, we will make use of the `map(...)` function to predict the
transcription of every test sample and to save the prediction in the
dataset itself. We will call the resulting dictionary `"results"`.
**Note**: we evaluate the test data set with `batch_size=1` on purpose
due to this [issue](https://github.com/pytorch/fairseq/issues/3227).
Since padded inputs don\'t yield the exact same output as non-padded
inputs, a better WER can be achieved by not padding the input at all.
```python
def map_to_result(batch):
with torch.no_grad():
input_values = torch.tensor(batch["input_values"], device="cuda").unsqueeze(0)
logits = model(input_values).logits
pred_ids = torch.argmax(logits, dim=-1)
batch["pred_str"] = processor.batch_decode(pred_ids)[0]
batch["text"] = processor.decode(batch["labels"], group_tokens=False)
return batch
results = timit["test"].map(map_to_result, remove_columns=timit["test"].column_names)
```
Let\'s compute the overall WER now.
```python
print("Test WER: {:.3f}".format(wer_metric.compute(predictions=results["pred_str"], references=results["text"])))
```
**Print Output:**
```bash
Test WER: 0.221
```
22.1% WER - not bad! Our demo model would have probably made it on the official [leaderboard](https://paperswithcode.com/sota/speech-recognition-on-timit).
Let's take a look at some predictions to see what errors are made by the model.
**Print Output:**
```python
show_random_elements(results.remove_columns(["speech", "sampling_rate"]))
```
| pred_str | target_text |
|----------|:-------------:|
| am to balence your employe you benefits package | aim to balance your employee benefit package |
| the fawlg prevented them from ariving on tom | the fog prevented them from arriving on time |
| young children should avoide exposure to contagieous diseases | young children should avoid exposure to contagious diseases |
| artifficial intelligence is for real | artificial intelligence is for real |
| their pcrops were two step latters a chair and a polmb fan | their props were two stepladders a chair and a palm fan |
| if people were more generous there would be no need for wealfare | if people were more generous there would be no need for welfare |
| the fish began to leep frantically on the surface of the small ac | the fish began to leap frantically on the surface of the small lake |
| her right hand eggs whenever the barametric pressur changes | her right hand aches whenever the barometric pressure changes |
| only lawyers loved miliunears | only lawyers love millionaires |
| the nearest cennagade may not be within wallkin distance | the nearest synagogue may not be within walking distance |
It becomes clear that the predicted transcriptions are acoustically very
similar to the target transcriptions, but often contain spelling or
grammatical errors. This shouldn\'t be very surprising though given that
we purely rely on Wav2Vec2 without making use of a language model.
Finally, to better understand how CTC works, it is worth taking a deeper
look at the exact output of the model. Let\'s run the first test sample
through the model, take the predicted ids and convert them to their
corresponding tokens.
```python
model.to("cuda")
with torch.no_grad():
logits = model(torch.tensor(timit["test"][:1]["input_values"], device="cuda")).logits
pred_ids = torch.argmax(logits, dim=-1)
# convert ids to tokens
" ".join(processor.tokenizer.convert_ids_to_tokens(pred_ids[0].tolist()))
```
**Print Output:**
```bash
[PAD] [PAD] [PAD] [PAD] [PAD] [PAD] t t h e e | | b b [PAD] u u n n n g g [PAD] a [PAD] [PAD] l l [PAD] o o o [PAD] | w w a a [PAD] s s | | [PAD] [PAD] p l l e e [PAD] [PAD] s s e n n t t t [PAD] l l y y | | | s s [PAD] i i [PAD] t t t [PAD] u u u u [PAD] [PAD] [PAD] a a [PAD] t t e e e d d d | n n e e a a a r | | t h h e | | s s h h h [PAD] o o o [PAD] o o r r [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD]
```
The output should make it a bit clearer how CTC works in practice. The
model is to some extent invariant to speaking rate since it has learned
to either just repeat the same token in case the speech chunk to be
classified still corresponds to the same token. This makes CTC a very
powerful algorithm for speech recognition since the speech file\'s
transcription is often very much independent of its length.
I again advise the reader to take a look at
[this](https://distill.pub/2017/ctc) very nice blog post to better
understand CTC.
| huggingface/blog/blob/main/fine-tune-wav2vec2-english.md |
Share a dataset using the CLI
At Hugging Face, we are on a mission to democratize good Machine Learning and we believe in the value of open source. That's why we designed 🤗 Datasets so that anyone can share a dataset with the greater ML community. There are currently thousands of datasets in over 100 languages in the Hugging Face Hub, and the Hugging Face team always welcomes new contributions!
Dataset repositories offer features such as:
- Free dataset hosting
- Dataset versioning
- Commit history and diffs
- Metadata for discoverability
- Dataset cards for documentation, licensing, limitations, etc.
This guide will show you how to share a dataset that can be easily accessed by anyone.
<a id='upload_dataset_repo'></a>
## Add a dataset
You can share your dataset with the community with a dataset repository on the Hugging Face Hub.
It can also be a private dataset if you want to control who has access to it.
In a dataset repository, you can host all your data files and [configure your dataset](./repository_structure#define-your-splits-in-yaml) to define which file goes to which split.
The following formats are supported: CSV, TSV, JSON, JSON lines, text, Parquet, Arrow, SQLite.
Many kinds of compressed file types are also supported: GZ, BZ2, LZ4, LZMA or ZSTD.
For example, your dataset can be made of `.json.gz` files.
On the other hand, if your dataset is not in a supported format or if you want more control over how your dataset is loaded, you can write your own dataset script.
When loading a dataset from the Hub, all the files in the supported formats are loaded, following the [repository structure](./repository_structure).
However if there's a dataset script, it is downloaded and executed to download and prepare the dataset instead.
For more information on how to load a dataset from the Hub, take a look at the [load a dataset from the Hub](./load_hub) tutorial.
### Create the repository
Sharing a community dataset will require you to create an account on [hf.co](https://huggingface.co/join) if you don't have one yet.
You can directly create a [new dataset repository](https://huggingface.co/login?next=%2Fnew-dataset) from your account on the Hugging Face Hub, but this guide will show you how to upload a dataset from the terminal.
1. Make sure you are in the virtual environment where you installed Datasets, and run the following command:
```
huggingface-cli login
```
2. Login using your Hugging Face Hub credentials, and create a new dataset repository:
```
huggingface-cli repo create your_dataset_name --type dataset
```
Add the `-organization` flag to create a repository under a specific organization:
```
huggingface-cli repo create your_dataset_name --type dataset --organization your-org-name
```
### Clone the repository
3. Install [Git LFS](https://git-lfs.github.com/) and clone your repository (refer to the [Git over SSH docs](https://huggingface.co/docs/hub/security-git-ssh) if you prefer cloning through SSH):
```
# Make sure you have git-lfs installed
# (https://git-lfs.github.com/)
git lfs install
git clone https://huggingface.co/datasets/namespace/your_dataset_name
```
Here the `namespace` is either your username or your organization name.
### Prepare your files
4. Now is a good time to check your directory to ensure the only files you're uploading are:
- The data files of the dataset
- The dataset card `README.md`
- (optional) `your_dataset_name.py` is your dataset loading script (optional if your data files are already in the supported formats csv/jsonl/json/parquet/txt). To create a dataset script, see the [dataset script](dataset_script) page.
### Upload your files
You can directly upload your files to your repository on the Hugging Face Hub, but this guide will show you how to upload the files from the terminal.
5. It is important to add the large data files first with `git lfs track` or else you will encounter an error later when you push your files:
```
cp /somewhere/data/*.json .
git lfs track *.json
git add .gitattributes
git add *.json
git commit -m "add json files"
```
6. (Optional) Add the dataset loading script:
```
cp /somewhere/data/load_script.py .
git add --all
```
7. Verify the files have been correctly staged. Then you can commit and push your files:
```
git status
git commit -m "First version of the your_dataset_name dataset."
git push
```
Congratulations, your dataset has now been uploaded to the Hugging Face Hub where anyone can load it in a single line of code! 🥳
```
dataset = load_dataset("namespace/your_dataset_name")
```
Finally, don't forget to enrich the dataset card to document your dataset and make it discoverable! Check out the [Create a dataset card](dataset_card) guide to learn more.
## Datasets on GitHub (legacy)
Datasets used to be hosted on our GitHub repository, but all datasets have now been migrated to the Hugging Face Hub.
The legacy GitHub datasets were added originally on our GitHub repository and therefore don't have a namespace on the Hub: "squad", "glue", etc. unlike the other datasets that are named "username/dataset_name" or "org/dataset_name".
<Tip>
The distinction between a Hub dataset within or without a namespace only comes from the legacy sharing workflow. It does not involve any ranking, decisioning, or opinion regarding the contents of the dataset itself.
</Tip>
Those datasets are now maintained on the Hub: if you think a fix is needed, please use their "Community" tab to open a discussion or create a Pull Request.
The code of these datasets is reviewed by the Hugging Face team.
| huggingface/datasets/blob/main/docs/source/share.mdx |
Community Examples
> **For more information about community pipelines, please have a look at [this issue](https://github.com/huggingface/diffusers/issues/841).**
**Community** examples consist of both inference and training examples that have been added by the community.
Please have a look at the following table to get an overview of all community examples. Click on the **Code Example** to get a copy-and-paste ready code example that you can try out.
If a community doesn't work as expected, please open an issue and ping the author on it.
| Example | Description | Code Example | Colab | Author |
|:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
| LLM-grounded Diffusion (LMD+) | LMD greatly improves the prompt following ability of text-to-image generation models by introducing an LLM as a front-end prompt parser and layout planner. [Project page.](https://llm-grounded-diffusion.github.io/) [See our full codebase (also with diffusers).](https://github.com/TonyLianLong/LLM-groundedDiffusion) | [LLM-grounded Diffusion (LMD+)](#llm-grounded-diffusion) | [Huggingface Demo](https://huggingface.co/spaces/longlian/llm-grounded-diffusion) [](https://colab.research.google.com/drive/1SXzMSeAB-LJYISb2yrUOdypLz4OYWUKj) | [Long (Tony) Lian](https://tonylian.com/) |
| CLIP Guided Stable Diffusion | Doing CLIP guidance for text to image generation with Stable Diffusion | [CLIP Guided Stable Diffusion](#clip-guided-stable-diffusion) | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/CLIP_Guided_Stable_diffusion_with_diffusers.ipynb) | [Suraj Patil](https://github.com/patil-suraj/) |
| One Step U-Net (Dummy) | Example showcasing of how to use Community Pipelines (see https://github.com/huggingface/diffusers/issues/841) | [One Step U-Net](#one-step-unet) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) |
| Stable Diffusion Interpolation | Interpolate the latent space of Stable Diffusion between different prompts/seeds | [Stable Diffusion Interpolation](#stable-diffusion-interpolation) | - | [Nate Raw](https://github.com/nateraw/) |
| Stable Diffusion Mega | **One** Stable Diffusion Pipeline with all functionalities of [Text2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py), [Image2Image](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) and [Inpainting](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | [Stable Diffusion Mega](#stable-diffusion-mega) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) |
| Long Prompt Weighting Stable Diffusion | **One** Stable Diffusion Pipeline without tokens length limit, and support parsing weighting in prompt. | [Long Prompt Weighting Stable Diffusion](#long-prompt-weighting-stable-diffusion) | - | [SkyTNT](https://github.com/SkyTNT) |
| Speech to Image | Using automatic-speech-recognition to transcribe text and Stable Diffusion to generate images | [Speech to Image](#speech-to-image) | - | [Mikail Duzenli](https://github.com/MikailINTech)
| Wild Card Stable Diffusion | Stable Diffusion Pipeline that supports prompts that contain wildcard terms (indicated by surrounding double underscores), with values instantiated randomly from a corresponding txt file or a dictionary of possible values | [Wildcard Stable Diffusion](#wildcard-stable-diffusion) | - | [Shyam Sudhakaran](https://github.com/shyamsn97) |
| [Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) | Stable Diffusion Pipeline that supports prompts that contain "|" in prompts (as an AND condition) and weights (separated by "|" as well) to positively / negatively weight prompts. | [Composable Stable Diffusion](#composable-stable-diffusion) | - | [Mark Rich](https://github.com/MarkRich) |
| Seed Resizing Stable Diffusion | Stable Diffusion Pipeline that supports resizing an image and retaining the concepts of the 512 by 512 generation. | [Seed Resizing](#seed-resizing) | - | [Mark Rich](https://github.com/MarkRich) |
| Imagic Stable Diffusion | Stable Diffusion Pipeline that enables writing a text prompt to edit an existing image | [Imagic Stable Diffusion](#imagic-stable-diffusion) | - | [Mark Rich](https://github.com/MarkRich) |
| Multilingual Stable Diffusion | Stable Diffusion Pipeline that supports prompts in 50 different languages. | [Multilingual Stable Diffusion](#multilingual-stable-diffusion-pipeline) | - | [Juan Carlos Piñeros](https://github.com/juancopi81) |
| Image to Image Inpainting Stable Diffusion | Stable Diffusion Pipeline that enables the overlaying of two images and subsequent inpainting | [Image to Image Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | - | [Alex McKinney](https://github.com/vvvm23) |
| Text Based Inpainting Stable Diffusion | Stable Diffusion Inpainting Pipeline that enables passing a text prompt to generate the mask for inpainting | [Text Based Inpainting Stable Diffusion](#image-to-image-inpainting-stable-diffusion) | - | [Dhruv Karan](https://github.com/unography) |
| Bit Diffusion | Diffusion on discrete data | [Bit Diffusion](#bit-diffusion) | - | [Stuti R.](https://github.com/kingstut) |
| K-Diffusion Stable Diffusion | Run Stable Diffusion with any of [K-Diffusion's samplers](https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py) | [Stable Diffusion with K Diffusion](#stable-diffusion-with-k-diffusion) | - | [Patrick von Platen](https://github.com/patrickvonplaten/) |
| Checkpoint Merger Pipeline | Diffusion Pipeline that enables merging of saved model checkpoints | [Checkpoint Merger Pipeline](#checkpoint-merger-pipeline) | - | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
Stable Diffusion v1.1-1.4 Comparison | Run all 4 model checkpoints for Stable Diffusion and compare their results together | [Stable Diffusion Comparison](#stable-diffusion-comparisons) | - | [Suvaditya Mukherjee](https://github.com/suvadityamuk) |
MagicMix | Diffusion Pipeline for semantic mixing of an image and a text prompt | [MagicMix](#magic-mix) | - | [Partho Das](https://github.com/daspartho) |
| Stable UnCLIP | Diffusion Pipeline for combining prior model (generate clip image embedding from text, UnCLIPPipeline `"kakaobrain/karlo-v1-alpha"`) and decoder pipeline (decode clip image embedding to image, StableDiffusionImageVariationPipeline `"lambdalabs/sd-image-variations-diffusers"` ). | [Stable UnCLIP](#stable-unclip) | - | [Ray Wang](https://wrong.wang) |
| UnCLIP Text Interpolation Pipeline | Diffusion Pipeline that allows passing two prompts and produces images while interpolating between the text-embeddings of the two prompts | [UnCLIP Text Interpolation Pipeline](#unclip-text-interpolation-pipeline) | - | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
| UnCLIP Image Interpolation Pipeline | Diffusion Pipeline that allows passing two images/image_embeddings and produces images while interpolating between their image-embeddings | [UnCLIP Image Interpolation Pipeline](#unclip-image-interpolation-pipeline) | - | [Naga Sai Abhinay Devarinti](https://github.com/Abhinay1997/) |
| DDIM Noise Comparative Analysis Pipeline | Investigating how the diffusion models learn visual concepts from each noise level (which is a contribution of [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227)) | [DDIM Noise Comparative Analysis Pipeline](#ddim-noise-comparative-analysis-pipeline) | - | [Aengus (Duc-Anh)](https://github.com/aengusng8) |
| CLIP Guided Img2Img Stable Diffusion Pipeline | Doing CLIP guidance for image to image generation with Stable Diffusion | [CLIP Guided Img2Img Stable Diffusion](#clip-guided-img2img-stable-diffusion) | - | [Nipun Jindal](https://github.com/nipunjindal/) |
| TensorRT Stable Diffusion Text to Image Pipeline | Accelerates the Stable Diffusion Text2Image Pipeline using TensorRT | [TensorRT Stable Diffusion Text to Image Pipeline](#tensorrt-text2image-stable-diffusion-pipeline) | - | [Asfiya Baig](https://github.com/asfiyab-nvidia) |
| EDICT Image Editing Pipeline | Diffusion pipeline for text-guided image editing | [EDICT Image Editing Pipeline](#edict-image-editing-pipeline) | - | [Joqsan Azocar](https://github.com/Joqsan) |
| Stable Diffusion RePaint | Stable Diffusion pipeline using [RePaint](https://arxiv.org/abs/2201.0986) for inpainting. | [Stable Diffusion RePaint](#stable-diffusion-repaint ) | - | [Markus Pobitzer](https://github.com/Markus-Pobitzer) |
| TensorRT Stable Diffusion Image to Image Pipeline | Accelerates the Stable Diffusion Image2Image Pipeline using TensorRT | [TensorRT Stable Diffusion Image to Image Pipeline](#tensorrt-image2image-stable-diffusion-pipeline) | - | [Asfiya Baig](https://github.com/asfiyab-nvidia) |
| Stable Diffusion IPEX Pipeline | Accelerate Stable Diffusion inference pipeline with BF16/FP32 precision on Intel Xeon CPUs with [IPEX](https://github.com/intel/intel-extension-for-pytorch) | [Stable Diffusion on IPEX](#stable-diffusion-on-ipex) | - | [Yingjie Han](https://github.com/yingjie-han/) |
| CLIP Guided Images Mixing Stable Diffusion Pipeline | Сombine images using usual diffusion models. | [CLIP Guided Images Mixing Using Stable Diffusion](#clip-guided-images-mixing-with-stable-diffusion) | - | [Karachev Denis](https://github.com/TheDenk) |
| TensorRT Stable Diffusion Inpainting Pipeline | Accelerates the Stable Diffusion Inpainting Pipeline using TensorRT | [TensorRT Stable Diffusion Inpainting Pipeline](#tensorrt-inpainting-stable-diffusion-pipeline) | - | [Asfiya Baig](https://github.com/asfiyab-nvidia) |
| IADB Pipeline | Implementation of [Iterative α-(de)Blending: a Minimalist Deterministic Diffusion Model](https://arxiv.org/abs/2305.03486) | [IADB Pipeline](#iadb-pipeline) | - | [Thomas Chambon](https://github.com/tchambon)
| Zero1to3 Pipeline | Implementation of [Zero-1-to-3: Zero-shot One Image to 3D Object](https://arxiv.org/abs/2303.11328) | [Zero1to3 Pipeline](#Zero1to3-pipeline) | - | [Xin Kong](https://github.com/kxhit) |
| Stable Diffusion XL Long Weighted Prompt Pipeline | A pipeline support unlimited length of prompt and negative prompt, use A1111 style of prompt weighting | [Stable Diffusion XL Long Weighted Prompt Pipeline](#stable-diffusion-xl-long-weighted-prompt-pipeline) | [](https://colab.research.google.com/drive/1LsqilswLR40XLLcp6XFOl5nKb_wOe26W?usp=sharing) | [Andrew Zhu](https://xhinker.medium.com/) |
FABRIC - Stable Diffusion with feedback Pipeline | pipeline supports feedback from liked and disliked images | [Stable Diffusion Fabric Pipeline](#stable-diffusion-fabric-pipeline) | - | [Shauray Singh](https://shauray8.github.io/about_shauray/) |
sketch inpaint - Inpainting with non-inpaint Stable Diffusion | sketch inpaint much like in automatic1111 | [Masked Im2Im Stable Diffusion Pipeline](#stable-diffusion-masked-im2im) | - | [Anatoly Belikov](https://github.com/noskill) |
prompt-to-prompt | change parts of a prompt and retain image structure (see [paper page](https://prompt-to-prompt.github.io/)) | [Prompt2Prompt Pipeline](#prompt2prompt-pipeline) | - | [Umer H. Adil](https://twitter.com/UmerHAdil) |
| Latent Consistency Pipeline | Implementation of [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378) | [Latent Consistency Pipeline](#latent-consistency-pipeline) | - | [Simian Luo](https://github.com/luosiallen) |
| Latent Consistency Img2img Pipeline | Img2img pipeline for Latent Consistency Models | [Latent Consistency Img2Img Pipeline](#latent-consistency-img2img-pipeline) | - | [Logan Zoellner](https://github.com/nagolinc) |
| Latent Consistency Interpolation Pipeline | Interpolate the latent space of Latent Consistency Models with multiple prompts | [Latent Consistency Interpolation Pipeline](#latent-consistency-interpolation-pipeline) | [](https://colab.research.google.com/drive/1pK3NrLWJSiJsBynLns1K1-IDTW9zbPvl?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) |
| SDE Drag Pipeline | The pipeline supports drag editing of images using stochastic differential equations | [SDE Drag Pipeline](#sde-drag-pipeline) | - | [NieShen](https://github.com/NieShenRuc) [Fengqi Zhu](https://github.com/Monohydroxides) |
| Regional Prompting Pipeline | Assign multiple prompts for different regions | [Regional Prompting Pipeline](#regional-prompting-pipeline) | - | [hako-mikan](https://github.com/hako-mikan) |
| LDM3D-sr (LDM3D upscaler) | Upscale low resolution RGB and depth inputs to high resolution | [StableDiffusionUpscaleLDM3D Pipeline](https://github.com/estelleafl/diffusers/tree/ldm3d_upscaler_community/examples/community#stablediffusionupscaleldm3d-pipeline) | - | [Estelle Aflalo](https://github.com/estelleafl) |
| AnimateDiff ControlNet Pipeline | Combines AnimateDiff with precise motion control using ControlNets | [AnimateDiff ControlNet Pipeline](#animatediff-controlnet-pipeline) | [](https://colab.research.google.com/drive/1SKboYeGjEQmQPWoFC0aLYpBlYdHXkvAu?usp=sharing) | [Aryan V S](https://github.com/a-r-r-o-w) and [Edoardo Botta](https://github.com/EdoardoBotta) |
| DemoFusion Pipeline | Implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973) | [DemoFusion Pipeline](#DemoFusion) | - | [Ruoyi Du](https://github.com/RuoyiDu) |
To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
```py
pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="filename_in_the_community_folder")
```
## Example usages
### LLM-grounded Diffusion
LMD and LMD+ greatly improves the prompt understanding ability of text-to-image generation models by introducing an LLM as a front-end prompt parser and layout planner. It improves spatial reasoning, the understanding of negation, attribute binding, generative numeracy, etc. in a unified manner without explicitly aiming for each. LMD is completely training-free (i.e., uses SD model off-the-shelf). LMD+ takes in additional adapters for better control. This is a reproduction of LMD+ model used in our work. [Project page.](https://llm-grounded-diffusion.github.io/) [See our full codebase (also with diffusers).](https://github.com/TonyLianLong/LLM-groundedDiffusion)
This pipeline can be used with an LLM or on its own. We provide a parser that parses LLM outputs to the layouts. You can obtain the prompt to input to the LLM for layout generation [here](https://github.com/TonyLianLong/LLM-groundedDiffusion/blob/main/prompt.py). After feeding the prompt to an LLM (e.g., GPT-4 on ChatGPT website), you can feed the LLM response into our pipeline.
The following code has been tested on 1x RTX 4090, but it should also support GPUs with lower GPU memory.
#### Use this pipeline with an LLM
```python
import torch
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"longlian/lmd_plus",
custom_pipeline="llm_grounded_diffusion",
custom_revision="main",
variant="fp16", torch_dtype=torch.float16
)
pipe.enable_model_cpu_offload()
# Generate directly from a text prompt and an LLM response
prompt = "a waterfall and a modern high speed train in a beautiful forest with fall foliage"
phrases, boxes, bg_prompt, neg_prompt = pipe.parse_llm_response("""
[('a waterfall', [71, 105, 148, 258]), ('a modern high speed train', [255, 223, 181, 149])]
Background prompt: A beautiful forest with fall foliage
Negative prompt:
""")
images = pipe(
prompt=prompt,
negative_prompt=neg_prompt,
phrases=phrases,
boxes=boxes,
gligen_scheduled_sampling_beta=0.4,
output_type="pil",
num_inference_steps=50,
lmd_guidance_kwargs={}
).images
images[0].save("./lmd_plus_generation.jpg")
```
#### Use this pipeline on its own for layout generation
```python
import torch
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"longlian/lmd_plus",
custom_pipeline="llm_grounded_diffusion",
variant="fp16", torch_dtype=torch.float16
)
pipe.enable_model_cpu_offload()
# Generate an image described by the prompt and
# insert objects described by text at the region defined by bounding boxes
prompt = "a waterfall and a modern high speed train in a beautiful forest with fall foliage"
boxes = [[0.1387, 0.2051, 0.4277, 0.7090], [0.4980, 0.4355, 0.8516, 0.7266]]
phrases = ["a waterfall", "a modern high speed train"]
images = pipe(
prompt=prompt,
phrases=phrases,
boxes=boxes,
gligen_scheduled_sampling_beta=0.4,
output_type="pil",
num_inference_steps=50,
lmd_guidance_kwargs={}
).images
images[0].save("./lmd_plus_generation.jpg")
```
### CLIP Guided Stable Diffusion
CLIP guided stable diffusion can help to generate more realistic images
by guiding stable diffusion at every denoising step with an additional CLIP model.
The following code requires roughly 12GB of GPU RAM.
```python
from diffusers import DiffusionPipeline
from transformers import CLIPImageProcessor, CLIPModel
import torch
feature_extractor = CLIPImageProcessor.from_pretrained("laion/CLIP-ViT-B-32-laion2B-s34B-b79K")
clip_model = CLIPModel.from_pretrained("laion/CLIP-ViT-B-32-laion2B-s34B-b79K", torch_dtype=torch.float16)
guided_pipeline = DiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
custom_pipeline="clip_guided_stable_diffusion",
clip_model=clip_model,
feature_extractor=feature_extractor,
torch_dtype=torch.float16,
)
guided_pipeline.enable_attention_slicing()
guided_pipeline = guided_pipeline.to("cuda")
prompt = "fantasy book cover, full moon, fantasy forest landscape, golden vector elements, fantasy magic, dark light night, intricate, elegant, sharp focus, illustration, highly detailed, digital painting, concept art, matte, art by WLOP and Artgerm and Albert Bierstadt, masterpiece"
generator = torch.Generator(device="cuda").manual_seed(0)
images = []
for i in range(4):
image = guided_pipeline(
prompt,
num_inference_steps=50,
guidance_scale=7.5,
clip_guidance_scale=100,
num_cutouts=4,
use_cutouts=False,
generator=generator,
).images[0]
images.append(image)
# save images locally
for i, img in enumerate(images):
img.save(f"./clip_guided_sd/image_{i}.png")
```
The `images` list contains a list of PIL images that can be saved locally or displayed directly in a google colab.
Generated images tend to be of higher qualtiy than natively using stable diffusion. E.g. the above script generates the following images:
.
### One Step Unet
The dummy "one-step-unet" can be run as follows:
```python
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="one_step_unet")
pipe()
```
**Note**: This community pipeline is not useful as a feature, but rather just serves as an example of how community pipelines can be added (see https://github.com/huggingface/diffusers/issues/841).
### Stable Diffusion Interpolation
The following code can be run on a GPU of at least 8GB VRAM and should take approximately 5 minutes.
```python
from diffusers import DiffusionPipeline
import torch
pipe = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
revision='fp16',
torch_dtype=torch.float16,
safety_checker=None, # Very important for videos...lots of false positives while interpolating
custom_pipeline="interpolate_stable_diffusion",
).to('cuda')
pipe.enable_attention_slicing()
frame_filepaths = pipe.walk(
prompts=['a dog', 'a cat', 'a horse'],
seeds=[42, 1337, 1234],
num_interpolation_steps=16,
output_dir='./dreams',
batch_size=4,
height=512,
width=512,
guidance_scale=8.5,
num_inference_steps=50,
)
```
The output of the `walk(...)` function returns a list of images saved under the folder as defined in `output_dir`. You can use these images to create videos of stable diffusion.
> **Please have a look at https://github.com/nateraw/stable-diffusion-videos for more in-detail information on how to create videos using stable diffusion as well as more feature-complete functionality.**
### Stable Diffusion Mega
The Stable Diffusion Mega Pipeline lets you use the main use cases of the stable diffusion pipeline in a single class.
```python
#!/usr/bin/env python3
from diffusers import DiffusionPipeline
import PIL
import requests
from io import BytesIO
import torch
def download_image(url):
response = requests.get(url)
return PIL.Image.open(BytesIO(response.content)).convert("RGB")
pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_mega", torch_dtype=torch.float16, revision="fp16")
pipe.to("cuda")
pipe.enable_attention_slicing()
### Text-to-Image
images = pipe.text2img("An astronaut riding a horse").images
### Image-to-Image
init_image = download_image("https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg")
prompt = "A fantasy landscape, trending on artstation"
images = pipe.img2img(prompt=prompt, image=init_image, strength=0.75, guidance_scale=7.5).images
### Inpainting
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
init_image = download_image(img_url).resize((512, 512))
mask_image = download_image(mask_url).resize((512, 512))
prompt = "a cat sitting on a bench"
images = pipe.inpaint(prompt=prompt, image=init_image, mask_image=mask_image, strength=0.75).images
```
As shown above this one pipeline can run all both "text-to-image", "image-to-image", and "inpainting" in one pipeline.
### Long Prompt Weighting Stable Diffusion
Features of this custom pipeline:
- Input a prompt without the 77 token length limit.
- Includes tx2img, img2img. and inpainting pipelines.
- Emphasize/weigh part of your prompt with parentheses as so: `a baby deer with (big eyes)`
- De-emphasize part of your prompt as so: `a [baby] deer with big eyes`
- Precisely weigh part of your prompt as so: `a baby deer with (big eyes:1.3)`
Prompt weighting equivalents:
- `a baby deer with` == `(a baby deer with:1.0)`
- `(big eyes)` == `(big eyes:1.1)`
- `((big eyes))` == `(big eyes:1.21)`
- `[big eyes]` == `(big eyes:0.91)`
You can run this custom pipeline as so:
#### pytorch
```python
from diffusers import DiffusionPipeline
import torch
pipe = DiffusionPipeline.from_pretrained(
'hakurei/waifu-diffusion',
custom_pipeline="lpw_stable_diffusion",
torch_dtype=torch.float16
)
pipe=pipe.to("cuda")
prompt = "best_quality (1girl:1.3) bow bride brown_hair closed_mouth frilled_bow frilled_hair_tubes frills (full_body:1.3) fox_ear hair_bow hair_tubes happy hood japanese_clothes kimono long_sleeves red_bow smile solo tabi uchikake white_kimono wide_sleeves cherry_blossoms"
neg_prompt = "lowres, bad_anatomy, error_body, error_hair, error_arm, error_hands, bad_hands, error_fingers, bad_fingers, missing_fingers, error_legs, bad_legs, multiple_legs, missing_legs, error_lighting, error_shadow, error_reflection, text, error, extra_digit, fewer_digits, cropped, worst_quality, low_quality, normal_quality, jpeg_artifacts, signature, watermark, username, blurry"
pipe.text2img(prompt, negative_prompt=neg_prompt, width=512,height=512,max_embeddings_multiples=3).images[0]
```
#### onnxruntime
```python
from diffusers import DiffusionPipeline
import torch
pipe = DiffusionPipeline.from_pretrained(
'CompVis/stable-diffusion-v1-4',
custom_pipeline="lpw_stable_diffusion_onnx",
revision="onnx",
provider="CUDAExecutionProvider"
)
prompt = "a photo of an astronaut riding a horse on mars, best quality"
neg_prompt = "lowres, bad anatomy, error body, error hair, error arm, error hands, bad hands, error fingers, bad fingers, missing fingers, error legs, bad legs, multiple legs, missing legs, error lighting, error shadow, error reflection, text, error, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry"
pipe.text2img(prompt,negative_prompt=neg_prompt, width=512, height=512, max_embeddings_multiples=3).images[0]
```
if you see `Token indices sequence length is longer than the specified maximum sequence length for this model ( *** > 77 ) . Running this sequence through the model will result in indexing errors`. Do not worry, it is normal.
### Speech to Image
The following code can generate an image from an audio sample using pre-trained OpenAI whisper-small and Stable Diffusion.
```Python
import torch
import matplotlib.pyplot as plt
from datasets import load_dataset
from diffusers import DiffusionPipeline
from transformers import (
WhisperForConditionalGeneration,
WhisperProcessor,
)
device = "cuda" if torch.cuda.is_available() else "cpu"
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
audio_sample = ds[3]
text = audio_sample["text"].lower()
speech_data = audio_sample["audio"]["array"]
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small").to(device)
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
diffuser_pipeline = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
custom_pipeline="speech_to_image_diffusion",
speech_model=model,
speech_processor=processor,
torch_dtype=torch.float16,
)
diffuser_pipeline.enable_attention_slicing()
diffuser_pipeline = diffuser_pipeline.to(device)
output = diffuser_pipeline(speech_data)
plt.imshow(output.images[0])
```
This example produces the following image:
### Wildcard Stable Diffusion
Following the great examples from https://github.com/jtkelm2/stable-diffusion-webui-1/blob/master/scripts/wildcards.py and https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts#wildcards, here's a minimal implementation that allows for users to add "wildcards", denoted by `__wildcard__` to prompts that are used as placeholders for randomly sampled values given by either a dictionary or a `.txt` file. For example:
Say we have a prompt:
```
prompt = "__animal__ sitting on a __object__ wearing a __clothing__"
```
We can then define possible values to be sampled for `animal`, `object`, and `clothing`. These can either be from a `.txt` with the same name as the category.
The possible values can also be defined / combined by using a dictionary like: `{"animal":["dog", "cat", mouse"]}`.
The actual pipeline works just like `StableDiffusionPipeline`, except the `__call__` method takes in:
`wildcard_files`: list of file paths for wild card replacement
`wildcard_option_dict`: dict with key as `wildcard` and values as a list of possible replacements
`num_prompt_samples`: number of prompts to sample, uniformly sampling wildcards
A full example:
create `animal.txt`, with contents like:
```
dog
cat
mouse
```
create `object.txt`, with contents like:
```
chair
sofa
bench
```
```python
from diffusers import DiffusionPipeline
import torch
pipe = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
custom_pipeline="wildcard_stable_diffusion",
torch_dtype=torch.float16,
)
prompt = "__animal__ sitting on a __object__ wearing a __clothing__"
out = pipe(
prompt,
wildcard_option_dict={
"clothing":["hat", "shirt", "scarf", "beret"]
},
wildcard_files=["object.txt", "animal.txt"],
num_prompt_samples=1
)
```
### Composable Stable diffusion
[Composable Stable Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/) proposes conjunction and negation (negative prompts) operators for compositional generation with conditional diffusion models.
```python
import torch as th
import numpy as np
import torchvision.utils as tvu
from diffusers import DiffusionPipeline
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--prompt", type=str, default="mystical trees | A magical pond | dark",
help="use '|' as the delimiter to compose separate sentences.")
parser.add_argument("--steps", type=int, default=50)
parser.add_argument("--scale", type=float, default=7.5)
parser.add_argument("--weights", type=str, default="7.5 | 7.5 | -7.5")
parser.add_argument("--seed", type=int, default=2)
parser.add_argument("--model_path", type=str, default="CompVis/stable-diffusion-v1-4")
parser.add_argument("--num_images", type=int, default=1)
args = parser.parse_args()
has_cuda = th.cuda.is_available()
device = th.device('cpu' if not has_cuda else 'cuda')
prompt = args.prompt
scale = args.scale
steps = args.steps
pipe = DiffusionPipeline.from_pretrained(
args.model_path,
custom_pipeline="composable_stable_diffusion",
).to(device)
pipe.safety_checker = None
images = []
generator = th.Generator("cuda").manual_seed(args.seed)
for i in range(args.num_images):
image = pipe(prompt, guidance_scale=scale, num_inference_steps=steps,
weights=args.weights, generator=generator).images[0]
images.append(th.from_numpy(np.array(image)).permute(2, 0, 1) / 255.)
grid = tvu.make_grid(th.stack(images, dim=0), nrow=4, padding=0)
tvu.save_image(grid, f'{prompt}_{args.weights}' + '.png')
```
### Imagic Stable Diffusion
Allows you to edit an image using stable diffusion.
```python
import requests
from PIL import Image
from io import BytesIO
import torch
import os
from diffusers import DiffusionPipeline, DDIMScheduler
has_cuda = torch.cuda.is_available()
device = torch.device('cpu' if not has_cuda else 'cuda')
pipe = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
safety_checker=None,
custom_pipeline="imagic_stable_diffusion",
scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
).to(device)
generator = torch.Generator("cuda").manual_seed(0)
seed = 0
prompt = "A photo of Barack Obama smiling with a big grin"
url = 'https://www.dropbox.com/s/6tlwzr73jd1r9yk/obama.png?dl=1'
response = requests.get(url)
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))
res = pipe.train(
prompt,
image=init_image,
generator=generator)
res = pipe(alpha=1, guidance_scale=7.5, num_inference_steps=50)
os.makedirs("imagic", exist_ok=True)
image = res.images[0]
image.save('./imagic/imagic_image_alpha_1.png')
res = pipe(alpha=1.5, guidance_scale=7.5, num_inference_steps=50)
image = res.images[0]
image.save('./imagic/imagic_image_alpha_1_5.png')
res = pipe(alpha=2, guidance_scale=7.5, num_inference_steps=50)
image = res.images[0]
image.save('./imagic/imagic_image_alpha_2.png')
```
### Seed Resizing
Test seed resizing. Originally generate an image in 512 by 512, then generate image with same seed at 512 by 592 using seed resizing. Finally, generate 512 by 592 using original stable diffusion pipeline.
```python
import torch as th
import numpy as np
from diffusers import DiffusionPipeline
has_cuda = th.cuda.is_available()
device = th.device('cpu' if not has_cuda else 'cuda')
pipe = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
custom_pipeline="seed_resize_stable_diffusion"
).to(device)
def dummy(images, **kwargs):
return images, False
pipe.safety_checker = dummy
images = []
th.manual_seed(0)
generator = th.Generator("cuda").manual_seed(0)
seed = 0
prompt = "A painting of a futuristic cop"
width = 512
height = 512
res = pipe(
prompt,
guidance_scale=7.5,
num_inference_steps=50,
height=height,
width=width,
generator=generator)
image = res.images[0]
image.save('./seed_resize/seed_resize_{w}_{h}_image.png'.format(w=width, h=height))
th.manual_seed(0)
generator = th.Generator("cuda").manual_seed(0)
pipe = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
custom_pipeline="/home/mark/open_source/diffusers/examples/community/"
).to(device)
width = 512
height = 592
res = pipe(
prompt,
guidance_scale=7.5,
num_inference_steps=50,
height=height,
width=width,
generator=generator)
image = res.images[0]
image.save('./seed_resize/seed_resize_{w}_{h}_image.png'.format(w=width, h=height))
pipe_compare = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
custom_pipeline="/home/mark/open_source/diffusers/examples/community/"
).to(device)
res = pipe_compare(
prompt,
guidance_scale=7.5,
num_inference_steps=50,
height=height,
width=width,
generator=generator
)
image = res.images[0]
image.save('./seed_resize/seed_resize_{w}_{h}_image_compare.png'.format(w=width, h=height))
```
### Multilingual Stable Diffusion Pipeline
The following code can generate an images from texts in different languages using the pre-trained [mBART-50 many-to-one multilingual machine translation model](https://huggingface.co/facebook/mbart-large-50-many-to-one-mmt) and Stable Diffusion.
```python
from PIL import Image
import torch
from diffusers import DiffusionPipeline
from transformers import (
pipeline,
MBart50TokenizerFast,
MBartForConditionalGeneration,
)
device = "cuda" if torch.cuda.is_available() else "cpu"
device_dict = {"cuda": 0, "cpu": -1}
# helper function taken from: https://huggingface.co/blog/stable_diffusion
def image_grid(imgs, rows, cols):
assert len(imgs) == rows*cols
w, h = imgs[0].size
grid = Image.new('RGB', size=(cols*w, rows*h))
grid_w, grid_h = grid.size
for i, img in enumerate(imgs):
grid.paste(img, box=(i%cols*w, i//cols*h))
return grid
# Add language detection pipeline
language_detection_model_ckpt = "papluca/xlm-roberta-base-language-detection"
language_detection_pipeline = pipeline("text-classification",
model=language_detection_model_ckpt,
device=device_dict[device])
# Add model for language translation
trans_tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-one-mmt")
trans_model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-one-mmt").to(device)
diffuser_pipeline = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
custom_pipeline="multilingual_stable_diffusion",
detection_pipeline=language_detection_pipeline,
translation_model=trans_model,
translation_tokenizer=trans_tokenizer,
torch_dtype=torch.float16,
)
diffuser_pipeline.enable_attention_slicing()
diffuser_pipeline = diffuser_pipeline.to(device)
prompt = ["a photograph of an astronaut riding a horse",
"Una casa en la playa",
"Ein Hund, der Orange isst",
"Un restaurant parisien"]
output = diffuser_pipeline(prompt)
images = output.images
grid = image_grid(images, rows=2, cols=2)
```
This example produces the following images:
### Image to Image Inpainting Stable Diffusion
Similar to the standard stable diffusion inpainting example, except with the addition of an `inner_image` argument.
`image`, `inner_image`, and `mask` should have the same dimensions. `inner_image` should have an alpha (transparency) channel.
The aim is to overlay two images, then mask out the boundary between `image` and `inner_image` to allow stable diffusion to make the connection more seamless.
For example, this could be used to place a logo on a shirt and make it blend seamlessly.
```python
import PIL
import torch
from diffusers import DiffusionPipeline
image_path = "./path-to-image.png"
inner_image_path = "./path-to-inner-image.png"
mask_path = "./path-to-mask.png"
init_image = PIL.Image.open(image_path).convert("RGB").resize((512, 512))
inner_image = PIL.Image.open(inner_image_path).convert("RGBA").resize((512, 512))
mask_image = PIL.Image.open(mask_path).convert("RGB").resize((512, 512))
pipe = DiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-inpainting",
custom_pipeline="img2img_inpainting",
torch_dtype=torch.float16
)
pipe = pipe.to("cuda")
prompt = "Your prompt here!"
image = pipe(prompt=prompt, image=init_image, inner_image=inner_image, mask_image=mask_image).images[0]
```
### Text Based Inpainting Stable Diffusion
Use a text prompt to generate the mask for the area to be inpainted.
Currently uses the CLIPSeg model for mask generation, then calls the standard Stable Diffusion Inpainting pipeline to perform the inpainting.
```python
from transformers import CLIPSegProcessor, CLIPSegForImageSegmentation
from diffusers import DiffusionPipeline
from PIL import Image
import requests
processor = CLIPSegProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
pipe = DiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-inpainting",
custom_pipeline="text_inpainting",
segmentation_model=model,
segmentation_processor=processor
)
pipe = pipe.to("cuda")
url = "https://github.com/timojl/clipseg/blob/master/example_image.jpg?raw=true"
image = Image.open(requests.get(url, stream=True).raw).resize((512, 512))
text = "a glass" # will mask out this text
prompt = "a cup" # the masked out region will be replaced with this
image = pipe(image=image, text=text, prompt=prompt).images[0]
```
### Bit Diffusion
Based https://arxiv.org/abs/2208.04202, this is used for diffusion on discrete data - eg, discreate image data, DNA sequence data. An unconditional discreate image can be generated like this:
```python
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("google/ddpm-cifar10-32", custom_pipeline="bit_diffusion")
image = pipe().images[0]
```
### Stable Diffusion with K Diffusion
Make sure you have @crowsonkb's https://github.com/crowsonkb/k-diffusion installed:
```
pip install k-diffusion
```
You can use the community pipeline as follows:
```python
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="sd_text2img_k_diffusion")
pipe = pipe.to("cuda")
prompt = "an astronaut riding a horse on mars"
pipe.set_scheduler("sample_heun")
generator = torch.Generator(device="cuda").manual_seed(seed)
image = pipe(prompt, generator=generator, num_inference_steps=20).images[0]
image.save("./astronaut_heun_k_diffusion.png")
```
To make sure that K Diffusion and `diffusers` yield the same results:
**Diffusers**:
```python
from diffusers import DiffusionPipeline, EulerDiscreteScheduler
seed = 33
pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4")
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to("cuda")
generator = torch.Generator(device="cuda").manual_seed(seed)
image = pipe(prompt, generator=generator, num_inference_steps=50).images[0]
```
**K Diffusion**:
```python
from diffusers import DiffusionPipeline, EulerDiscreteScheduler
seed = 33
pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="sd_text2img_k_diffusion")
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to("cuda")
pipe.set_scheduler("sample_euler")
generator = torch.Generator(device="cuda").manual_seed(seed)
image = pipe(prompt, generator=generator, num_inference_steps=50).images[0]
```
### Checkpoint Merger Pipeline
Based on the AUTOMATIC1111/webui for checkpoint merging. This is a custom pipeline that merges upto 3 pretrained model checkpoints as long as they are in the HuggingFace model_index.json format.
The checkpoint merging is currently memory intensive as it modifies the weights of a DiffusionPipeline object in place. Expect atleast 13GB RAM Usage on Kaggle GPU kernels and
on colab you might run out of the 12GB memory even while merging two checkpoints.
Usage:-
```python
from diffusers import DiffusionPipeline
#Return a CheckpointMergerPipeline class that allows you to merge checkpoints.
#The checkpoint passed here is ignored. But still pass one of the checkpoints you plan to
#merge for convenience
pipe = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", custom_pipeline="checkpoint_merger")
#There are multiple possible scenarios:
#The pipeline with the merged checkpoints is returned in all the scenarios
#Compatible checkpoints a.k.a matched model_index.json files. Ignores the meta attributes in model_index.json during comparison.( attrs with _ as prefix )
merged_pipe = pipe.merge(["CompVis/stable-diffusion-v1-4","CompVis/stable-diffusion-v1-2"], interp = "sigmoid", alpha = 0.4)
#Incompatible checkpoints in model_index.json but merge might be possible. Use force = True to ignore model_index.json compatibility
merged_pipe_1 = pipe.merge(["CompVis/stable-diffusion-v1-4","hakurei/waifu-diffusion"], force = True, interp = "sigmoid", alpha = 0.4)
#Three checkpoint merging. Only "add_difference" method actually works on all three checkpoints. Using any other options will ignore the 3rd checkpoint.
merged_pipe_2 = pipe.merge(["CompVis/stable-diffusion-v1-4","hakurei/waifu-diffusion","prompthero/openjourney"], force = True, interp = "add_difference", alpha = 0.4)
prompt = "An astronaut riding a horse on Mars"
image = merged_pipe(prompt).images[0]
```
Some examples along with the merge details:
1. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" ; Sigmoid interpolation; alpha = 0.8
2. "hakurei/waifu-diffusion" + "prompthero/openjourney" ; Inverse Sigmoid interpolation; alpha = 0.8
3. "CompVis/stable-diffusion-v1-4" + "hakurei/waifu-diffusion" + "prompthero/openjourney"; Add Difference interpolation; alpha = 0.5
### Stable Diffusion Comparisons
This Community Pipeline enables the comparison between the 4 checkpoints that exist for Stable Diffusion. They can be found through the following links:
1. [Stable Diffusion v1.1](https://huggingface.co/CompVis/stable-diffusion-v1-1)
2. [Stable Diffusion v1.2](https://huggingface.co/CompVis/stable-diffusion-v1-2)
3. [Stable Diffusion v1.3](https://huggingface.co/CompVis/stable-diffusion-v1-3)
4. [Stable Diffusion v1.4](https://huggingface.co/CompVis/stable-diffusion-v1-4)
```python
from diffusers import DiffusionPipeline
import matplotlib.pyplot as plt
pipe = DiffusionPipeline.from_pretrained('CompVis/stable-diffusion-v1-4', custom_pipeline='suvadityamuk/StableDiffusionComparison')
pipe.enable_attention_slicing()
pipe = pipe.to('cuda')
prompt = "an astronaut riding a horse on mars"
output = pipe(prompt)
plt.subplots(2,2,1)
plt.imshow(output.images[0])
plt.title('Stable Diffusion v1.1')
plt.axis('off')
plt.subplots(2,2,2)
plt.imshow(output.images[1])
plt.title('Stable Diffusion v1.2')
plt.axis('off')
plt.subplots(2,2,3)
plt.imshow(output.images[2])
plt.title('Stable Diffusion v1.3')
plt.axis('off')
plt.subplots(2,2,4)
plt.imshow(output.images[3])
plt.title('Stable Diffusion v1.4')
plt.axis('off')
plt.show()
```
As a result, you can look at a grid of all 4 generated images being shown together, that captures a difference the advancement of the training between the 4 checkpoints.
### Magic Mix
Implementation of the [MagicMix: Semantic Mixing with Diffusion Models](https://arxiv.org/abs/2210.16056) paper. This is a Diffusion Pipeline for semantic mixing of an image and a text prompt to create a new concept while preserving the spatial layout and geometry of the subject in the image. The pipeline takes an image that provides the layout semantics and a prompt that provides the content semantics for the mixing process.
There are 3 parameters for the method-
- `mix_factor`: It is the interpolation constant used in the layout generation phase. The greater the value of `mix_factor`, the greater the influence of the prompt on the layout generation process.
- `kmax` and `kmin`: These determine the range for the layout and content generation process. A higher value of kmax results in loss of more information about the layout of the original image and a higher value of kmin results in more steps for content generation process.
Here is an example usage-
```python
from diffusers import DiffusionPipeline, DDIMScheduler
from PIL import Image
pipe = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
custom_pipeline="magic_mix",
scheduler = DDIMScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler"),
).to('cuda')
img = Image.open('phone.jpg')
mix_img = pipe(
img,
prompt = 'bed',
kmin = 0.3,
kmax = 0.5,
mix_factor = 0.5,
)
mix_img.save('phone_bed_mix.jpg')
```
The `mix_img` is a PIL image that can be saved locally or displayed directly in a google colab. Generated image is a mix of the layout semantics of the given image and the content semantics of the prompt.
E.g. the above script generates the following image:
`phone.jpg`
`phone_bed_mix.jpg`
For more example generations check out this [demo notebook](https://github.com/daspartho/MagicMix/blob/main/demo.ipynb).
### Stable UnCLIP
UnCLIPPipeline("kakaobrain/karlo-v1-alpha") provide a prior model that can generate clip image embedding from text.
StableDiffusionImageVariationPipeline("lambdalabs/sd-image-variations-diffusers") provide a decoder model than can generate images from clip image embedding.
```python
import torch
from diffusers import DiffusionPipeline
device = torch.device("cpu" if not torch.cuda.is_available() else "cuda")
pipeline = DiffusionPipeline.from_pretrained(
"kakaobrain/karlo-v1-alpha",
torch_dtype=torch.float16,
custom_pipeline="stable_unclip",
decoder_pipe_kwargs=dict(
image_encoder=None,
),
)
pipeline.to(device)
prompt = "a shiba inu wearing a beret and black turtleneck"
random_generator = torch.Generator(device=device).manual_seed(1000)
output = pipeline(
prompt=prompt,
width=512,
height=512,
generator=random_generator,
prior_guidance_scale=4,
prior_num_inference_steps=25,
decoder_guidance_scale=8,
decoder_num_inference_steps=50,
)
image = output.images[0]
image.save("./shiba-inu.jpg")
# debug
# `pipeline.decoder_pipe` is a regular StableDiffusionImageVariationPipeline instance.
# It is used to convert clip image embedding to latents, then fed into VAE decoder.
print(pipeline.decoder_pipe.__class__)
# <class 'diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_image_variation.StableDiffusionImageVariationPipeline'>
# this pipeline only use prior module in "kakaobrain/karlo-v1-alpha"
# It is used to convert clip text embedding to clip image embedding.
print(pipeline)
# StableUnCLIPPipeline {
# "_class_name": "StableUnCLIPPipeline",
# "_diffusers_version": "0.12.0.dev0",
# "prior": [
# "diffusers",
# "PriorTransformer"
# ],
# "prior_scheduler": [
# "diffusers",
# "UnCLIPScheduler"
# ],
# "text_encoder": [
# "transformers",
# "CLIPTextModelWithProjection"
# ],
# "tokenizer": [
# "transformers",
# "CLIPTokenizer"
# ]
# }
# pipeline.prior_scheduler is the scheduler used for prior in UnCLIP.
print(pipeline.prior_scheduler)
# UnCLIPScheduler {
# "_class_name": "UnCLIPScheduler",
# "_diffusers_version": "0.12.0.dev0",
# "clip_sample": true,
# "clip_sample_range": 5.0,
# "num_train_timesteps": 1000,
# "prediction_type": "sample",
# "variance_type": "fixed_small_log"
# }
```
`shiba-inu.jpg`
### UnCLIP Text Interpolation Pipeline
This Diffusion Pipeline takes two prompts and interpolates between the two input prompts using spherical interpolation ( slerp ). The input prompts are converted to text embeddings by the pipeline's text_encoder and the interpolation is done on the resulting text_embeddings over the number of steps specified. Defaults to 5 steps.
```python
import torch
from diffusers import DiffusionPipeline
device = torch.device("cpu" if not torch.cuda.is_available() else "cuda")
pipe = DiffusionPipeline.from_pretrained(
"kakaobrain/karlo-v1-alpha",
torch_dtype=torch.float16,
custom_pipeline="unclip_text_interpolation"
)
pipe.to(device)
start_prompt = "A photograph of an adult lion"
end_prompt = "A photograph of a lion cub"
#For best results keep the prompts close in length to each other. Of course, feel free to try out with differing lengths.
generator = torch.Generator(device=device).manual_seed(42)
output = pipe(start_prompt, end_prompt, steps = 6, generator = generator, enable_sequential_cpu_offload=False)
for i,image in enumerate(output.images):
img.save('result%s.jpg' % i)
```
The resulting images in order:-
### UnCLIP Image Interpolation Pipeline
This Diffusion Pipeline takes two images or an image_embeddings tensor of size 2 and interpolates between their embeddings using spherical interpolation ( slerp ). The input images/image_embeddings are converted to image embeddings by the pipeline's image_encoder and the interpolation is done on the resulting image_embeddings over the number of steps specified. Defaults to 5 steps.
```python
import torch
from diffusers import DiffusionPipeline
from PIL import Image
device = torch.device("cpu" if not torch.cuda.is_available() else "cuda")
dtype = torch.float16 if torch.cuda.is_available() else torch.bfloat16
pipe = DiffusionPipeline.from_pretrained(
"kakaobrain/karlo-v1-alpha-image-variations",
torch_dtype=dtype,
custom_pipeline="unclip_image_interpolation"
)
pipe.to(device)
images = [Image.open('./starry_night.jpg'), Image.open('./flowers.jpg')]
#For best results keep the prompts close in length to each other. Of course, feel free to try out with differing lengths.
generator = torch.Generator(device=device).manual_seed(42)
output = pipe(image = images ,steps = 6, generator = generator)
for i,image in enumerate(output.images):
image.save('starry_to_flowers_%s.jpg' % i)
```
The original images:-
The resulting images in order:-
### DDIM Noise Comparative Analysis Pipeline
#### **Research question: What visual concepts do the diffusion models learn from each noise level during training?**
The [P2 weighting (CVPR 2022)](https://arxiv.org/abs/2204.00227) paper proposed an approach to answer the above question, which is their second contribution.
The approach consists of the following steps:
1. The input is an image x0.
2. Perturb it to xt using a diffusion process q(xt|x0).
- `strength` is a value between 0.0 and 1.0, that controls the amount of noise that is added to the input image. Values that approach 1.0 allow for lots of variations but will also produce images that are not semantically consistent with the input.
3. Reconstruct the image with the learned denoising process pθ(ˆx0|xt).
4. Compare x0 and ˆx0 among various t to show how each step contributes to the sample.
The authors used [openai/guided-diffusion](https://github.com/openai/guided-diffusion) model to denoise images in FFHQ dataset. This pipeline extends their second contribution by investigating DDIM on any input image.
```python
import torch
from PIL import Image
import numpy as np
image_path = "path/to/your/image" # images from CelebA-HQ might be better
image_pil = Image.open(image_path)
image_name = image_path.split("/")[-1].split(".")[0]
device = torch.device("cpu" if not torch.cuda.is_available() else "cuda")
pipe = DiffusionPipeline.from_pretrained(
"google/ddpm-ema-celebahq-256",
custom_pipeline="ddim_noise_comparative_analysis",
)
pipe = pipe.to(device)
for strength in np.linspace(0.1, 1, 25):
denoised_image, latent_timestep = pipe(
image_pil, strength=strength, return_dict=False
)
denoised_image = denoised_image[0]
denoised_image.save(
f"noise_comparative_analysis_{image_name}_{latent_timestep}.png"
)
```
Here is the result of this pipeline (which is DDIM) on CelebA-HQ dataset.
### CLIP Guided Img2Img Stable Diffusion
CLIP guided Img2Img stable diffusion can help to generate more realistic images with an initial image
by guiding stable diffusion at every denoising step with an additional CLIP model.
The following code requires roughly 12GB of GPU RAM.
```python
from io import BytesIO
import requests
import torch
from diffusers import DiffusionPipeline
from PIL import Image
from transformers import CLIPFeatureExtractor, CLIPModel
feature_extractor = CLIPFeatureExtractor.from_pretrained(
"laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
)
clip_model = CLIPModel.from_pretrained(
"laion/CLIP-ViT-B-32-laion2B-s34B-b79K", torch_dtype=torch.float16
)
guided_pipeline = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
# custom_pipeline="clip_guided_stable_diffusion",
custom_pipeline="/home/njindal/diffusers/examples/community/clip_guided_stable_diffusion.py",
clip_model=clip_model,
feature_extractor=feature_extractor,
torch_dtype=torch.float16,
)
guided_pipeline.enable_attention_slicing()
guided_pipeline = guided_pipeline.to("cuda")
prompt = "fantasy book cover, full moon, fantasy forest landscape, golden vector elements, fantasy magic, dark light night, intricate, elegant, sharp focus, illustration, highly detailed, digital painting, concept art, matte, art by WLOP and Artgerm and Albert Bierstadt, masterpiece"
url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
response = requests.get(url)
init_image = Image.open(BytesIO(response.content)).convert("RGB")
image = guided_pipeline(
prompt=prompt,
num_inference_steps=30,
image=init_image,
strength=0.75,
guidance_scale=7.5,
clip_guidance_scale=100,
num_cutouts=4,
use_cutouts=False,
).images[0]
display(image)
```
Init Image
Output Image
### TensorRT Text2Image Stable Diffusion Pipeline
The TensorRT Pipeline can be used to accelerate the Text2Image Stable Diffusion Inference run.
NOTE: The ONNX conversions and TensorRT engine build may take up to 30 minutes.
```python
import torch
from diffusers import DDIMScheduler
from diffusers.pipelines.stable_diffusion import StableDiffusionPipeline
# Use the DDIMScheduler scheduler here instead
scheduler = DDIMScheduler.from_pretrained("stabilityai/stable-diffusion-2-1",
subfolder="scheduler")
pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1",
custom_pipeline="stable_diffusion_tensorrt_txt2img",
revision='fp16',
torch_dtype=torch.float16,
scheduler=scheduler,)
# re-use cached folder to save ONNX models and TensorRT Engines
pipe.set_cached_folder("stabilityai/stable-diffusion-2-1", revision='fp16',)
pipe = pipe.to("cuda")
prompt = "a beautiful photograph of Mt. Fuji during cherry blossom"
image = pipe(prompt).images[0]
image.save('tensorrt_mt_fuji.png')
```
### EDICT Image Editing Pipeline
This pipeline implements the text-guided image editing approach from the paper [EDICT: Exact Diffusion Inversion via Coupled Transformations](https://arxiv.org/abs/2211.12446). You have to pass:
- (`PIL`) `image` you want to edit.
- `base_prompt`: the text prompt describing the current image (before editing).
- `target_prompt`: the text prompt describing with the edits.
```python
from diffusers import DiffusionPipeline, DDIMScheduler
from transformers import CLIPTextModel
import torch, PIL, requests
from io import BytesIO
from IPython.display import display
def center_crop_and_resize(im):
width, height = im.size
d = min(width, height)
left = (width - d) / 2
upper = (height - d) / 2
right = (width + d) / 2
lower = (height + d) / 2
return im.crop((left, upper, right, lower)).resize((512, 512))
torch_dtype = torch.float16
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# scheduler and text_encoder param values as in the paper
scheduler = DDIMScheduler(
num_train_timesteps=1000,
beta_start=0.00085,
beta_end=0.012,
beta_schedule="scaled_linear",
set_alpha_to_one=False,
clip_sample=False,
)
text_encoder = CLIPTextModel.from_pretrained(
pretrained_model_name_or_path="openai/clip-vit-large-patch14",
torch_dtype=torch_dtype,
)
# initialize pipeline
pipeline = DiffusionPipeline.from_pretrained(
pretrained_model_name_or_path="CompVis/stable-diffusion-v1-4",
custom_pipeline="edict_pipeline",
revision="fp16",
scheduler=scheduler,
text_encoder=text_encoder,
leapfrog_steps=True,
torch_dtype=torch_dtype,
).to(device)
# download image
image_url = "https://huggingface.co/datasets/Joqsan/images/resolve/main/imagenet_dog_1.jpeg"
response = requests.get(image_url)
image = PIL.Image.open(BytesIO(response.content))
# preprocess it
cropped_image = center_crop_and_resize(image)
# define the prompts
base_prompt = "A dog"
target_prompt = "A golden retriever"
# run the pipeline
result_image = pipeline(
base_prompt=base_prompt,
target_prompt=target_prompt,
image=cropped_image,
)
display(result_image)
```
Init Image
Output Image
### Stable Diffusion RePaint
This pipeline uses the [RePaint](https://arxiv.org/abs/2201.09865) logic on the latent space of stable diffusion. It can
be used similarly to other image inpainting pipelines but does not rely on a specific inpainting model. This means you can use
models that are not specifically created for inpainting.
Make sure to use the ```RePaintScheduler``` as shown in the example below.
Disclaimer: The mask gets transferred into latent space, this may lead to unexpected changes on the edge of the masked part.
The inference time is a lot slower.
```py
import PIL
import requests
import torch
from io import BytesIO
from diffusers import StableDiffusionPipeline, RePaintScheduler
def download_image(url):
response = requests.get(url)
return PIL.Image.open(BytesIO(response.content)).convert("RGB")
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
init_image = download_image(img_url).resize((512, 512))
mask_image = download_image(mask_url).resize((512, 512))
mask_image = PIL.ImageOps.invert(mask_image)
pipe = StableDiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16, custom_pipeline="stable_diffusion_repaint",
)
pipe.scheduler = RePaintScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to("cuda")
prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
```
### TensorRT Image2Image Stable Diffusion Pipeline
The TensorRT Pipeline can be used to accelerate the Image2Image Stable Diffusion Inference run.
NOTE: The ONNX conversions and TensorRT engine build may take up to 30 minutes.
```python
import requests
from io import BytesIO
from PIL import Image
import torch
from diffusers import DDIMScheduler
from diffusers.pipelines.stable_diffusion import StableDiffusionImg2ImgPipeline
# Use the DDIMScheduler scheduler here instead
scheduler = DDIMScheduler.from_pretrained("stabilityai/stable-diffusion-2-1",
subfolder="scheduler")
pipe = StableDiffusionImg2ImgPipeline.from_pretrained("stabilityai/stable-diffusion-2-1",
custom_pipeline="stable_diffusion_tensorrt_img2img",
revision='fp16',
torch_dtype=torch.float16,
scheduler=scheduler,)
# re-use cached folder to save ONNX models and TensorRT Engines
pipe.set_cached_folder("stabilityai/stable-diffusion-2-1", revision='fp16',)
pipe = pipe.to("cuda")
url = "https://pajoca.com/wp-content/uploads/2022/09/tekito-yamakawa-1.png"
response = requests.get(url)
input_image = Image.open(BytesIO(response.content)).convert("RGB")
prompt = "photorealistic new zealand hills"
image = pipe(prompt, image=input_image, strength=0.75,).images[0]
image.save('tensorrt_img2img_new_zealand_hills.png')
```
### Stable Diffusion Reference
This pipeline uses the Reference Control. Refer to the [sd-webui-controlnet discussion: Reference-only Control](https://github.com/Mikubill/sd-webui-controlnet/discussions/1236)[sd-webui-controlnet discussion: Reference-adain Control](https://github.com/Mikubill/sd-webui-controlnet/discussions/1280).
Based on [this issue](https://github.com/huggingface/diffusers/issues/3566),
- `EulerAncestralDiscreteScheduler` got poor results.
```py
import torch
from diffusers import UniPCMultistepScheduler
from diffusers.utils import load_image
input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
pipe = StableDiffusionReferencePipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
safety_checker=None,
torch_dtype=torch.float16
).to('cuda:0')
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
result_img = pipe(ref_image=input_image,
prompt="1girl",
num_inference_steps=20,
reference_attn=True,
reference_adain=True).images[0]
```
Reference Image
Output Image of `reference_attn=True` and `reference_adain=False`
Output Image of `reference_attn=False` and `reference_adain=True`
Output Image of `reference_attn=True` and `reference_adain=True`
### Stable Diffusion ControlNet Reference
This pipeline uses the Reference Control with ControlNet. Refer to the [sd-webui-controlnet discussion: Reference-only Control](https://github.com/Mikubill/sd-webui-controlnet/discussions/1236)[sd-webui-controlnet discussion: Reference-adain Control](https://github.com/Mikubill/sd-webui-controlnet/discussions/1280).
Based on [this issue](https://github.com/huggingface/diffusers/issues/3566),
- `EulerAncestralDiscreteScheduler` got poor results.
- `guess_mode=True` works well for ControlNet v1.1
```py
import cv2
import torch
import numpy as np
from PIL import Image
from diffusers import UniPCMultistepScheduler
from diffusers.utils import load_image
input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
# get canny image
image = cv2.Canny(np.array(input_image), 100, 200)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
canny_image = Image.fromarray(image)
controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
pipe = StableDiffusionControlNetReferencePipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
controlnet=controlnet,
safety_checker=None,
torch_dtype=torch.float16
).to('cuda:0')
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
result_img = pipe(ref_image=input_image,
prompt="1girl",
image=canny_image,
num_inference_steps=20,
reference_attn=True,
reference_adain=True).images[0]
```
Reference Image
Output Image
### Stable Diffusion on IPEX
This diffusion pipeline aims to accelarate the inference of Stable-Diffusion on Intel Xeon CPUs with BF16/FP32 precision using [IPEX](https://github.com/intel/intel-extension-for-pytorch).
To use this pipeline, you need to:
1. Install [IPEX](https://github.com/intel/intel-extension-for-pytorch)
**Note:** For each PyTorch release, there is a corresponding release of the IPEX. Here is the mapping relationship. It is recommended to install Pytorch/IPEX2.0 to get the best performance.
|PyTorch Version|IPEX Version|
|--|--|
|[v2.0.\*](https://github.com/pytorch/pytorch/tree/v2.0.1 "v2.0.1")|[v2.0.\*](https://github.com/intel/intel-extension-for-pytorch/tree/v2.0.100+cpu)|
|[v1.13.\*](https://github.com/pytorch/pytorch/tree/v1.13.0 "v1.13.0")|[v1.13.\*](https://github.com/intel/intel-extension-for-pytorch/tree/v1.13.100+cpu)|
You can simply use pip to install IPEX with the latest version.
```python
python -m pip install intel_extension_for_pytorch
```
**Note:** To install a specific version, run with the following command:
```
python -m pip install intel_extension_for_pytorch==<version_name> -f https://developer.intel.com/ipex-whl-stable-cpu
```
2. After pipeline initialization, `prepare_for_ipex()` should be called to enable IPEX accelaration. Supported inference datatypes are Float32 and BFloat16.
**Note:** The setting of generated image height/width for `prepare_for_ipex()` should be same as the setting of pipeline inference.
```python
pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", custom_pipeline="stable_diffusion_ipex")
# For Float32
pipe.prepare_for_ipex(prompt, dtype=torch.float32, height=512, width=512) #value of image height/width should be consistent with the pipeline inference
# For BFloat16
pipe.prepare_for_ipex(prompt, dtype=torch.bfloat16, height=512, width=512) #value of image height/width should be consistent with the pipeline inference
```
Then you can use the ipex pipeline in a similar way to the default stable diffusion pipeline.
```python
# For Float32
image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] #value of image height/width should be consistent with 'prepare_for_ipex()'
# For BFloat16
with torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
image = pipe(prompt, num_inference_steps=20, height=512, width=512).images[0] #value of image height/width should be consistent with 'prepare_for_ipex()'
```
The following code compares the performance of the original stable diffusion pipeline with the ipex-optimized pipeline.
```python
import torch
import intel_extension_for_pytorch as ipex
from diffusers import StableDiffusionPipeline
import time
prompt = "sailing ship in storm by Rembrandt"
model_id = "runwayml/stable-diffusion-v1-5"
# Helper function for time evaluation
def elapsed_time(pipeline, nb_pass=3, num_inference_steps=20):
# warmup
for _ in range(2):
images = pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512).images
#time evaluation
start = time.time()
for _ in range(nb_pass):
pipeline(prompt, num_inference_steps=num_inference_steps, height=512, width=512)
end = time.time()
return (end - start) / nb_pass
############## bf16 inference performance ###############
# 1. IPEX Pipeline initialization
pipe = DiffusionPipeline.from_pretrained(model_id, custom_pipeline="stable_diffusion_ipex")
pipe.prepare_for_ipex(prompt, dtype=torch.bfloat16, height=512, width=512)
# 2. Original Pipeline initialization
pipe2 = StableDiffusionPipeline.from_pretrained(model_id)
# 3. Compare performance between Original Pipeline and IPEX Pipeline
with torch.cpu.amp.autocast(enabled=True, dtype=torch.bfloat16):
latency = elapsed_time(pipe)
print("Latency of StableDiffusionIPEXPipeline--bf16", latency)
latency = elapsed_time(pipe2)
print("Latency of StableDiffusionPipeline--bf16",latency)
############## fp32 inference performance ###############
# 1. IPEX Pipeline initialization
pipe3 = DiffusionPipeline.from_pretrained(model_id, custom_pipeline="stable_diffusion_ipex")
pipe3.prepare_for_ipex(prompt, dtype=torch.float32, height=512, width=512)
# 2. Original Pipeline initialization
pipe4 = StableDiffusionPipeline.from_pretrained(model_id)
# 3. Compare performance between Original Pipeline and IPEX Pipeline
latency = elapsed_time(pipe3)
print("Latency of StableDiffusionIPEXPipeline--fp32", latency)
latency = elapsed_time(pipe4)
print("Latency of StableDiffusionPipeline--fp32",latency)
```
### CLIP Guided Images Mixing With Stable Diffusion
CLIP guided stable diffusion images mixing pipeline allows to combine two images using standard diffusion models.
This approach is using (optional) CoCa model to avoid writing image description.
[More code examples](https://github.com/TheDenk/images_mixing)
### Stable Diffusion XL Long Weighted Prompt Pipeline
This SDXL pipeline support unlimited length prompt and negative prompt, compatible with A1111 prompt weighted style.
You can provide both `prompt` and `prompt_2`. If only one prompt is provided, `prompt_2` will be a copy of the provided `prompt`. Here is a sample code to use this pipeline.
```python
from diffusers import DiffusionPipeline
from diffusers.utils import load_image
import torch
pipe = DiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0"
, torch_dtype = torch.float16
, use_safetensors = True
, variant = "fp16"
, custom_pipeline = "lpw_stable_diffusion_xl",
)
prompt = "photo of a cute (white) cat running on the grass" * 20
prompt2 = "chasing (birds:1.5)" * 20
prompt = f"{prompt},{prompt2}"
neg_prompt = "blur, low quality, carton, animate"
pipe.to("cuda")
# text2img
t2i_images = pipe(
prompt=prompt,
negative_prompt=neg_prompt,
).images # alternatively, you can call the .text2img() function
# img2img
input_image = load_image("/path/to/local/image.png") # or URL to your input image
i2i_images = pipe.img2img(
prompt=prompt,
negative_prompt=neg_prompt,
image=input_image,
strength=0.8, # higher strength will result in more variation compared to original image
).images
# inpaint
input_mask = load_image("/path/to/local/mask.png") # or URL to your input inpainting mask
inpaint_images = pipe.inpaint(
prompt="photo of a cute (black) cat running on the grass" * 20,
negative_prompt=neg_prompt,
image=input_image,
mask=input_mask,
strength=0.6, # higher strength will result in more variation compared to original image
).images
pipe.to("cpu")
torch.cuda.empty_cache()
from IPython.display import display # assuming you are using this code in a notebook
display(t2i_images[0])
display(i2i_images[0])
display(inpaint_images[0])
```
In the above code, the `prompt2` is appended to the `prompt`, which is more than 77 tokens. "birds" are showing up in the result.
For more results, checkout [PR #6114](https://github.com/huggingface/diffusers/pull/6114).
## Example Images Mixing (with CoCa)
```python
import requests
from io import BytesIO
import PIL
import torch
import open_clip
from open_clip import SimpleTokenizer
from diffusers import DiffusionPipeline
from transformers import CLIPFeatureExtractor, CLIPModel
def download_image(url):
response = requests.get(url)
return PIL.Image.open(BytesIO(response.content)).convert("RGB")
# Loading additional models
feature_extractor = CLIPFeatureExtractor.from_pretrained(
"laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
)
clip_model = CLIPModel.from_pretrained(
"laion/CLIP-ViT-B-32-laion2B-s34B-b79K", torch_dtype=torch.float16
)
coca_model = open_clip.create_model('coca_ViT-L-14', pretrained='laion2B-s13B-b90k').to('cuda')
coca_model.dtype = torch.float16
coca_transform = open_clip.image_transform(
coca_model.visual.image_size,
is_train = False,
mean = getattr(coca_model.visual, 'image_mean', None),
std = getattr(coca_model.visual, 'image_std', None),
)
coca_tokenizer = SimpleTokenizer()
# Pipeline creating
mixing_pipeline = DiffusionPipeline.from_pretrained(
"CompVis/stable-diffusion-v1-4",
custom_pipeline="clip_guided_images_mixing_stable_diffusion",
clip_model=clip_model,
feature_extractor=feature_extractor,
coca_model=coca_model,
coca_tokenizer=coca_tokenizer,
coca_transform=coca_transform,
torch_dtype=torch.float16,
)
mixing_pipeline.enable_attention_slicing()
mixing_pipeline = mixing_pipeline.to("cuda")
# Pipeline running
generator = torch.Generator(device="cuda").manual_seed(17)
def download_image(url):
response = requests.get(url)
return PIL.Image.open(BytesIO(response.content)).convert("RGB")
content_image = download_image("https://huggingface.co/datasets/TheDenk/images_mixing/resolve/main/boromir.jpg")
style_image = download_image("https://huggingface.co/datasets/TheDenk/images_mixing/resolve/main/gigachad.jpg")
pipe_images = mixing_pipeline(
num_inference_steps=50,
content_image=content_image,
style_image=style_image,
noise_strength=0.65,
slerp_latent_style_strength=0.9,
slerp_prompt_style_strength=0.1,
slerp_clip_image_style_strength=0.1,
guidance_scale=9.0,
batch_size=1,
clip_guidance_scale=100,
generator=generator,
).images
```
### Stable Diffusion Mixture Tiling
This pipeline uses the Mixture. Refer to the [Mixture](https://arxiv.org/abs/2302.02412) paper for more details.
```python
from diffusers import LMSDiscreteScheduler, DiffusionPipeline
# Creater scheduler and model (similar to StableDiffusionPipeline)
scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
pipeline = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", scheduler=scheduler, custom_pipeline="mixture_tiling")
pipeline.to("cuda")
# Mixture of Diffusers generation
image = pipeline(
prompt=[[
"A charming house in the countryside, by jakub rozalski, sunset lighting, elegant, highly detailed, smooth, sharp focus, artstation, stunning masterpiece",
"A dirt road in the countryside crossing pastures, by jakub rozalski, sunset lighting, elegant, highly detailed, smooth, sharp focus, artstation, stunning masterpiece",
"An old and rusty giant robot lying on a dirt road, by jakub rozalski, dark sunset lighting, elegant, highly detailed, smooth, sharp focus, artstation, stunning masterpiece"
]],
tile_height=640,
tile_width=640,
tile_row_overlap=0,
tile_col_overlap=256,
guidance_scale=8,
seed=7178915308,
num_inference_steps=50,
)["images"][0]
```
### TensorRT Inpainting Stable Diffusion Pipeline
The TensorRT Pipeline can be used to accelerate the Inpainting Stable Diffusion Inference run.
NOTE: The ONNX conversions and TensorRT engine build may take up to 30 minutes.
```python
import requests
from io import BytesIO
from PIL import Image
import torch
from diffusers import PNDMScheduler
from diffusers.pipelines.stable_diffusion import StableDiffusionInpaintPipeline
# Use the PNDMScheduler scheduler here instead
scheduler = PNDMScheduler.from_pretrained("stabilityai/stable-diffusion-2-inpainting", subfolder="scheduler")
pipe = StableDiffusionInpaintPipeline.from_pretrained("stabilityai/stable-diffusion-2-inpainting",
custom_pipeline="stable_diffusion_tensorrt_inpaint",
revision='fp16',
torch_dtype=torch.float16,
scheduler=scheduler,
)
# re-use cached folder to save ONNX models and TensorRT Engines
pipe.set_cached_folder("stabilityai/stable-diffusion-2-inpainting", revision='fp16',)
pipe = pipe.to("cuda")
url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
response = requests.get(url)
input_image = Image.open(BytesIO(response.content)).convert("RGB")
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
response = requests.get(mask_url)
mask_image = Image.open(BytesIO(response.content)).convert("RGB")
prompt = "a mecha robot sitting on a bench"
image = pipe(prompt, image=input_image, mask_image=mask_image, strength=0.75,).images[0]
image.save('tensorrt_inpaint_mecha_robot.png')
```
### Stable Diffusion Mixture Canvas
This pipeline uses the Mixture. Refer to the [Mixture](https://arxiv.org/abs/2302.02412) paper for more details.
```python
from PIL import Image
from diffusers import LMSDiscreteScheduler, DiffusionPipeline
from diffusers.pipelines.pipeline_utils import Image2ImageRegion, Text2ImageRegion, preprocess_image
# Load and preprocess guide image
iic_image = preprocess_image(Image.open("input_image.png").convert("RGB"))
# Creater scheduler and model (similar to StableDiffusionPipeline)
scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
pipeline = DiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", scheduler=scheduler).to("cuda:0", custom_pipeline="mixture_canvas")
pipeline.to("cuda")
# Mixture of Diffusers generation
output = pipeline(
canvas_height=800,
canvas_width=352,
regions=[
Text2ImageRegion(0, 800, 0, 352, guidance_scale=8,
prompt=f"best quality, masterpiece, WLOP, sakimichan, art contest winner on pixiv, 8K, intricate details, wet effects, rain drops, ethereal, mysterious, futuristic, UHD, HDR, cinematic lighting, in a beautiful forest, rainy day, award winning, trending on artstation, beautiful confident cheerful young woman, wearing a futuristic sleeveless dress, ultra beautiful detailed eyes, hyper-detailed face, complex, perfect, model, textured, chiaroscuro, professional make-up, realistic, figure in frame, "),
Image2ImageRegion(352-800, 352, 0, 352, reference_image=iic_image, strength=1.0),
],
num_inference_steps=100,
seed=5525475061,
)["images"][0]
```
### IADB pipeline
This pipeline is the implementation of the [α-(de)Blending: a Minimalist Deterministic Diffusion Model](https://arxiv.org/abs/2305.03486) paper.
It is a simple and minimalist diffusion model.
The following code shows how to use the IADB pipeline to generate images using a pretrained celebahq-256 model.
```python
pipeline_iadb = DiffusionPipeline.from_pretrained("thomasc4/iadb-celebahq-256", custom_pipeline='iadb')
pipeline_iadb = pipeline_iadb.to('cuda')
output = pipeline_iadb(batch_size=4,num_inference_steps=128)
for i in range(len(output[0])):
plt.imshow(output[0][i])
plt.show()
```
Sampling with the IADB formulation is easy, and can be done in a few lines (the pipeline already implements it):
```python
def sample_iadb(model, x0, nb_step):
x_alpha = x0
for t in range(nb_step):
alpha = (t/nb_step)
alpha_next =((t+1)/nb_step)
d = model(x_alpha, torch.tensor(alpha, device=x_alpha.device))['sample']
x_alpha = x_alpha + (alpha_next-alpha)*d
return x_alpha
```
The training loop is also straightforward:
```python
# Training loop
while True:
x0 = sample_noise()
x1 = sample_dataset()
alpha = torch.rand(batch_size)
# Blend
x_alpha = (1-alpha) * x0 + alpha * x1
# Loss
loss = torch.sum((D(x_alpha, alpha)- (x1-x0))**2)
optimizer.zero_grad()
loss.backward()
optimizer.step()
```
### Zero1to3 pipeline
This pipeline is the implementation of the [Zero-1-to-3: Zero-shot One Image to 3D Object](https://arxiv.org/abs/2303.11328) paper.
The original pytorch-lightning [repo](https://github.com/cvlab-columbia/zero123) and a diffusers [repo](https://github.com/kxhit/zero123-hf).
The following code shows how to use the Zero1to3 pipeline to generate novel view synthesis images using a pretrained stable diffusion model.
```python
import os
import torch
from pipeline_zero1to3 import Zero1to3StableDiffusionPipeline
from diffusers.utils import load_image
model_id = "kxic/zero123-165000" # zero123-105000, zero123-165000, zero123-xl
pipe = Zero1to3StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe.enable_xformers_memory_efficient_attention()
pipe.enable_vae_tiling()
pipe.enable_attention_slicing()
pipe = pipe.to("cuda")
num_images_per_prompt = 4
# test inference pipeline
# x y z, Polar angle (vertical rotation in degrees) Azimuth angle (horizontal rotation in degrees) Zoom (relative distance from center)
query_pose1 = [-75.0, 100.0, 0.0]
query_pose2 = [-20.0, 125.0, 0.0]
query_pose3 = [-55.0, 90.0, 0.0]
# load image
# H, W = (256, 256) # H, W = (512, 512) # zero123 training is 256,256
# for batch input
input_image1 = load_image("./demo/4_blackarm.png") #load_image("https://cvlab-zero123-live.hf.space/file=/home/user/app/configs/4_blackarm.png")
input_image2 = load_image("./demo/8_motor.png") #load_image("https://cvlab-zero123-live.hf.space/file=/home/user/app/configs/8_motor.png")
input_image3 = load_image("./demo/7_london.png") #load_image("https://cvlab-zero123-live.hf.space/file=/home/user/app/configs/7_london.png")
input_images = [input_image1, input_image2, input_image3]
query_poses = [query_pose1, query_pose2, query_pose3]
# # for single input
# H, W = (256, 256)
# input_images = [input_image2.resize((H, W), PIL.Image.NEAREST)]
# query_poses = [query_pose2]
# better do preprocessing
from gradio_new import preprocess_image, create_carvekit_interface
import numpy as np
import PIL.Image as Image
pre_images = []
models = dict()
print('Instantiating Carvekit HiInterface...')
models['carvekit'] = create_carvekit_interface()
if not isinstance(input_images, list):
input_images = [input_images]
for raw_im in input_images:
input_im = preprocess_image(models, raw_im, True)
H, W = input_im.shape[:2]
pre_images.append(Image.fromarray((input_im * 255.0).astype(np.uint8)))
input_images = pre_images
# infer pipeline, in original zero123 num_inference_steps=76
images = pipe(input_imgs=input_images, prompt_imgs=input_images, poses=query_poses, height=H, width=W,
guidance_scale=3.0, num_images_per_prompt=num_images_per_prompt, num_inference_steps=50).images
# save imgs
log_dir = "logs"
os.makedirs(log_dir, exist_ok=True)
bs = len(input_images)
i = 0
for obj in range(bs):
for idx in range(num_images_per_prompt):
images[i].save(os.path.join(log_dir,f"obj{obj}_{idx}.jpg"))
i += 1
```
### Stable Diffusion XL Reference
This pipeline uses the Reference . Refer to the [stable_diffusion_reference](https://github.com/huggingface/diffusers/blob/main/examples/community/README.md#stable-diffusion-reference).
```py
import torch
from PIL import Image
from diffusers.utils import load_image
from diffusers import DiffusionPipeline
from diffusers.schedulers import UniPCMultistepScheduler
input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")
# pipe = DiffusionPipeline.from_pretrained(
# "stabilityai/stable-diffusion-xl-base-1.0",
# custom_pipeline="stable_diffusion_xl_reference",
# torch_dtype=torch.float16,
# use_safetensors=True,
# variant="fp16").to('cuda:0')
pipe = StableDiffusionXLReferencePipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16,
use_safetensors=True,
variant="fp16").to('cuda:0')
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
result_img = pipe(ref_image=input_image,
prompt="1girl",
num_inference_steps=20,
reference_attn=True,
reference_adain=True).images[0]
```
Reference Image
Output Image
`prompt: 1 girl`
`reference_attn=True, reference_adain=True, num_inference_steps=20`
Reference Image
Output Image
`prompt: A dog`
`reference_attn=True, reference_adain=False, num_inference_steps=20`
Reference Image
Output Image
`prompt: An astronaut riding a lion`
`reference_attn=True, reference_adain=True, num_inference_steps=20`
### Stable diffusion fabric pipeline
FABRIC approach applicable to a wide range of popular diffusion models, which exploits
the self-attention layer present in the most widely used architectures to condition
the diffusion process on a set of feedback images.
```python
import requests
import torch
from PIL import Image
from io import BytesIO
from diffusers import DiffusionPipeline
# load the pipeline
# make sure you're logged in with `huggingface-cli login`
model_id_or_path = "runwayml/stable-diffusion-v1-5"
#can also be used with dreamlike-art/dreamlike-photoreal-2.0
pipe = DiffusionPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16, custom_pipeline="pipeline_fabric").to("cuda")
# let's specify a prompt
prompt = "An astronaut riding an elephant"
negative_prompt = "lowres, cropped"
# call the pipeline
image = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
num_inference_steps=20,
generator=torch.manual_seed(12)
).images[0]
image.save("horse_to_elephant.jpg")
# let's try another example with feedback
url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/A%20black%20colored%20car.png"
response = requests.get(url)
init_image = Image.open(BytesIO(response.content)).convert("RGB")
prompt = "photo, A blue colored car, fish eye"
liked = [init_image]
## same goes with disliked
# call the pipeline
torch.manual_seed(0)
image = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
liked = liked,
num_inference_steps=20,
).images[0]
image.save("black_to_blue.png")
```
*With enough feedbacks you can create very similar high quality images.*
The original codebase can be found at [sd-fabric/fabric](https://github.com/sd-fabric/fabric), and available checkpoints are [dreamlike-art/dreamlike-photoreal-2.0](https://huggingface.co/dreamlike-art/dreamlike-photoreal-2.0), [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5), and [stabilityai/stable-diffusion-2-1](https://huggingface.co/stabilityai/stable-diffusion-2-1) (may give unexpected results).
Let's have a look at the images (*512X512*)
| Without Feedback | With Feedback (1st image) |
|---------------------|---------------------|
|  |  |
### Masked Im2Im Stable Diffusion Pipeline
This pipeline reimplements sketch inpaint feature from A1111 for non-inpaint models. The following code reads two images, original and one with mask painted over it. It computes mask as a difference of two images and does the inpainting in the area defined by the mask.
```python
img = PIL.Image.open("./mech.png")
# read image with mask painted over
img_paint = PIL.Image.open("./mech_painted.png")
neq = numpy.any(numpy.array(img) != numpy.array(img_paint), axis=-1)
mask = neq / neq.max()
pipeline = MaskedStableDiffusionImg2ImgPipeline.from_pretrained("frankjoshua/icbinpICantBelieveIts_v8")
# works best with EulerAncestralDiscreteScheduler
pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(pipeline.scheduler.config)
generator = torch.Generator(device="cpu").manual_seed(4)
prompt = "a man wearing a mask"
result = pipeline(prompt=prompt, image=img_paint, mask=mask, strength=0.75,
generator=generator)
result.images[0].save("result.png")
```
original image mech.png
<img src=https://github.com/noskill/diffusers/assets/733626/10ad972d-d655-43cb-8de1-039e3d79e849 width="25%" >
image with mask mech_painted.png
<img src=https://github.com/noskill/diffusers/assets/733626/c334466a-67fe-4377-9ff7-f46021b9c224 width="25%" >
result:
<img src=https://github.com/noskill/diffusers/assets/733626/23a0a71d-51db-471e-926a-107ac62512a8 width="25%" >
### Prompt2Prompt Pipeline
Prompt2Prompt allows the following edits:
- ReplaceEdit (change words in prompt)
- ReplaceEdit with local blend (change words in prompt, keep image part unrelated to changes constant)
- RefineEdit (add words to prompt)
- RefineEdit with local blend (add words to prompt, keep image part unrelated to changes constant)
- ReweightEdit (modulate importance of words)
Here's a full example for `ReplaceEdit``:
```python
import torch
import numpy as np
import matplotlib.pyplot as plt
from diffusers.pipelines import Prompt2PromptPipeline
pipe = Prompt2PromptPipeline.from_pretrained("CompVis/stable-diffusion-v1-4").to("cuda")
prompts = ["A turtle playing with a ball",
"A monkey playing with a ball"]
cross_attention_kwargs = {
"edit_type": "replace",
"cross_replace_steps": 0.4,
"self_replace_steps": 0.4
}
outputs = pipe(prompt=prompts, height=512, width=512, num_inference_steps=50, cross_attention_kwargs=cross_attention_kwargs)
```
And abbreviated examples for the other edits:
`ReplaceEdit with local blend`
```python
prompts = ["A turtle playing with a ball",
"A monkey playing with a ball"]
cross_attention_kwargs = {
"edit_type": "replace",
"cross_replace_steps": 0.4,
"self_replace_steps": 0.4,
"local_blend_words": ["turtle", "monkey"]
}
```
`RefineEdit`
```python
prompts = ["A turtle",
"A turtle in a forest"]
cross_attention_kwargs = {
"edit_type": "refine",
"cross_replace_steps": 0.4,
"self_replace_steps": 0.4,
}
```
`RefineEdit with local blend`
```python
prompts = ["A turtle",
"A turtle in a forest"]
cross_attention_kwargs = {
"edit_type": "refine",
"cross_replace_steps": 0.4,
"self_replace_steps": 0.4,
"local_blend_words": ["in", "a" , "forest"]
}
```
`ReweightEdit`
```python
prompts = ["A smiling turtle"] * 2
edit_kcross_attention_kwargswargs = {
"edit_type": "reweight",
"cross_replace_steps": 0.4,
"self_replace_steps": 0.4,
"equalizer_words": ["smiling"],
"equalizer_strengths": [5]
}
```
Side note: See [this GitHub gist](https://gist.github.com/UmerHA/b65bb5fb9626c9c73f3ade2869e36164) if you want to visualize the attention maps.
### Latent Consistency Pipeline
Latent Consistency Models was proposed in [Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference](https://arxiv.org/abs/2310.04378) by *Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, Hang Zhao* from Tsinghua University.
The abstract of the paper reads as follows:
*Latent Diffusion models (LDMs) have achieved remarkable results in synthesizing high-resolution images. However, the iterative sampling process is computationally intensive and leads to slow generation. Inspired by Consistency Models (song et al.), we propose Latent Consistency Models (LCMs), enabling swift inference with minimal steps on any pre-trained LDMs, including Stable Diffusion (rombach et al). Viewing the guided reverse diffusion process as solving an augmented probability flow ODE (PF-ODE), LCMs are designed to directly predict the solution of such ODE in latent space, mitigating the need for numerous iterations and allowing rapid, high-fidelity sampling. Efficiently distilled from pre-trained classifier-free guided diffusion models, a high-quality 768 x 768 2~4-step LCM takes only 32 A100 GPU hours for training. Furthermore, we introduce Latent Consistency Fine-tuning (LCF), a novel method that is tailored for fine-tuning LCMs on customized image datasets. Evaluation on the LAION-5B-Aesthetics dataset demonstrates that LCMs achieve state-of-the-art text-to-image generation performance with few-step inference. Project Page: [this https URL](https://latent-consistency-models.github.io/)*
The model can be used with `diffusers` as follows:
- *1. Load the model from the community pipeline.*
```py
from diffusers import DiffusionPipeline
import torch
pipe = DiffusionPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7", custom_pipeline="latent_consistency_txt2img", custom_revision="main")
# To save GPU memory, torch.float16 can be used, but it may compromise image quality.
pipe.to(torch_device="cuda", torch_dtype=torch.float32)
```
- 2. Run inference with as little as 4 steps:
```py
prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
# Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
num_inference_steps = 4
images = pipe(prompt=prompt, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
```
For any questions or feedback, feel free to reach out to [Simian Luo](https://github.com/luosiallen).
You can also try this pipeline directly in the [🚀 official spaces](https://huggingface.co/spaces/SimianLuo/Latent_Consistency_Model).
### Latent Consistency Img2img Pipeline
This pipeline extends the Latent Consistency Pipeline to allow it to take an input image.
```py
from diffusers import DiffusionPipeline
import torch
pipe = DiffusionPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7", custom_pipeline="latent_consistency_img2img")
# To save GPU memory, torch.float16 can be used, but it may compromise image quality.
pipe.to(torch_device="cuda", torch_dtype=torch.float32)
```
- 2. Run inference with as little as 4 steps:
```py
prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
input_image=Image.open("myimg.png")
strength = 0.5 #strength =0 (no change) strength=1 (completely overwrite image)
# Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
num_inference_steps = 4
images = pipe(prompt=prompt, image=input_image, strength=strength, num_inference_steps=num_inference_steps, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil").images
```
### Latent Consistency Interpolation Pipeline
This pipeline extends the Latent Consistency Pipeline to allow for interpolation of the latent space between multiple prompts. It is similar to the [Stable Diffusion Interpolate](https://github.com/huggingface/diffusers/blob/main/examples/community/interpolate_stable_diffusion.py) and [unCLIP Interpolate](https://github.com/huggingface/diffusers/blob/main/examples/community/unclip_text_interpolation.py) community pipelines.
```py
import torch
import numpy as np
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7", custom_pipeline="latent_consistency_interpolate")
# To save GPU memory, torch.float16 can be used, but it may compromise image quality.
pipe.to(torch_device="cuda", torch_dtype=torch.float32)
prompts = [
"Self-portrait oil painting, a beautiful cyborg with golden hair, Margot Robbie, 8k",
"Self-portrait oil painting, an extremely strong man, body builder, Huge Jackman, 8k",
"An astronaut floating in space, renaissance art, realistic, high quality, 8k",
"Oil painting of a cat, cute, dream-like",
"Hugging face emoji, cute, realistic"
]
num_inference_steps = 4
num_interpolation_steps = 60
seed = 1337
torch.manual_seed(seed)
np.random.seed(seed)
images = pipe(
prompt=prompts,
height=512,
width=512,
num_inference_steps=num_inference_steps,
num_interpolation_steps=num_interpolation_steps,
guidance_scale=8.0,
embedding_interpolation_type="lerp",
latent_interpolation_type="slerp",
process_batch_size=4, # Make it higher or lower based on your GPU memory
generator=torch.Generator(seed),
)
assert len(images) == (len(prompts) - 1) * num_interpolation_steps
```
### StableDiffusionUpscaleLDM3D Pipeline
[LDM3D-VR](https://arxiv.org/pdf/2311.03226.pdf) is an extended version of LDM3D.
The abstract from the paper is:
*Latent diffusion models have proven to be state-of-the-art in the creation and manipulation of visual outputs. However, as far as we know, the generation of depth maps jointly with RGB is still limited. We introduce LDM3D-VR, a suite of diffusion models targeting virtual reality development that includes LDM3D-pano and LDM3D-SR. These models enable the generation of panoramic RGBD based on textual prompts and the upscaling of low-resolution inputs to high-resolution RGBD, respectively. Our models are fine-tuned from existing pretrained models on datasets containing panoramic/high-resolution RGB images, depth maps and captions. Both models are evaluated in comparison to existing related methods*
Two checkpoints are available for use:
- [ldm3d-pano](https://huggingface.co/Intel/ldm3d-pano). This checkpoint enables the generation of panoramic images and requires the StableDiffusionLDM3DPipeline pipeline to be used.
- [ldm3d-sr](https://huggingface.co/Intel/ldm3d-sr). This checkpoint enables the upscaling of RGB and depth images. Can be used in cascade after the original LDM3D pipeline using the StableDiffusionUpscaleLDM3DPipeline pipeline.
'''py
from PIL import Image
import os
import torch
from diffusers import StableDiffusionLDM3DPipeline, DiffusionPipeline
#Generate a rgb/depth output from LDM3D
pipe_ldm3d = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c")
pipe_ldm3d.to("cuda")
prompt =f"A picture of some lemons on a table"
output = pipe_ldm3d(prompt)
rgb_image, depth_image = output.rgb, output.depth
rgb_image[0].save(f"lemons_ldm3d_rgb.jpg")
depth_image[0].save(f"lemons_ldm3d_depth.png")
#Upscale the previous output to a resolution of (1024, 1024)
pipe_ldm3d_upscale = DiffusionPipeline.from_pretrained("Intel/ldm3d-sr", custom_pipeline="pipeline_stable_diffusion_upscale_ldm3d")
pipe_ldm3d_upscale.to("cuda")
low_res_img = Image.open(f"lemons_ldm3d_rgb.jpg").convert("RGB")
low_res_depth = Image.open(f"lemons_ldm3d_depth.png").convert("L")
outputs = pipe_ldm3d_upscale(prompt="high quality high resolution uhd 4k image", rgb=low_res_img, depth=low_res_depth, num_inference_steps=50, target_res=[1024, 1024])
upscaled_rgb, upscaled_depth =outputs.rgb[0], outputs.depth[0]
upscaled_rgb.save(f"upscaled_lemons_rgb.png")
upscaled_depth.save(f"upscaled_lemons_depth.png")
'''
### ControlNet + T2I Adapter Pipeline
This pipelines combines both ControlNet and T2IAdapter into a single pipeline, where the forward pass is executed once.
It receives `control_image` and `adapter_image`, as well as `controlnet_conditioning_scale` and `adapter_conditioning_scale`, for the ControlNet and Adapter modules, respectively. Whenever `adapter_conditioning_scale = 0` or `controlnet_conditioning_scale = 0`, it will act as a full ControlNet module or as a full T2IAdapter module, respectively.
```py
import cv2
import numpy as np
import torch
from controlnet_aux.midas import MidasDetector
from PIL import Image
from diffusers import AutoencoderKL, ControlNetModel, MultiAdapter, T2IAdapter
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from diffusers.utils import load_image
from examples.community.pipeline_stable_diffusion_xl_controlnet_adapter import (
StableDiffusionXLControlNetAdapterPipeline,
)
controlnet_depth = ControlNetModel.from_pretrained(
"diffusers/controlnet-depth-sdxl-1.0",
torch_dtype=torch.float16,
variant="fp16",
use_safetensors=True
)
adapter_depth = T2IAdapter.from_pretrained(
"TencentARC/t2i-adapter-depth-midas-sdxl-1.0", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
)
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16, use_safetensors=True)
pipe = StableDiffusionXLControlNetAdapterPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
controlnet=controlnet_depth,
adapter=adapter_depth,
vae=vae,
variant="fp16",
use_safetensors=True,
torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")
pipe.enable_xformers_memory_efficient_attention()
# pipe.enable_freeu(s1=0.6, s2=0.4, b1=1.1, b2=1.2)
midas_depth = MidasDetector.from_pretrained(
"valhalla/t2iadapter-aux-models", filename="dpt_large_384.pt", model_type="dpt_large"
).to("cuda")
prompt = "a tiger sitting on a park bench"
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
image = load_image(img_url).resize((1024, 1024))
depth_image = midas_depth(
image, detect_resolution=512, image_resolution=1024
)
strength = 0.5
images = pipe(
prompt,
control_image=depth_image,
adapter_image=depth_image,
num_inference_steps=30,
controlnet_conditioning_scale=strength,
adapter_conditioning_scale=strength,
).images
images[0].save("controlnet_and_adapter.png")
```
### ControlNet + T2I Adapter + Inpainting Pipeline
```py
import cv2
import numpy as np
import torch
from controlnet_aux.midas import MidasDetector
from PIL import Image
from diffusers import AutoencoderKL, ControlNetModel, MultiAdapter, T2IAdapter
from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
from diffusers.utils import load_image
from examples.community.pipeline_stable_diffusion_xl_controlnet_adapter_inpaint import (
StableDiffusionXLControlNetAdapterInpaintPipeline,
)
controlnet_depth = ControlNetModel.from_pretrained(
"diffusers/controlnet-depth-sdxl-1.0",
torch_dtype=torch.float16,
variant="fp16",
use_safetensors=True
)
adapter_depth = T2IAdapter.from_pretrained(
"TencentARC/t2i-adapter-depth-midas-sdxl-1.0", torch_dtype=torch.float16, variant="fp16", use_safetensors=True
)
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16, use_safetensors=True)
pipe = StableDiffusionXLControlNetAdapterInpaintPipeline.from_pretrained(
"diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
controlnet=controlnet_depth,
adapter=adapter_depth,
vae=vae,
variant="fp16",
use_safetensors=True,
torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")
pipe.enable_xformers_memory_efficient_attention()
# pipe.enable_freeu(s1=0.6, s2=0.4, b1=1.1, b2=1.2)
midas_depth = MidasDetector.from_pretrained(
"valhalla/t2iadapter-aux-models", filename="dpt_large_384.pt", model_type="dpt_large"
).to("cuda")
prompt = "a tiger sitting on a park bench"
img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
image = load_image(img_url).resize((1024, 1024))
mask_image = load_image(mask_url).resize((1024, 1024))
depth_image = midas_depth(
image, detect_resolution=512, image_resolution=1024
)
strength = 0.4
images = pipe(
prompt,
image=image,
mask_image=mask_image,
control_image=depth_image,
adapter_image=depth_image,
num_inference_steps=30,
controlnet_conditioning_scale=strength,
adapter_conditioning_scale=strength,
strength=0.7,
).images
images[0].save("controlnet_and_adapter_inpaint.png")
```
### Regional Prompting Pipeline
This pipeline is a port of the [Regional Prompter extension](https://github.com/hako-mikan/sd-webui-regional-prompter) for [Stable Diffusion web UI](https://github.com/AUTOMATIC1111/stable-diffusion-webui) to diffusers.
This code implements a pipeline for the Stable Diffusion model, enabling the division of the canvas into multiple regions, with different prompts applicable to each region. Users can specify regions in two ways: using `Cols` and `Rows` modes for grid-like divisions, or the `Prompt` mode for regions calculated based on prompts.
### Usage
### Sample Code
```
from from examples.community.regional_prompting_stable_diffusion import RegionalPromptingStableDiffusionPipeline
pipe = RegionalPromptingStableDiffusionPipeline.from_single_file(model_path, vae=vae)
rp_args = {
"mode":"rows",
"div": "1;1;1"
}
prompt ="""
green hair twintail BREAK
red blouse BREAK
blue skirt
"""
images = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
guidance_scale=7.5,
height = 768,
width = 512,
num_inference_steps =20,
num_images_per_prompt = 1,
rp_args = rp_args
).images
time = time.strftime(r"%Y%m%d%H%M%S")
i = 1
for image in images:
i += 1
fileName = f'img-{time}-{i+1}.png'
image.save(fileName)
```
### Cols, Rows mode
In the Cols, Rows mode, you can split the screen vertically and horizontally and assign prompts to each region. The split ratio can be specified by 'div', and you can set the division ratio like '3;3;2' or '0.1;0.5'. Furthermore, as will be described later, you can also subdivide the split Cols, Rows to specify more complex regions.
In this image, the image is divided into three parts, and a separate prompt is applied to each. The prompts are divided by 'BREAK', and each is applied to the respective region.
```
green hair twintail BREAK
red blouse BREAK
blue skirt
```
### 2-Dimentional division
The prompt consists of instructions separated by the term `BREAK` and is assigned to different regions of a two-dimensional space. The image is initially split in the main splitting direction, which in this case is rows, due to the presence of a single semicolon`;`, dividing the space into an upper and a lower section. Additional sub-splitting is then applied, indicated by commas. The upper row is split into ratios of `2:1:1`, while the lower row is split into a ratio of `4:6`. Rows themselves are split in a `1:2` ratio. According to the reference image, the blue sky is designated as the first region, green hair as the second, the bookshelf as the third, and so on, in a sequence based on their position from the top left. The terrarium is placed on the desk in the fourth region, and the orange dress and sofa are in the fifth region, conforming to their respective splits.
```
rp_args = {
"mode":"rows",
"div": "1,2,1,1;2,4,6"
}
prompt ="""
blue sky BREAK
green hair BREAK
book shelf BREAK
terrarium on desk BREAK
orange dress and sofa
"""
```
### Prompt Mode
There are limitations to methods of specifying regions in advance. This is because specifying regions can be a hindrance when designating complex shapes or dynamic compositions. In the region specified by the prompt, the regions is determined after the image generation has begun. This allows us to accommodate compositions and complex regions.
For further infomagen, see [here](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/main/prompt_en.md).
### syntax
```
baseprompt target1 target2 BREAK
effect1, target1 BREAK
effect2 ,target2
```
First, write the base prompt. In the base prompt, write the words (target1, target2) for which you want to create a mask. Next, separate them with BREAK. Next, write the prompt corresponding to target1. Then enter a comma and write target1. The order of the targets in the base prompt and the order of the BREAK-separated targets can be back to back.
```
target2 baseprompt target1 BREAK
effect1, target1 BREAK
effect2 ,target2
```
is also effective.
### Sample
In this example, masks are calculated for shirt, tie, skirt, and color prompts are specified only for those regions.
```
rp_args = {
"mode":"prompt-ex",
"save_mask":True,
"th": "0.4,0.6,0.6",
}
prompt ="""
a girl in street with shirt, tie, skirt BREAK
red, shirt BREAK
green, tie BREAK
blue , skirt
"""
```
### threshold
The threshold used to determine the mask created by the prompt. This can be set as many times as there are masks, as the range varies widely depending on the target prompt. If multiple regions are used, enter them separated by commas. For example, hair tends to be ambiguous and requires a small value, while face tends to be large and requires a small value. These should be ordered by BREAK.
```
a lady ,hair, face BREAK
red, hair BREAK
tanned ,face
```
`threshold : 0.4,0.6`
If only one input is given for multiple regions, they are all assumed to be the same value.
### Prompt and Prompt-EX
The difference is that in Prompt, duplicate regions are added, whereas in Prompt-EX, duplicate regions are overwritten sequentially. Since they are processed in order, setting a TARGET with a large regions first makes it easier for the effect of small regions to remain unmuffled.
### Accuracy
In the case of a 512 x 512 image, Attention mode reduces the size of the region to about 8 x 8 pixels deep in the U-Net, so that small regions get mixed up; Latent mode calculates 64*64, so that the region is exact.
```
girl hair twintail frills,ribbons, dress, face BREAK
girl, ,face
```
### Mask
When an image is generated, the generated mask is displayed. It is generated at the same size as the image, but is actually used at a much smaller size.
### Use common prompt
You can attach the prompt up to ADDCOMM to all prompts by separating it first with ADDCOMM. This is useful when you want to include elements common to all regions. For example, when generating pictures of three people with different appearances, it's necessary to include the instruction of 'three people' in all regions. It's also useful when inserting quality tags and other things."For example, if you write as follows:
```
best quality, 3persons in garden, ADDCOMM
a girl white dress BREAK
a boy blue shirt BREAK
an old man red suit
```
If common is enabled, this prompt is converted to the following:
```
best quality, 3persons in garden, a girl white dress BREAK
best quality, 3persons in garden, a boy blue shirt BREAK
best quality, 3persons in garden, an old man red suit
```
### Negative prompt
Negative prompts are equally effective across all regions, but it is possible to set region-specific prompts for negative prompts as well. The number of BREAKs must be the same as the number of prompts. If the number of prompts does not match, the negative prompts will be used without being divided into regions.
### Parameters
To activate Regional Prompter, it is necessary to enter settings in rp_args. The items that can be set are as follows. rp_args is a dictionary type.
### Input Parameters
Parameters are specified through the `rp_arg`(dictionary type).
```
rp_args = {
"mode":"rows",
"div": "1;1;1"
}
pipe(prompt =prompt, rp_args = rp_args)
```
### Required Parameters
- `mode`: Specifies the method for defining regions. Choose from `Cols`, `Rows`, `Prompt` or `Prompt-Ex`. This parameter is case-insensitive.
- `divide`: Used in `Cols` and `Rows` modes. Details on how to specify this are provided under the respective `Cols` and `Rows` sections.
- `th`: Used in `Prompt` mode. The method of specification is detailed under the `Prompt` section.
### Optional Parameters
- `save_mask`: In `Prompt` mode, choose whether to output the generated mask along with the image. The default is `False`.
The Pipeline supports `compel` syntax. Input prompts using the `compel` structure will be automatically applied and processed.
## Diffusion Posterior Sampling Pipeline
* Reference paper
```
@article{chung2022diffusion,
title={Diffusion posterior sampling for general noisy inverse problems},
author={Chung, Hyungjin and Kim, Jeongsol and Mccann, Michael T and Klasky, Marc L and Ye, Jong Chul},
journal={arXiv preprint arXiv:2209.14687},
year={2022}
}
```
* This pipeline allows zero-shot conditional sampling from the posterior distribution $p(x|y)$, given observation on $y$, unconditional generative model $p(x)$ and differentiable operator $y=f(x)$.
* For example, $f(.)$ can be downsample operator, then $y$ is a downsampled image, and the pipeline becomes a super-resolution pipeline.
* To use this pipeline, you need to know your operator $f(.)$ and corrupted image $y$, and pass them during the call. For example, as in the main function of dps_pipeline.py, you need to first define the Gaussian blurring operator $f(.)$. The operator should be a callable nn.Module, with all the parameter gradient disabled:
```python
import torch.nn.functional as F
import scipy
from torch import nn
# define the Gaussian blurring operator first
class GaussialBlurOperator(nn.Module):
def __init__(self, kernel_size, intensity):
super().__init__()
class Blurkernel(nn.Module):
def __init__(self, blur_type='gaussian', kernel_size=31, std=3.0):
super().__init__()
self.blur_type = blur_type
self.kernel_size = kernel_size
self.std = std
self.seq = nn.Sequential(
nn.ReflectionPad2d(self.kernel_size//2),
nn.Conv2d(3, 3, self.kernel_size, stride=1, padding=0, bias=False, groups=3)
)
self.weights_init()
def forward(self, x):
return self.seq(x)
def weights_init(self):
if self.blur_type == "gaussian":
n = np.zeros((self.kernel_size, self.kernel_size))
n[self.kernel_size // 2,self.kernel_size // 2] = 1
k = scipy.ndimage.gaussian_filter(n, sigma=self.std)
k = torch.from_numpy(k)
self.k = k
for name, f in self.named_parameters():
f.data.copy_(k)
elif self.blur_type == "motion":
k = Kernel(size=(self.kernel_size, self.kernel_size), intensity=self.std).kernelMatrix
k = torch.from_numpy(k)
self.k = k
for name, f in self.named_parameters():
f.data.copy_(k)
def update_weights(self, k):
if not torch.is_tensor(k):
k = torch.from_numpy(k)
for name, f in self.named_parameters():
f.data.copy_(k)
def get_kernel(self):
return self.k
self.kernel_size = kernel_size
self.conv = Blurkernel(blur_type='gaussian',
kernel_size=kernel_size,
std=intensity)
self.kernel = self.conv.get_kernel()
self.conv.update_weights(self.kernel.type(torch.float32))
for param in self.parameters():
param.requires_grad=False
def forward(self, data, **kwargs):
return self.conv(data)
def transpose(self, data, **kwargs):
return data
def get_kernel(self):
return self.kernel.view(1, 1, self.kernel_size, self.kernel_size)
```
* Next, you should obtain the corrupted image $y$ by the operator. In this example, we generate $y$ from the source image $x$. However in practice, having the operator $f(.)$ and corrupted image $y$ is enough:
```python
# set up source image
src = Image.open('sample.png')
# read image into [1,3,H,W]
src = torch.from_numpy(np.array(src, dtype=np.float32)).permute(2,0,1)[None]
# normalize image to [-1,1]
src = (src / 127.5) - 1.0
src = src.to("cuda")
# set up operator and measurement
operator = GaussialBlurOperator(kernel_size=61, intensity=3.0).to("cuda")
measurement = operator(src)
# save the source and corrupted images
save_image((src+1.0)/2.0, "dps_src.png")
save_image((measurement+1.0)/2.0, "dps_mea.png")
```
* We provide an example pair of saved source and corrupted images, using the Gaussian blur operator above
* Source image:
* 
* Gaussian blurred image:
* 
* You can download those image to run the example on your own.
* Next, we need to define a loss function used for diffusion posterior sample. For most of the cases, the RMSE is fine:
```python
def RMSELoss(yhat, y):
return torch.sqrt(torch.sum((yhat-y)**2))
```
* And next, as any other diffusion models, we need the score estimator and scheduler. As we are working with $256x256$ face images, we use ddmp-celebahq-256:
```python
# set up scheduler
scheduler = DDPMScheduler.from_pretrained("google/ddpm-celebahq-256")
scheduler.set_timesteps(1000)
# set up model
model = UNet2DModel.from_pretrained("google/ddpm-celebahq-256").to("cuda")
```
* And finally, run the pipeline:
```python
# finally, the pipeline
dpspipe = DPSPipeline(model, scheduler)
image = dpspipe(
measurement = measurement,
operator = operator,
loss_fn = RMSELoss,
zeta = 1.0,
).images[0]
image.save("dps_generated_image.png")
```
* The zeta is a hyperparameter that is in range of $[0,1]$. It need to be tuned for best effect. By setting zeta=1, you should be able to have the reconstructed result:
* Reconstructed image:
* 
* The reconstruction is perceptually similar to the source image, but different in details.
* In dps_pipeline.py, we also provide a super-resolution example, which should produce:
* Downsampled image:
* 
* Reconstructed image:
* 
### AnimateDiff ControlNet Pipeline
This pipeline combines AnimateDiff and ControlNet. Enjoy precise motion control for your videos! Refer to [this](https://github.com/huggingface/diffusers/issues/5866) issue for more details.
```py
import torch
from diffusers import AutoencoderKL, ControlNetModel, MotionAdapter
from diffusers.pipelines import DiffusionPipeline
from diffusers.schedulers import DPMSolverMultistepScheduler
from PIL import Image
motion_id = "guoyww/animatediff-motion-adapter-v1-5-2"
adapter = MotionAdapter.from_pretrained(motion_id)
controlnet = ControlNetModel.from_pretrained("lllyasviel/control_v11p_sd15_openpose", torch_dtype=torch.float16)
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", torch_dtype=torch.float16)
model_id = "SG161222/Realistic_Vision_V5.1_noVAE"
pipe = DiffusionPipeline.from_pretrained(
model_id,
motion_adapter=adapter,
controlnet=controlnet,
vae=vae,
custom_pipeline="pipeline_animatediff_controlnet",
).to(device="cuda", dtype=torch.float16)
pipe.scheduler = DPMSolverMultistepScheduler.from_pretrained(
model_id, subfolder="scheduler", clip_sample=False, timestep_spacing="linspace", steps_offset=1
)
pipe.enable_vae_slicing()
conditioning_frames = []
for i in range(1, 16 + 1):
conditioning_frames.append(Image.open(f"frame_{i}.png"))
prompt = "astronaut in space, dancing"
negative_prompt = "bad quality, worst quality, jpeg artifacts, ugly"
result = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
width=512,
height=768,
conditioning_frames=conditioning_frames,
num_inference_steps=12,
).frames[0]
from diffusers.utils import export_to_gif
export_to_gif(result.frames[0], "result.gif")
```
<table>
<tr><td colspan="2" align=center><b>Conditioning Frames</b></td></tr>
<tr align=center>
<td align=center><img src="https://user-images.githubusercontent.com/7365912/265043418-23291941-864d-495a-8ba8-d02e05756396.gif" alt="input-frames"></td>
</tr>
<tr><td colspan="2" align=center><b>AnimateDiff model: SG161222/Realistic_Vision_V5.1_noVAE</b></td></tr>
<tr>
<td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/baf301e2-d03c-4129-bd84-203a1de2b2be" alt="gif-1"></td>
<td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/9f923475-ecaf-452b-92c8-4e42171182d8" alt="gif-2"></td>
</tr>
<tr><td colspan="2" align=center><b>AnimateDiff model: CardosAnime</b></td></tr>
<tr>
<td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/b2c41028-38a0-45d6-86ed-fec7446b87f7" alt="gif-1"></td>
<td align=center><img src="https://github.com/huggingface/diffusers/assets/72266394/eb7d2952-72e4-44fa-b664-077c79b4fc70" alt="gif-2"></td>
</tr>
</table>
### DemoFusion
This pipeline is the official implementation of [DemoFusion: Democratising High-Resolution Image Generation With No $$$](https://arxiv.org/abs/2311.16973).
The original repo can be found at [repo](https://github.com/PRIS-CV/DemoFusion).
- `view_batch_size` (`int`, defaults to 16):
The batch size for multiple denoising paths. Typically, a larger batch size can result in higher efficiency but comes with increased GPU memory requirements.
- `stride` (`int`, defaults to 64):
The stride of moving local patches. A smaller stride is better for alleviating seam issues, but it also introduces additional computational overhead and inference time.
- `cosine_scale_1` (`float`, defaults to 3):
Control the strength of skip-residual. For specific impacts, please refer to Appendix C in the DemoFusion paper.
- `cosine_scale_2` (`float`, defaults to 1):
Control the strength of dilated sampling. For specific impacts, please refer to Appendix C in the DemoFusion paper.
- `cosine_scale_3` (`float`, defaults to 1):
Control the strength of the Gaussian filter. For specific impacts, please refer to Appendix C in the DemoFusion paper.
- `sigma` (`float`, defaults to 1):
The standard value of the Gaussian filter. Larger sigma promotes the global guidance of dilated sampling, but has the potential of over-smoothing.
- `multi_decoder` (`bool`, defaults to True):
Determine whether to use a tiled decoder. Generally, when the resolution exceeds 3072x3072, a tiled decoder becomes necessary.
- `show_image` (`bool`, defaults to False):
Determine whether to show intermediate results during generation.
```py
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
custom_pipeline="pipeline_demofusion_sdxl",
custom_revision="main",
torch_dtype=torch.float16,
)
pipe = pipe.to("cuda")
prompt = "Envision a portrait of an elderly woman, her face a canvas of time, framed by a headscarf with muted tones of rust and cream. Her eyes, blue like faded denim. Her attire, simple yet dignified."
negative_prompt = "blurry, ugly, duplicate, poorly drawn, deformed, mosaic"
images = pipe(
prompt,
negative_prompt=negative_prompt,
height=3072,
width=3072,
view_batch_size=16,
stride=64,
num_inference_steps=50,
guidance_scale=7.5,
cosine_scale_1=3,
cosine_scale_2=1,
cosine_scale_3=1,
sigma=0.8,
multi_decoder=True,
show_image=True
)
```
You can display and save the generated images as:
```py
def image_grid(imgs, save_path=None):
w = 0
for i, img in enumerate(imgs):
h_, w_ = imgs[i].size
w += w_
h = h_
grid = Image.new('RGB', size=(w, h))
grid_w, grid_h = grid.size
w = 0
for i, img in enumerate(imgs):
h_, w_ = imgs[i].size
grid.paste(img, box=(w, h - h_))
if save_path != None:
img.save(save_path + "/img_{}.jpg".format((i + 1) * 1024))
w += w_
return grid
image_grid(images, save_path="./outputs/")
```
### SDE Drag pipeline
This pipeline provides drag-and-drop image editing using stochastic differential equations. It enables image editing by inputting prompt, image, mask_image, source_points, and target_points.
See [paper](https://arxiv.org/abs/2311.01410), [paper page](https://ml-gsai.github.io/SDE-Drag-demo/), [original repo](https://github.com/ML-GSAI/SDE-Drag) for more infomation.
```py
import PIL
import torch
from diffusers import DDIMScheduler, DiffusionPipeline
# Load the pipeline
model_path = "runwayml/stable-diffusion-v1-5"
scheduler = DDIMScheduler.from_pretrained(model_path, subfolder="scheduler")
pipe = DiffusionPipeline.from_pretrained(model_path, scheduler=scheduler, custom_pipeline="sde_drag")
pipe.to('cuda')
# To save GPU memory, torch.float16 can be used, but it may compromise image quality.
# If not training LoRA, please avoid using torch.float16
# pipe.to(torch.float16)
# Provide prompt, image, mask image, and the starting and target points for drag editing.
prompt = "prompt of the image"
image = PIL.Image.open('/path/to/image')
mask_image = PIL.Image.open('/path/to/mask_image')
source_points = [[123, 456]]
target_points = [[234, 567]]
# train_lora is optional, and in most cases, using train_lora can better preserve consistency with the original image.
pipe.train_lora(prompt, image)
output = pipe(prompt, image, mask_image, source_points, target_points)
output_image = PIL.Image.fromarray(output)
output_image.save("./output.png")
```
| huggingface/diffusers/blob/main/examples/community/README.md |
ResNeXt
A **ResNeXt** repeats a [building block](https://paperswithcode.com/method/resnext-block) that aggregates a set of transformations with the same topology. Compared to a [ResNet](https://paperswithcode.com/method/resnet), it exposes a new dimension, *cardinality* (the size of the set of transformations) \\( C \\), as an essential factor in addition to the dimensions of depth and width.
## How do I use this model on an image?
To load a pretrained model:
```py
>>> import timm
>>> model = timm.create_model('resnext101_32x8d', pretrained=True)
>>> model.eval()
```
To load and preprocess the image:
```py
>>> import urllib
>>> from PIL import Image
>>> from timm.data import resolve_data_config
>>> from timm.data.transforms_factory import create_transform
>>> config = resolve_data_config({}, model=model)
>>> transform = create_transform(**config)
>>> url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
>>> urllib.request.urlretrieve(url, filename)
>>> img = Image.open(filename).convert('RGB')
>>> tensor = transform(img).unsqueeze(0) # transform and add batch dimension
```
To get the model predictions:
```py
>>> import torch
>>> with torch.no_grad():
... out = model(tensor)
>>> probabilities = torch.nn.functional.softmax(out[0], dim=0)
>>> print(probabilities.shape)
>>> # prints: torch.Size([1000])
```
To get the top-5 predictions class names:
```py
>>> # Get imagenet class mappings
>>> url, filename = ("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt", "imagenet_classes.txt")
>>> urllib.request.urlretrieve(url, filename)
>>> with open("imagenet_classes.txt", "r") as f:
... categories = [s.strip() for s in f.readlines()]
>>> # Print top categories per image
>>> top5_prob, top5_catid = torch.topk(probabilities, 5)
>>> for i in range(top5_prob.size(0)):
... print(categories[top5_catid[i]], top5_prob[i].item())
>>> # prints class names and probabilities like:
>>> # [('Samoyed', 0.6425196528434753), ('Pomeranian', 0.04062102362513542), ('keeshond', 0.03186424449086189), ('white wolf', 0.01739676296710968), ('Eskimo dog', 0.011717947199940681)]
```
Replace the model name with the variant you want to use, e.g. `resnext101_32x8d`. You can find the IDs in the model summaries at the top of this page.
To extract image features with this model, follow the [timm feature extraction examples](../feature_extraction), just change the name of the model you want to use.
## How do I finetune this model?
You can finetune any of the pre-trained models just by changing the classifier (the last layer).
```py
>>> model = timm.create_model('resnext101_32x8d', pretrained=True, num_classes=NUM_FINETUNE_CLASSES)
```
To finetune on your own dataset, you have to write a training loop or adapt [timm's training
script](https://github.com/rwightman/pytorch-image-models/blob/master/train.py) to use your dataset.
## How do I train this model?
You can follow the [timm recipe scripts](../scripts) for training a new model afresh.
## Citation
```BibTeX
@article{DBLP:journals/corr/XieGDTH16,
author = {Saining Xie and
Ross B. Girshick and
Piotr Doll{\'{a}}r and
Zhuowen Tu and
Kaiming He},
title = {Aggregated Residual Transformations for Deep Neural Networks},
journal = {CoRR},
volume = {abs/1611.05431},
year = {2016},
url = {http://arxiv.org/abs/1611.05431},
archivePrefix = {arXiv},
eprint = {1611.05431},
timestamp = {Mon, 13 Aug 2018 16:45:58 +0200},
biburl = {https://dblp.org/rec/journals/corr/XieGDTH16.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
```
<!--
Type: model-index
Collections:
- Name: ResNeXt
Paper:
Title: Aggregated Residual Transformations for Deep Neural Networks
URL: https://paperswithcode.com/paper/aggregated-residual-transformations-for-deep
Models:
- Name: resnext101_32x8d
In Collection: ResNeXt
Metadata:
FLOPs: 21180417024
Parameters: 88790000
File Size: 356082095
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: resnext101_32x8d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bilinear
Code: https://github.com/rwightman/pytorch-image-models/blob/b9843f954b0457af2db4f9dea41a8538f51f5d78/timm/models/resnet.py#L877
Weights: https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.3%
Top 5 Accuracy: 94.53%
- Name: resnext50_32x4d
In Collection: ResNeXt
Metadata:
FLOPs: 5472648192
Parameters: 25030000
File Size: 100435887
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: resnext50_32x4d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/b9843f954b0457af2db4f9dea41a8538f51f5d78/timm/models/resnet.py#L851
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnext50_32x4d_ra-d733960d.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.79%
Top 5 Accuracy: 94.61%
- Name: resnext50d_32x4d
In Collection: ResNeXt
Metadata:
FLOPs: 5781119488
Parameters: 25050000
File Size: 100515304
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: resnext50d_32x4d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/b9843f954b0457af2db4f9dea41a8538f51f5d78/timm/models/resnet.py#L869
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/resnext50d_32x4d-103e99f8.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.67%
Top 5 Accuracy: 94.87%
- Name: tv_resnext50_32x4d
In Collection: ResNeXt
Metadata:
FLOPs: 5472648192
Parameters: 25030000
File Size: 100441675
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Techniques:
- SGD with Momentum
- Weight Decay
Training Data:
- ImageNet
ID: tv_resnext50_32x4d
LR: 0.1
Epochs: 90
Crop Pct: '0.875'
LR Gamma: 0.1
Momentum: 0.9
Batch Size: 32
Image Size: '224'
LR Step Size: 30
Weight Decay: 0.0001
Interpolation: bilinear
Code: https://github.com/rwightman/pytorch-image-models/blob/9a25fdf3ad0414b4d66da443fe60ae0aa14edc84/timm/models/resnet.py#L842
Weights: https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 77.61%
Top 5 Accuracy: 93.68%
-->
| huggingface/pytorch-image-models/blob/main/hfdocs/source/models/resnext.mdx |
Gradio Demo: gpt2_xl_unified
```
!pip install -q gradio
```
```
import gradio as gr
component = gr.Textbox(lines=5, label="Text")
api = gr.load("huggingface/gpt2-xl")
demo = gr.Interface(
fn=lambda x: x[:-50] + api(x[-50:]),
inputs=component,
outputs=component,
title="gpt2-xl",
)
if __name__ == "__main__":
demo.launch()
```
| gradio-app/gradio/blob/main/demo/gpt2_xl_unified/run.ipynb |
!-- DISABLE-FRONTMATTER-SECTIONS -->
# End-of-chapter quiz[[end-of-chapter-quiz]]
<CourseFloatingBanner
chapter={5}
classNames="absolute z-10 right-0 top-0"
/>
This chapter covered a lot of ground! Don't worry if you didn't grasp all the details; the next chapters will help you understand how things work under the hood.
Before moving on, though, let's test what you learned in this chapter.
### 1. The `load_dataset()` function in 🤗 Datasets allows you to load a dataset from which of the following locations?
<Question
choices={[
{
text: "Locally, e.g. on your laptop",
explain: "Correct! You can pass the paths of local files to the <code>data_files</code> argument of <code>load_dataset()</code> to load local datasets.",
correct: true
},
{
text: "The Hugging Face Hub",
explain: "Correct! You can load datasets on the Hub by providing the dataset ID, e.g. <code>load_dataset('emotion')</code>.",
correct: true
},
{
text: "A remote server",
explain: "Correct! You can pass URLs to the <code>data_files</code> argument of <code>load_dataset()</code> to load remote files.",
correct: true
},
]}
/>
### 2. Suppose you load one of the GLUE tasks as follows:
```py
from datasets import load_dataset
dataset = load_dataset("glue", "mrpc", split="train")
```
Which of the following commands will produce a random sample of 50 elements from `dataset`?
<Question
choices={[
{
text: "<code>dataset.sample(50)</code>",
explain: "This is incorrect -- there is no <code>Dataset.sample()</code> method."
},
{
text: "<code>dataset.shuffle().select(range(50))</code>",
explain: "Correct! As you saw in this chapter, you first shuffle the dataset and then select the samples from it.",
correct: true
},
{
text: "<code>dataset.select(range(50)).shuffle()</code>",
explain: "This is incorrect -- although the code will run, it will only shuffle the first 50 elements in the dataset."
}
]}
/>
### 3. Suppose you have a dataset about household pets called `pets_dataset`, which has a `name` column that denotes the name of each pet. Which of the following approaches would allow you to filter the dataset for all pets whose names start with the letter "L"?
<Question
choices={[
{
text: "<code>pets_dataset.filter(lambda x : x['name'].startswith('L'))</code>",
explain: "Correct! Using a Python lambda function for these quick filters is a great idea. Can you think of another solution?",
correct: true
},
{
text: "<code>pets_dataset.filter(lambda x['name'].startswith('L'))</code>",
explain: "This is incorrect -- a lambda function takes the general form <code>lambda *arguments* : *expression*</code>, so you need to provide arguments in this case."
},
{
text: "Create a function like <code>def filter_names(x): return x['name'].startswith('L')</code> and run <code>pets_dataset.filter(filter_names)</code>.",
explain: "Correct! Just like with <code>Dataset.map()</code>, you can pass explicit functions to <code>Dataset.filter()</code>. This is useful when you have some complex logic that isn't suitable for a short lambda function. Which of the other solutions would work?",
correct: true
}
]}
/>
### 4. What is memory mapping?
<Question
choices={[
{
text: "A mapping between CPU and GPU RAM",
explain: "That's not it -- try again!",
},
{
text: "A mapping between RAM and filesystem storage",
explain: "Correct! 🤗 Datasets treats each dataset as a memory-mapped file. This allows the library to access and operate on elements of the dataset without needing to fully load it into memory.",
correct: true
},
{
text: "A mapping between two files in the 🤗 Datasets cache",
explain: "This is not correct - try again!"
}
]}
/>
### 5. Which of the following are the main benefits of memory mapping?
<Question
choices={[
{
text: "Accessing memory-mapped files is faster than reading from or writing to disk.",
explain: "Correct! This allows 🤗 Datasets to be blazing fast. That's not the only benefit, though.",
correct: true
},
{
text: "Applications can access segments of data in an extremely large file without having to read the whole file into RAM first.",
explain: "Correct! This allows 🤗 Datasets to load multi-gigabyte datasets on your laptop without blowing up your CPU. What other advantage does memory mapping offer?",
correct: true
},
{
text: "It consumes less energy, so your battery lasts longer.",
explain: "This is not correct -- try again!"
}
]}
/>
### 6. Why does the following code fail?
```py
from datasets import load_dataset
dataset = load_dataset("allocine", streaming=True, split="train")
dataset[0]
```
<Question
choices={[
{
text: "It tries to stream a dataset that's too large to fit in RAM.",
explain: "This is not correct -- streaming datasets are decompressed on the fly, and you can process terabyte-sized datasets with very little RAM!",
},
{
text: "It tries to access an <code>IterableDataset</code>.",
explain: "Correct! An <code>IterableDataset</code> is a generator, not a container, so you should access its elements using <code>next(iter(dataset))</code>.",
correct: true
},
{
text: "The <code>allocine</code> dataset doesn't have a <code>train</code> split.",
explain: "This is incorrect -- check out the [<code>allocine</code> dataset card](https://huggingface.co/datasets/allocine) on the Hub to see which splits it contains."
}
]}
/>
### 7. Which of the following are the main benefits of creating a dataset card?
<Question
choices={[
{
text: "It provides information about the intended use and supported tasks of the dataset so others in the community can make an informed decision about using it.",
explain: "Correct! Undocumented datasets may be used to train models that may not reflect the intentions of the dataset creators, or may produce models whose legal status is murky if they're trained on data that violates privacy or licensing restrictions. This isn't the only benefit, though!",
correct : true
},
{
text: "It helps draw attention to the biases that are present in a corpus.",
explain: "Correct! Almost all datasets have some form of bias, which can produce negative consequences downstream. Being aware of them helps model builders understand how to address the inherent biases. What else do dataset cards help with?",
correct : true
},
{
text: "It improves the chances that others in the community will use my dataset.",
explain: "Correct! A well-written dataset card will tend to lead to higher usage of your precious dataset. What other benefits does it offer?",
correct: true
},
]}
/>
### 8. What is semantic search?
<Question
choices={[
{
text: "A way to search for exact matches between the words in a query and the documents in a corpus",
explain: "This is incorrect -- this type of search is called *lexical search*, and it's what you typically see with traditional search engines."
},
{
text: "A way to search for matching documents by understanding the contextual meaning of a query",
explain: "Correct! Semantic search uses embedding vectors to represent queries and documents, and uses a similarity metric to measure the amount of overlap between them. How else might you describe it?",
correct: true
},
{
text: "A way to improve search accuracy",
explain: "Correct! Semantic search engines can capture the intent of a query much better than keyword matching and typically retrieve documents with higher precision. But this isn't the only right answer - what else does semantic search provide?",
correct: true
}
]}
/>
### 9. For asymmetric semantic search, you usually have:
<Question
choices={[
{
text: "A short query and a longer paragraph that answers the query",
explain: "Correct!",
correct : true
},
{
text: "Queries and paragraphs that are of about the same length",
explain: "This is actually an example of symmetric semantic search -- try again!"
},
{
text: "A long query and a shorter paragraph that answers the query",
explain: "This is incorrect -- try again!"
}
]}
/>
### 10. Can I use 🤗 Datasets to load data for use in other domains, like speech processing?
<Question
choices={[
{
text: "No",
explain: "This is incorrect -- 🤗 Datasets currently supports tabular data, audio, and computer vision. Check out the <a href='https://huggingface.co/datasets/mnist'>MNIST dataset</a> on the Hub for a computer vision example."
},
{
text: "Yes",
explain: "Correct! Check out the exciting developments with speech and vision in the 🤗 Transformers library to see how 🤗 Datasets is used in these domains.",
correct : true
},
]}
/>
| huggingface/course/blob/main/chapters/en/chapter5/8.mdx |
Quiz [[quiz]]
The best way to learn and [to avoid the illusion of competence](https://www.coursera.org/lecture/learning-how-to-learn/illusions-of-competence-BuFzf) **is to test yourself.** This will help you to find **where you need to reinforce your knowledge**.
### Q1: We mentioned Q Learning is a tabular method. What are tabular methods?
<details>
<summary>Solution</summary>
*Tabular methods* is a type of problem in which the state and actions spaces are small enough to approximate value functions to be **represented as arrays and tables**. For instance, **Q-Learning is a tabular method** since we use a table to represent the state, and action value pairs.
</details>
### Q2: Why can't we use a classical Q-Learning to solve an Atari Game?
<Question
choices={[
{
text: "Atari environments are too fast for Q-Learning",
explain: ""
},
{
text: "Atari environments have a big observation space. So creating an updating the Q-Table would not be efficient",
explain: "",
correct: true
}
]}
/>
### Q3: Why do we stack four frames together when we use frames as input in Deep Q-Learning?
<details>
<summary>Solution</summary>
We stack frames together because it helps us **handle the problem of temporal limitation**: one frame is not enough to capture temporal information.
For instance, in pong, our agent **will be unable to know the ball direction if it gets only one frame**.
<img src="https://huggingface.co/datasets/huggingface-deep-rl-course/course-images/resolve/main/en/unit4/temporal-limitation.jpg" alt="Temporal limitation"/>
<img src="https://huggingface.co/datasets/huggingface-deep-rl-course/course-images/resolve/main/en/unit4/temporal-limitation-2.jpg" alt="Temporal limitation"/>
</details>
### Q4: What are the two phases of Deep Q-Learning?
<Question
choices={[
{
text: "Sampling",
explain: "We perform actions and store the observed experiences tuples in a replay memory.",
correct: true,
},
{
text: "Shuffling",
explain: "",
},
{
text: "Reranking",
explain: "",
},
{
text: "Training",
explain: "We select the small batch of tuple randomly and learn from it using a gradient descent update step.",
correct: true,
}
]}
/>
### Q5: Why do we create a replay memory in Deep Q-Learning?
<details>
<summary>Solution</summary>
**1. Make more efficient use of the experiences during the training**
Usually, in online reinforcement learning, the agent interacts in the environment, gets experiences (state, action, reward, and next state), learns from them (updates the neural network), and discards them. This is not efficient.
But, with experience replay, **we create a replay buffer that saves experience samples that we can reuse during the training**.
**2. Avoid forgetting previous experiences and reduce the correlation between experiences**
The problem we get if we give sequential samples of experiences to our neural network is that it **tends to forget the previous experiences as it overwrites new experiences**. For instance, if we are in the first level and then the second, which is different, our agent can forget how to behave and play in the first level.
</details>
### Q6: How do we use Double Deep Q-Learning?
<details>
<summary>Solution</summary>
When we compute the Q target, we use two networks to decouple the action selection from the target Q value generation. We:
- Use our *DQN network* to **select the best action to take for the next state** (the action with the highest Q value).
- Use our *Target network* to calculate **the target Q value of taking that action at the next state**.
</details>
Congrats on finishing this Quiz 🥳, if you missed some elements, take time to read again the chapter to reinforce (😏) your knowledge.
| huggingface/deep-rl-class/blob/main/units/en/unit3/quiz.mdx |
DreamBooth training example
[DreamBooth](https://arxiv.org/abs/2208.12242) is a method to personalize text2image models like stable diffusion given just a few(3~5) images of a subject.
The `train_dreambooth.py` script shows how to implement the training procedure and adapt it for stable diffusion.
## Running locally with PyTorch
### Installing the dependencies
Before running the scripts, make sure to install the library's training dependencies:
**Important**
To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
```bash
git clone https://github.com/huggingface/diffusers
cd diffusers
pip install -e .
```
Then cd in the example folder and run
```bash
pip install -r requirements.txt
```
And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
```bash
accelerate config
```
Or for a default accelerate configuration without answering questions about your environment
```bash
accelerate config default
```
Or if your environment doesn't support an interactive shell e.g. a notebook
```python
from accelerate.utils import write_basic_config
write_basic_config()
```
When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.6.0` installed in your environment.
### Dog toy example
Now let's get our dataset. For this example we will use some dog images: https://huggingface.co/datasets/diffusers/dog-example.
Let's first download it locally:
```python
from huggingface_hub import snapshot_download
local_dir = "./dog"
snapshot_download(
"diffusers/dog-example",
local_dir=local_dir, repo_type="dataset",
ignore_patterns=".gitattributes",
)
```
And launch the training using:
**___Note: Change the `resolution` to 768 if you are using the [stable-diffusion-2](https://huggingface.co/stabilityai/stable-diffusion-2) 768x768 model.___**
```bash
export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="dog"
export OUTPUT_DIR="path-to-save-model"
accelerate launch train_dreambooth.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--output_dir=$OUTPUT_DIR \
--instance_prompt="a photo of sks dog" \
--resolution=512 \
--train_batch_size=1 \
--gradient_accumulation_steps=1 \
--learning_rate=5e-6 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=400 \
--push_to_hub
```
### Training with prior-preservation loss
Prior-preservation is used to avoid overfitting and language-drift. Refer to the paper to learn more about it. For prior-preservation we first generate images using the model with a class prompt and then use those during training along with our data.
According to the paper, it's recommended to generate `num_epochs * num_samples` images for prior-preservation. 200-300 works well for most cases. The `num_class_images` flag sets the number of images to generate with the class prompt. You can place existing images in `class_data_dir`, and the training script will generate any additional images so that `num_class_images` are present in `class_data_dir` during training time.
```bash
export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="dog"
export CLASS_DIR="path-to-class-images"
export OUTPUT_DIR="path-to-save-model"
accelerate launch train_dreambooth.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--class_data_dir=$CLASS_DIR \
--output_dir=$OUTPUT_DIR \
--with_prior_preservation --prior_loss_weight=1.0 \
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog" \
--resolution=512 \
--train_batch_size=1 \
--gradient_accumulation_steps=1 \
--learning_rate=5e-6 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--num_class_images=200 \
--max_train_steps=800 \
--push_to_hub
```
### Training on a 16GB GPU:
With the help of gradient checkpointing and the 8-bit optimizer from bitsandbytes it's possible to run train dreambooth on a 16GB GPU.
To install `bitsandbytes` please refer to this [readme](https://github.com/TimDettmers/bitsandbytes#requirements--installation).
```bash
export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="dog"
export CLASS_DIR="path-to-class-images"
export OUTPUT_DIR="path-to-save-model"
accelerate launch train_dreambooth.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--class_data_dir=$CLASS_DIR \
--output_dir=$OUTPUT_DIR \
--with_prior_preservation --prior_loss_weight=1.0 \
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog" \
--resolution=512 \
--train_batch_size=1 \
--gradient_accumulation_steps=2 --gradient_checkpointing \
--use_8bit_adam \
--learning_rate=5e-6 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--num_class_images=200 \
--max_train_steps=800 \
--push_to_hub
```
### Training on a 12GB GPU:
It is possible to run dreambooth on a 12GB GPU by using the following optimizations:
- [gradient checkpointing and the 8-bit optimizer](#training-on-a-16gb-gpu)
- [xformers](#training-with-xformers)
- [setting grads to none](#set-grads-to-none)
```bash
export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="dog"
export CLASS_DIR="path-to-class-images"
export OUTPUT_DIR="path-to-save-model"
accelerate launch train_dreambooth.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--class_data_dir=$CLASS_DIR \
--output_dir=$OUTPUT_DIR \
--with_prior_preservation --prior_loss_weight=1.0 \
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog" \
--resolution=512 \
--train_batch_size=1 \
--gradient_accumulation_steps=1 --gradient_checkpointing \
--use_8bit_adam \
--enable_xformers_memory_efficient_attention \
--set_grads_to_none \
--learning_rate=2e-6 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--num_class_images=200 \
--max_train_steps=800 \
--push_to_hub
```
### Training on a 8 GB GPU:
By using [DeepSpeed](https://www.deepspeed.ai/) it's possible to offload some
tensors from VRAM to either CPU or NVME allowing to train with less VRAM.
DeepSpeed needs to be enabled with `accelerate config`. During configuration
answer yes to "Do you want to use DeepSpeed?". With DeepSpeed stage 2, fp16
mixed precision and offloading both parameters and optimizer state to cpu it's
possible to train on under 8 GB VRAM with a drawback of requiring significantly
more RAM (about 25 GB). See [documentation](https://huggingface.co/docs/accelerate/usage_guides/deepspeed) for more DeepSpeed configuration options.
Changing the default Adam optimizer to DeepSpeed's special version of Adam
`deepspeed.ops.adam.DeepSpeedCPUAdam` gives a substantial speedup but enabling
it requires CUDA toolchain with the same version as pytorch. 8-bit optimizer
does not seem to be compatible with DeepSpeed at the moment.
```bash
export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="dog"
export CLASS_DIR="path-to-class-images"
export OUTPUT_DIR="path-to-save-model"
accelerate launch --mixed_precision="fp16" train_dreambooth.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--class_data_dir=$CLASS_DIR \
--output_dir=$OUTPUT_DIR \
--with_prior_preservation --prior_loss_weight=1.0 \
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog" \
--resolution=512 \
--train_batch_size=1 \
--sample_batch_size=1 \
--gradient_accumulation_steps=1 --gradient_checkpointing \
--learning_rate=5e-6 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--num_class_images=200 \
--max_train_steps=800 \
--push_to_hub
```
### Fine-tune text encoder with the UNet.
The script also allows to fine-tune the `text_encoder` along with the `unet`. It's been observed experimentally that fine-tuning `text_encoder` gives much better results especially on faces.
Pass the `--train_text_encoder` argument to the script to enable training `text_encoder`.
___Note: Training text encoder requires more memory, with this option the training won't fit on 16GB GPU. It needs at least 24GB VRAM.___
```bash
export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export INSTANCE_DIR="dog"
export CLASS_DIR="path-to-class-images"
export OUTPUT_DIR="path-to-save-model"
accelerate launch train_dreambooth.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_text_encoder \
--instance_data_dir=$INSTANCE_DIR \
--class_data_dir=$CLASS_DIR \
--output_dir=$OUTPUT_DIR \
--with_prior_preservation --prior_loss_weight=1.0 \
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog" \
--resolution=512 \
--train_batch_size=1 \
--use_8bit_adam \
--gradient_checkpointing \
--learning_rate=2e-6 \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--num_class_images=200 \
--max_train_steps=800 \
--push_to_hub
```
### Using DreamBooth for pipelines other than Stable Diffusion
The [AltDiffusion pipeline](https://huggingface.co/docs/diffusers/api/pipelines/alt_diffusion) also supports dreambooth fine-tuning. The process is the same as above, all you need to do is replace the `MODEL_NAME` like this:
```
export MODEL_NAME="CompVis/stable-diffusion-v1-4" --> export MODEL_NAME="BAAI/AltDiffusion-m9"
or
export MODEL_NAME="CompVis/stable-diffusion-v1-4" --> export MODEL_NAME="BAAI/AltDiffusion"
```
### Inference
Once you have trained a model using the above command, you can run inference simply using the `StableDiffusionPipeline`. Make sure to include the `identifier` (e.g. sks in above example) in your prompt.
```python
from diffusers import StableDiffusionPipeline
import torch
model_id = "path-to-your-trained-model"
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
prompt = "A photo of sks dog in a bucket"
image = pipe(prompt, num_inference_steps=50, guidance_scale=7.5).images[0]
image.save("dog-bucket.png")
```
### Inference from a training checkpoint
You can also perform inference from one of the checkpoints saved during the training process, if you used the `--checkpointing_steps` argument. Please, refer to [the documentation](https://huggingface.co/docs/diffusers/main/en/training/dreambooth#performing-inference-using-a-saved-checkpoint) to see how to do it.
## Training with Low-Rank Adaptation of Large Language Models (LoRA)
Low-Rank Adaption of Large Language Models was first introduced by Microsoft in [LoRA: Low-Rank Adaptation of Large Language Models](https://arxiv.org/abs/2106.09685) by *Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen*
In a nutshell, LoRA allows to adapt pretrained models by adding pairs of rank-decomposition matrices to existing weights and **only** training those newly added weights. This has a couple of advantages:
- Previous pretrained weights are kept frozen so that the model is not prone to [catastrophic forgetting](https://www.pnas.org/doi/10.1073/pnas.1611835114)
- Rank-decomposition matrices have significantly fewer parameters than the original model, which means that trained LoRA weights are easily portable.
- LoRA attention layers allow to control to which extent the model is adapted towards new training images via a `scale` parameter.
[cloneofsimo](https://github.com/cloneofsimo) was the first to try out LoRA training for Stable Diffusion in
the popular [lora](https://github.com/cloneofsimo/lora) GitHub repository.
### Training
Let's get started with a simple example. We will re-use the dog example of the [previous section](#dog-toy-example).
First, you need to set-up your dreambooth training example as is explained in the [installation section](#Installing-the-dependencies).
Next, let's download the dog dataset. Download images from [here](https://drive.google.com/drive/folders/1BO_dyz-p65qhBRRMRA4TbZ8qW4rB99JZ) and save them in a directory. Make sure to set `INSTANCE_DIR` to the name of your directory further below. This will be our training data.
Now, you can launch the training. Here we will use [Stable Diffusion 1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5).
**___Note: Change the `resolution` to 768 if you are using the [stable-diffusion-2](https://huggingface.co/stabilityai/stable-diffusion-2) 768x768 model.___**
**___Note: It is quite useful to monitor the training progress by regularly generating sample images during training. [wandb](https://docs.wandb.ai/quickstart) is a nice solution to easily see generating images during training. All you need to do is to run `pip install wandb` before training and pass `--report_to="wandb"` to automatically log images.___**
```bash
export MODEL_NAME="runwayml/stable-diffusion-v1-5"
export INSTANCE_DIR="dog"
export OUTPUT_DIR="path-to-save-model"
```
For this example we want to directly store the trained LoRA embeddings on the Hub, so
we need to be logged in and add the `--push_to_hub` flag.
```bash
huggingface-cli login
```
Now we can start training!
```bash
accelerate launch train_dreambooth_lora.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--output_dir=$OUTPUT_DIR \
--instance_prompt="a photo of sks dog" \
--resolution=512 \
--train_batch_size=1 \
--gradient_accumulation_steps=1 \
--checkpointing_steps=100 \
--learning_rate=1e-4 \
--report_to="wandb" \
--lr_scheduler="constant" \
--lr_warmup_steps=0 \
--max_train_steps=500 \
--validation_prompt="A photo of sks dog in a bucket" \
--validation_epochs=50 \
--seed="0" \
--push_to_hub
```
**___Note: When using LoRA we can use a much higher learning rate compared to vanilla dreambooth. Here we
use *1e-4* instead of the usual *2e-6*.___**
The final LoRA embedding weights have been uploaded to [patrickvonplaten/lora_dreambooth_dog_example](https://huggingface.co/patrickvonplaten/lora_dreambooth_dog_example). **___Note: [The final weights](https://huggingface.co/patrickvonplaten/lora/blob/main/pytorch_attn_procs.bin) are only 3 MB in size which is orders of magnitudes smaller than the original model.**
The training results are summarized [here](https://api.wandb.ai/report/patrickvonplaten/xm6cd5q5).
You can use the `Step` slider to see how the model learned the features of our subject while the model trained.
Optionally, we can also train additional LoRA layers for the text encoder. Specify the `--train_text_encoder` argument above for that. If you're interested to know more about how we
enable this support, check out this [PR](https://github.com/huggingface/diffusers/pull/2918).
With the default hyperparameters from the above, the training seems to go in a positive direction. Check out [this panel](https://wandb.ai/sayakpaul/dreambooth-lora/reports/test-23-04-17-17-00-13---Vmlldzo0MDkwNjMy). The trained LoRA layers are available [here](https://huggingface.co/sayakpaul/dreambooth).
### Inference
After training, LoRA weights can be loaded very easily into the original pipeline. First, you need to
load the original pipeline:
```python
from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
import torch
pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
pipe.to("cuda")
```
Next, we can load the adapter layers into the UNet with the [`load_attn_procs` function](https://huggingface.co/docs/diffusers/api/loaders#diffusers.loaders.UNet2DConditionLoadersMixin.load_attn_procs).
```python
pipe.unet.load_attn_procs("patrickvonplaten/lora_dreambooth_dog_example")
```
Finally, we can run the model in inference.
```python
image = pipe("A picture of a sks dog in a bucket", num_inference_steps=25).images[0]
```
If you are loading the LoRA parameters from the Hub and if the Hub repository has
a `base_model` tag (such as [this](https://huggingface.co/patrickvonplaten/lora_dreambooth_dog_example/blob/main/README.md?code=true#L4)), then
you can do:
```py
from huggingface_hub.repocard import RepoCard
lora_model_id = "patrickvonplaten/lora_dreambooth_dog_example"
card = RepoCard.load(lora_model_id)
base_model_id = card.data.to_dict()["base_model"]
pipe = StableDiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16)
...
```
If you used `--train_text_encoder` during training, then use `pipe.load_lora_weights()` to load the LoRA
weights. For example:
```python
from huggingface_hub.repocard import RepoCard
from diffusers import StableDiffusionPipeline
import torch
lora_model_id = "sayakpaul/dreambooth-text-encoder-test"
card = RepoCard.load(lora_model_id)
base_model_id = card.data.to_dict()["base_model"]
pipe = StableDiffusionPipeline.from_pretrained(base_model_id, torch_dtype=torch.float16)
pipe = pipe.to("cuda")
pipe.load_lora_weights(lora_model_id)
image = pipe("A picture of a sks dog in a bucket", num_inference_steps=25).images[0]
```
Note that the use of [`LoraLoaderMixin.load_lora_weights`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights) is preferred to [`UNet2DConditionLoadersMixin.load_attn_procs`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.UNet2DConditionLoadersMixin.load_attn_procs) for loading LoRA parameters. This is because
`LoraLoaderMixin.load_lora_weights` can handle the following situations:
* LoRA parameters that don't have separate identifiers for the UNet and the text encoder (such as [`"patrickvonplaten/lora_dreambooth_dog_example"`](https://huggingface.co/patrickvonplaten/lora_dreambooth_dog_example)). So, you can just do:
```py
pipe.load_lora_weights(lora_model_path)
```
* LoRA parameters that have separate identifiers for the UNet and the text encoder such as: [`"sayakpaul/dreambooth"`](https://huggingface.co/sayakpaul/dreambooth).
## Training with Flax/JAX
For faster training on TPUs and GPUs you can leverage the flax training example. Follow the instructions above to get the model and dataset before running the script.
____Note: The flax example don't yet support features like gradient checkpoint, gradient accumulation etc, so to use flax for faster training we will need >30GB cards.___
Before running the scripts, make sure to install the library's training dependencies:
```bash
pip install -U -r requirements_flax.txt
```
### Training without prior preservation loss
```bash
export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
export INSTANCE_DIR="dog"
export OUTPUT_DIR="path-to-save-model"
python train_dreambooth_flax.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--output_dir=$OUTPUT_DIR \
--instance_prompt="a photo of sks dog" \
--resolution=512 \
--train_batch_size=1 \
--learning_rate=5e-6 \
--max_train_steps=400
```
### Training with prior preservation loss
```bash
export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
export INSTANCE_DIR="dog"
export CLASS_DIR="path-to-class-images"
export OUTPUT_DIR="path-to-save-model"
python train_dreambooth_flax.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--class_data_dir=$CLASS_DIR \
--output_dir=$OUTPUT_DIR \
--with_prior_preservation --prior_loss_weight=1.0 \
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog" \
--resolution=512 \
--train_batch_size=1 \
--learning_rate=5e-6 \
--num_class_images=200 \
--max_train_steps=800
```
### Fine-tune text encoder with the UNet.
```bash
export MODEL_NAME="duongna/stable-diffusion-v1-4-flax"
export INSTANCE_DIR="dog"
export CLASS_DIR="path-to-class-images"
export OUTPUT_DIR="path-to-save-model"
python train_dreambooth_flax.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--train_text_encoder \
--instance_data_dir=$INSTANCE_DIR \
--class_data_dir=$CLASS_DIR \
--output_dir=$OUTPUT_DIR \
--with_prior_preservation --prior_loss_weight=1.0 \
--instance_prompt="a photo of sks dog" \
--class_prompt="a photo of dog" \
--resolution=512 \
--train_batch_size=1 \
--learning_rate=2e-6 \
--num_class_images=200 \
--max_train_steps=800
```
### Training with xformers:
You can enable memory efficient attention by [installing xFormers](https://github.com/facebookresearch/xformers#installing-xformers) and padding the `--enable_xformers_memory_efficient_attention` argument to the script. This is not available with the Flax/JAX implementation.
You can also use Dreambooth to train the specialized in-painting model. See [the script in the research folder for details](https://github.com/huggingface/diffusers/tree/main/examples/research_projects/dreambooth_inpaint).
### Set grads to none
To save even more memory, pass the `--set_grads_to_none` argument to the script. This will set grads to None instead of zero. However, be aware that it changes certain behaviors, so if you start experiencing any problems, remove this argument.
More info: https://pytorch.org/docs/stable/generated/torch.optim.Optimizer.zero_grad.html
### Experimental results
You can refer to [this blog post](https://huggingface.co/blog/dreambooth) that discusses some of DreamBooth experiments in detail. Specifically, it recommends a set of DreamBooth-specific tips and tricks that we have found to work well for a variety of subjects.
## IF
You can use the lora and full dreambooth scripts to train the text to image [IF model](https://huggingface.co/DeepFloyd/IF-I-XL-v1.0) and the stage II upscaler
[IF model](https://huggingface.co/DeepFloyd/IF-II-L-v1.0).
Note that IF has a predicted variance, and our finetuning scripts only train the models predicted error, so for finetuned IF models we switch to a fixed
variance schedule. The full finetuning scripts will update the scheduler config for the full saved model. However, when loading saved LoRA weights, you
must also update the pipeline's scheduler config.
```py
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained("DeepFloyd/IF-I-XL-v1.0")
pipe.load_lora_weights("<lora weights path>")
# Update scheduler config to fixed variance schedule
pipe.scheduler = pipe.scheduler.__class__.from_config(pipe.scheduler.config, variance_type="fixed_small")
```
Additionally, a few alternative cli flags are needed for IF.
`--resolution=64`: IF is a pixel space diffusion model. In order to operate on un-compressed pixels, the input images are of a much smaller resolution.
`--pre_compute_text_embeddings`: IF uses [T5](https://huggingface.co/docs/transformers/model_doc/t5) for its text encoder. In order to save GPU memory, we pre compute all text embeddings and then de-allocate
T5.
`--tokenizer_max_length=77`: T5 has a longer default text length, but the default IF encoding procedure uses a smaller number.
`--text_encoder_use_attention_mask`: T5 passes the attention mask to the text encoder.
### Tips and Tricks
We find LoRA to be sufficient for finetuning the stage I model as the low resolution of the model makes representing finegrained detail hard regardless.
For common and/or not-visually complex object concepts, you can get away with not-finetuning the upscaler. Just be sure to adjust the prompt passed to the
upscaler to remove the new token from the instance prompt. I.e. if your stage I prompt is "a sks dog", use "a dog" for your stage II prompt.
For finegrained detail like faces that aren't present in the original training set, we find that full finetuning of the stage II upscaler is better than
LoRA finetuning stage II.
For finegrained detail like faces, we find that lower learning rates along with larger batch sizes work best.
For stage II, we find that lower learning rates are also needed.
We found experimentally that the DDPM scheduler with the default larger number of denoising steps to sometimes work better than the DPM Solver scheduler
used in the training scripts.
### Stage II additional validation images
The stage II validation requires images to upscale, we can download a downsized version of the training set:
```py
from huggingface_hub import snapshot_download
local_dir = "./dog_downsized"
snapshot_download(
"diffusers/dog-example-downsized",
local_dir=local_dir,
repo_type="dataset",
ignore_patterns=".gitattributes",
)
```
### IF stage I LoRA Dreambooth
This training configuration requires ~28 GB VRAM.
```sh
export MODEL_NAME="DeepFloyd/IF-I-XL-v1.0"
export INSTANCE_DIR="dog"
export OUTPUT_DIR="dreambooth_dog_lora"
accelerate launch train_dreambooth_lora.py \
--report_to wandb \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--output_dir=$OUTPUT_DIR \
--instance_prompt="a sks dog" \
--resolution=64 \
--train_batch_size=4 \
--gradient_accumulation_steps=1 \
--learning_rate=5e-6 \
--scale_lr \
--max_train_steps=1200 \
--validation_prompt="a sks dog" \
--validation_epochs=25 \
--checkpointing_steps=100 \
--pre_compute_text_embeddings \
--tokenizer_max_length=77 \
--text_encoder_use_attention_mask
```
### IF stage II LoRA Dreambooth
`--validation_images`: These images are upscaled during validation steps.
`--class_labels_conditioning=timesteps`: Pass additional conditioning to the UNet needed for stage II.
`--learning_rate=1e-6`: Lower learning rate than stage I.
`--resolution=256`: The upscaler expects higher resolution inputs
```sh
export MODEL_NAME="DeepFloyd/IF-II-L-v1.0"
export INSTANCE_DIR="dog"
export OUTPUT_DIR="dreambooth_dog_upscale"
export VALIDATION_IMAGES="dog_downsized/image_1.png dog_downsized/image_2.png dog_downsized/image_3.png dog_downsized/image_4.png"
python train_dreambooth_lora.py \
--report_to wandb \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--output_dir=$OUTPUT_DIR \
--instance_prompt="a sks dog" \
--resolution=256 \
--train_batch_size=4 \
--gradient_accumulation_steps=1 \
--learning_rate=1e-6 \
--max_train_steps=2000 \
--validation_prompt="a sks dog" \
--validation_epochs=100 \
--checkpointing_steps=500 \
--pre_compute_text_embeddings \
--tokenizer_max_length=77 \
--text_encoder_use_attention_mask \
--validation_images $VALIDATION_IMAGES \
--class_labels_conditioning=timesteps
```
### IF Stage I Full Dreambooth
`--skip_save_text_encoder`: When training the full model, this will skip saving the entire T5 with the finetuned model. You can still load the pipeline
with a T5 loaded from the original model.
`use_8bit_adam`: Due to the size of the optimizer states, we recommend training the full XL IF model with 8bit adam.
`--learning_rate=1e-7`: For full dreambooth, IF requires very low learning rates. With higher learning rates model quality will degrade. Note that it is
likely the learning rate can be increased with larger batch sizes.
Using 8bit adam and a batch size of 4, the model can be trained in ~48 GB VRAM.
`--validation_scheduler`: Set a particular scheduler via a string. We found that it is better to use the DDPMScheduler for validation when training DeepFloyd IF.
```sh
export MODEL_NAME="DeepFloyd/IF-I-XL-v1.0"
export INSTANCE_DIR="dog"
export OUTPUT_DIR="dreambooth_if"
accelerate launch train_dreambooth.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--output_dir=$OUTPUT_DIR \
--instance_prompt="a photo of sks dog" \
--resolution=64 \
--train_batch_size=4 \
--gradient_accumulation_steps=1 \
--learning_rate=1e-7 \
--max_train_steps=150 \
--validation_prompt "a photo of sks dog" \
--validation_steps 25 \
--text_encoder_use_attention_mask \
--tokenizer_max_length 77 \
--pre_compute_text_embeddings \
--use_8bit_adam \
--set_grads_to_none \
--skip_save_text_encoder \
--validation_scheduler DDPMScheduler \
--push_to_hub
```
### IF Stage II Full Dreambooth
`--learning_rate=5e-6`: With a smaller effective batch size of 4, we found that we required learning rates as low as
1e-8.
`--resolution=256`: The upscaler expects higher resolution inputs
`--train_batch_size=2` and `--gradient_accumulation_steps=6`: We found that full training of stage II particularly with
faces required large effective batch sizes.
```sh
export MODEL_NAME="DeepFloyd/IF-II-L-v1.0"
export INSTANCE_DIR="dog"
export OUTPUT_DIR="dreambooth_dog_upscale"
export VALIDATION_IMAGES="dog_downsized/image_1.png dog_downsized/image_2.png dog_downsized/image_3.png dog_downsized/image_4.png"
accelerate launch train_dreambooth.py \
--report_to wandb \
--pretrained_model_name_or_path=$MODEL_NAME \
--instance_data_dir=$INSTANCE_DIR \
--output_dir=$OUTPUT_DIR \
--instance_prompt="a sks dog" \
--resolution=256 \
--train_batch_size=2 \
--gradient_accumulation_steps=6 \
--learning_rate=5e-6 \
--max_train_steps=2000 \
--validation_prompt="a sks dog" \
--validation_steps=150 \
--checkpointing_steps=500 \
--pre_compute_text_embeddings \
--tokenizer_max_length=77 \
--text_encoder_use_attention_mask \
--validation_images $VALIDATION_IMAGES \
--class_labels_conditioning timesteps \
--validation_scheduler DDPMScheduler\
--push_to_hub
```
## Stable Diffusion XL
We support fine-tuning of the UNet shipped in [Stable Diffusion XL](https://huggingface.co/papers/2307.01952) with DreamBooth and LoRA via the `train_dreambooth_lora_sdxl.py` script. Please refer to the docs [here](./README_sdxl.md).
| huggingface/diffusers/blob/main/examples/dreambooth/README.md |
How to Create a Chatbot with Gradio
Tags: NLP, TEXT, CHAT
## Introduction
Chatbots are a popular application of large language models. Using `gradio`, you can easily build a demo of your chatbot model and share that with your users, or try it yourself using an intuitive chatbot UI.
This tutorial uses `gr.ChatInterface()`, which is a high-level abstraction that allows you to create your chatbot UI fast, often with a single line of code. The chatbot interface that we create will look something like this:
$demo_chatinterface_streaming_echo
We'll start with a couple of simple examples, and then show how to use `gr.ChatInterface()` with real language models from several popular APIs and libraries, including `langchain`, `openai`, and Hugging Face.
**Prerequisites**: please make sure you are using the **latest version** version of Gradio:
```bash
$ pip install --upgrade gradio
```
## Defining a chat function
When working with `gr.ChatInterface()`, the first thing you should do is define your chat function. Your chat function should take two arguments: `message` and then `history` (the arguments can be named anything, but must be in this order).
- `message`: a `str` representing the user's input.
- `history`: a `list` of `list` representing the conversations up until that point. Each inner list consists of two `str` representing a pair: `[user input, bot response]`.
Your function should return a single string response, which is the bot's response to the particular user input `message`. Your function can take into account the `history` of messages, as well as the current message.
Let's take a look at a few examples.
## Example: a chatbot that responds yes or no
Let's write a chat function that responds `Yes` or `No` randomly.
Here's our chat function:
```python
import random
def random_response(message, history):
return random.choice(["Yes", "No"])
```
Now, we can plug this into `gr.ChatInterface()` and call the `.launch()` method to create the web interface:
```python
import gradio as gr
gr.ChatInterface(random_response).launch()
```
That's it! Here's our running demo, try it out:
$demo_chatinterface_random_response
## Another example using the user's input and history
Of course, the previous example was very simplistic, it didn't even take user input or the previous history into account! Here's another simple example showing how to incorporate a user's input as well as the history.
```python
import random
import gradio as gr
def alternatingly_agree(message, history):
if len(history) % 2 == 0:
return f"Yes, I do think that '{message}'"
else:
return "I don't think so"
gr.ChatInterface(alternatingly_agree).launch()
```
## Streaming chatbots
If in your chat function, you use `yield` to generate a sequence of responses, you'll end up with a streaming chatbot. It's that simple!
```python
import time
import gradio as gr
def slow_echo(message, history):
for i in range(len(message)):
time.sleep(0.3)
yield "You typed: " + message[: i+1]
gr.ChatInterface(slow_echo).launch()
```
Notice that we've [enabled queuing](/guides/key-features#queuing), which is required to use generator functions. While the response is streaming, the "Submit" button turns into a "Stop" button that can be used to stop the generator function. You can customize the appearance and behavior of the "Stop" button using the `stop_btn` parameter.
## Customizing your chatbot
If you're familiar with Gradio's `Interface` class, the `gr.ChatInterface` includes many of the same arguments that you can use to customize the look and feel of your Chatbot. For example, you can:
- add a title and description above your chatbot using `title` and `description` arguments.
- add a theme or custom css using `theme` and `css` arguments respectively.
- add `examples` and even enable `cache_examples`, which make it easier for users to try it out .
- You can change the text or disable each of the buttons that appear in the chatbot interface: `submit_btn`, `retry_btn`, `undo_btn`, `clear_btn`.
If you want to customize the `gr.Chatbot` or `gr.Textbox` that compose the `ChatInterface`, then you can pass in your own chatbot or textbox as well. Here's an example of how we can use these parameters:
```python
import gradio as gr
def yes_man(message, history):
if message.endswith("?"):
return "Yes"
else:
return "Ask me anything!"
gr.ChatInterface(
yes_man,
chatbot=gr.Chatbot(height=300),
textbox=gr.Textbox(placeholder="Ask me a yes or no question", container=False, scale=7),
title="Yes Man",
description="Ask Yes Man any question",
theme="soft",
examples=["Hello", "Am I cool?", "Are tomatoes vegetables?"],
cache_examples=True,
retry_btn=None,
undo_btn="Delete Previous",
clear_btn="Clear",
).launch()
```
## Additional Inputs
You may want to add additional parameters to your chatbot and expose them to your users through the Chatbot UI. For example, suppose you want to add a textbox for a system prompt, or a slider that sets the number of tokens in the chatbot's response. The `ChatInterface` class supports an `additional_inputs` parameter which can be used to add additional input components.
The `additional_inputs` parameters accepts a component or a list of components. You can pass the component instances directly, or use their string shortcuts (e.g. `"textbox"` instead of `gr.Textbox()`). If you pass in component instances, and they have _not_ already been rendered, then the components will appear underneath the chatbot (and any examples) within a `gr.Accordion()`. You can set the label of this accordion using the `additional_inputs_accordion_name` parameter.
Here's a complete example:
$code_chatinterface_system_prompt
If the components you pass into the `additional_inputs` have already been rendered in a parent `gr.Blocks()`, then they will _not_ be re-rendered in the accordion. This provides flexibility in deciding where to lay out the input components. In the example below, we position the `gr.Textbox()` on top of the Chatbot UI, while keeping the slider underneath.
```python
import gradio as gr
import time
def echo(message, history, system_prompt, tokens):
response = f"System prompt: {system_prompt}\n Message: {message}."
for i in range(min(len(response), int(tokens))):
time.sleep(0.05)
yield response[: i+1]
with gr.Blocks() as demo:
system_prompt = gr.Textbox("You are helpful AI.", label="System Prompt")
slider = gr.Slider(10, 100, render=False)
gr.ChatInterface(
echo, additional_inputs=[system_prompt, slider]
)
demo.launch()
```
If you need to create something even more custom, then its best to construct the chatbot UI using the low-level `gr.Blocks()` API. We have [a dedicated guide for that here](/guides/creating-a-custom-chatbot-with-blocks).
## Using your chatbot via an API
Once you've built your Gradio chatbot and are hosting it on [Hugging Face Spaces](https://hf.space) or somewhere else, then you can query it with a simple API at the `/chat` endpoint. The endpoint just expects the user's message (and potentially additional inputs if you have set any using the `additional_inputs` parameter), and will return the response, internally keeping track of the messages sent so far.
[](https://github.com/gradio-app/gradio/assets/1778297/7b10d6db-6476-4e2e-bebd-ecda802c3b8f)
To use the endpoint, you should use either the [Gradio Python Client](/guides/getting-started-with-the-python-client) or the [Gradio JS client](/guides/getting-started-with-the-js-client).
## A `langchain` example
Now, let's actually use the `gr.ChatInterface` with some real large language models. We'll start by using `langchain` on top of `openai` to build a general-purpose streaming chatbot application in 19 lines of code. You'll need to have an OpenAI key for this example (keep reading for the free, open-source equivalent!)
```python
from langchain.chat_models import ChatOpenAI
from langchain.schema import AIMessage, HumanMessage
import openai
import gradio as gr
os.environ["OPENAI_API_KEY"] = "sk-..." # Replace with your key
llm = ChatOpenAI(temperature=1.0, model='gpt-3.5-turbo-0613')
def predict(message, history):
history_langchain_format = []
for human, ai in history:
history_langchain_format.append(HumanMessage(content=human))
history_langchain_format.append(AIMessage(content=ai))
history_langchain_format.append(HumanMessage(content=message))
gpt_response = llm(history_langchain_format)
return gpt_response.content
gr.ChatInterface(predict).launch()
```
## A streaming example using `openai`
Of course, we could also use the `openai` library directy. Here a similar example, but this time with streaming results as well:
```python
import openai
import gradio as gr
openai.api_key = "sk-..." # Replace with your key
def predict(message, history):
history_openai_format = []
for human, assistant in history:
history_openai_format.append({"role": "user", "content": human })
history_openai_format.append({"role": "assistant", "content":assistant})
history_openai_format.append({"role": "user", "content": message})
response = openai.ChatCompletion.create(
model='gpt-3.5-turbo',
messages= history_openai_format,
temperature=1.0,
stream=True
)
partial_message = ""
for chunk in response:
if len(chunk['choices'][0]['delta']) != 0:
partial_message = partial_message + chunk['choices'][0]['delta']['content']
yield partial_message
gr.ChatInterface(predict).launch()
```
## Example using a local, open-source LLM with Hugging Face
Of course, in many cases you want to run a chatbot locally. Here's the equivalent example using Together's RedePajama model, from Hugging Face (this requires you to have a GPU with CUDA).
```python
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer
from threading import Thread
tokenizer = AutoTokenizer.from_pretrained("togethercomputer/RedPajama-INCITE-Chat-3B-v1")
model = AutoModelForCausalLM.from_pretrained("togethercomputer/RedPajama-INCITE-Chat-3B-v1", torch_dtype=torch.float16)
model = model.to('cuda:0')
class StopOnTokens(StoppingCriteria):
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
stop_ids = [29, 0]
for stop_id in stop_ids:
if input_ids[0][-1] == stop_id:
return True
return False
def predict(message, history):
history_transformer_format = history + [[message, ""]]
stop = StopOnTokens()
messages = "".join(["".join(["\n<human>:"+item[0], "\n<bot>:"+item[1]]) #curr_system_message +
for item in history_transformer_format])
model_inputs = tokenizer([messages], return_tensors="pt").to("cuda")
streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
generate_kwargs = dict(
model_inputs,
streamer=streamer,
max_new_tokens=1024,
do_sample=True,
top_p=0.95,
top_k=1000,
temperature=1.0,
num_beams=1,
stopping_criteria=StoppingCriteriaList([stop])
)
t = Thread(target=model.generate, kwargs=generate_kwargs)
t.start()
partial_message = ""
for new_token in streamer:
if new_token != '<':
partial_message += new_token
yield partial_message
gr.ChatInterface(predict).launch()
```
With those examples, you should be all set to create your own Gradio Chatbot demos soon! For building even more custom Chatbot applications, check out [a dedicated guide](/guides/creating-a-custom-chatbot-with-blocks) using the low-level `gr.Blocks()` API.
| gradio-app/gradio/blob/main/guides/04_chatbots/01_creating-a-chatbot-fast.md |
Create a Dashboard from Supabase Data
Tags: TABULAR, DASHBOARD, PLOTS
[Supabase](https://supabase.com/) is a cloud-based open-source backend that provides a PostgreSQL database, authentication, and other useful features for building web and mobile applications. In this tutorial, you will learn how to read data from Supabase and plot it in **real-time** on a Gradio Dashboard.
**Prerequisites:** To start, you will need a free Supabase account, which you can sign up for here: [https://app.supabase.com/](https://app.supabase.com/)
In this end-to-end guide, you will learn how to:
- Create tables in Supabase
- Write data to Supabase using the Supabase Python Client
- Visualize the data in a real-time dashboard using Gradio
If you already have data on Supabase that you'd like to visualize in a dashboard, you can skip the first two sections and go directly to [visualizing the data](#visualize-the-data-in-a-real-time-gradio-dashboard)!
## Create a table in Supabase
First of all, we need some data to visualize. Following this [excellent guide](https://supabase.com/blog/loading-data-supabase-python), we'll create fake commerce data and put it in Supabase.
1\. Start by creating a new project in Supabase. Once you're logged in, click the "New Project" button
2\. Give your project a name and database password. You can also choose a pricing plan (for our purposes, the Free Tier is sufficient!)
3\. You'll be presented with your API keys while the database spins up (can take up to 2 minutes).
4\. Click on "Table Editor" (the table icon) in the left pane to create a new table. We'll create a single table called `Product`, with the following schema:
<center>
<table>
<tr><td>product_id</td><td>int8</td></tr>
<tr><td>inventory_count</td><td>int8</td></tr>
<tr><td>price</td><td>float8</td></tr>
<tr><td>product_name</td><td>varchar</td></tr>
</table>
</center>
5\. Click Save to save the table schema.
Our table is now ready!
## Write data to Supabase
The next step is to write data to a Supabase dataset. We will use the Supabase Python library to do this.
6\. Install `supabase` by running the following command in your terminal:
```bash
pip install supabase
```
7\. Get your project URL and API key. Click the Settings (gear icon) on the left pane and click 'API'. The URL is listed in the Project URL box, while the API key is listed in Project API keys (with the tags `service_role`, `secret`)
8\. Now, run the following Python script to write some fake data to the table (note you have to put the values of `SUPABASE_URL` and `SUPABASE_SECRET_KEY` from step 7):
```python
import supabase
# Initialize the Supabase client
client = supabase.create_client('SUPABASE_URL', 'SUPABASE_SECRET_KEY')
# Define the data to write
import random
main_list = []
for i in range(10):
value = {'product_id': i,
'product_name': f"Item {i}",
'inventory_count': random.randint(1, 100),
'price': random.random()*100
}
main_list.append(value)
# Write the data to the table
data = client.table('Product').insert(main_list).execute()
```
Return to your Supabase dashboard and refresh the page, you should now see 10 rows populated in the `Product` table!
## Visualize the Data in a Real-Time Gradio Dashboard
Finally, we will read the data from the Supabase dataset using the same `supabase` Python library and create a realtime dashboard using `gradio`.
Note: We repeat certain steps in this section (like creating the Supabase client) in case you did not go through the previous sections. As described in Step 7, you will need the project URL and API Key for your database.
9\. Write a function that loads the data from the `Product` table and returns it as a pandas Dataframe:
```python
import supabase
import pandas as pd
client = supabase.create_client('SUPABASE_URL', 'SUPABASE_SECRET_KEY')
def read_data():
response = client.table('Product').select("*").execute()
df = pd.DataFrame(response.data)
return df
```
10\. Create a small Gradio Dashboard with 2 Barplots that plots the prices and inventories of all of the items every minute and updates in real-time:
```python
import gradio as gr
with gr.Blocks() as dashboard:
with gr.Row():
gr.BarPlot(read_data, x="product_id", y="price", title="Prices", every=60)
gr.BarPlot(read_data, x="product_id", y="inventory_count", title="Inventory", every=60)
dashboard.queue().launch()
```
Notice that by passing in a function to `gr.BarPlot()`, we have the BarPlot query the database as soon as the web app loads (and then again every 60 seconds because of the `every` parameter). Your final dashboard should look something like this:
<gradio-app space="abidlabs/supabase"></gradio-app>
## Conclusion
That's it! In this tutorial, you learned how to write data to a Supabase dataset, and then read that data and plot the results as bar plots. If you update the data in the Supabase database, you'll notice that the Gradio dashboard will update within a minute.
Try adding more plots and visualizations to this example (or with a different dataset) to build a more complex dashboard!
| gradio-app/gradio/blob/main/guides/07_tabular-data-science-and-plots/creating-a-dashboard-from-supabase-data.md |
Deep Q-Learning [[deep-q-learning]]
<img src="https://huggingface.co/datasets/huggingface-deep-rl-course/course-images/resolve/main/en/unit4/thumbnail.jpg" alt="Unit 3 thumbnail" width="100%">
In the last unit, we learned our first reinforcement learning algorithm: Q-Learning, **implemented it from scratch**, and trained it in two environments, FrozenLake-v1 ☃️ and Taxi-v3 🚕.
We got excellent results with this simple algorithm, but these environments were relatively simple because the **state space was discrete and small** (16 different states for FrozenLake-v1 and 500 for Taxi-v3). For comparison, the state space in Atari games can **contain \\(10^{9}\\) to \\(10^{11}\\) states**.
But as we'll see, producing and updating a **Q-table can become ineffective in large state space environments.**
So in this unit, **we'll study our first Deep Reinforcement Learning agent**: Deep Q-Learning. Instead of using a Q-table, Deep Q-Learning uses a Neural Network that takes a state and approximates Q-values for each action based on that state.
And **we'll train it to play Space Invaders and other Atari environments using [RL-Zoo](https://github.com/DLR-RM/rl-baselines3-zoo)**, a training framework for RL using Stable-Baselines that provides scripts for training, evaluating agents, tuning hyperparameters, plotting results, and recording videos.
<img src="https://huggingface.co/datasets/huggingface-deep-rl-course/course-images/resolve/main/en/unit4/atari-envs.gif" alt="Environments"/>
So let’s get started! 🚀
| huggingface/deep-rl-class/blob/main/units/en/unit3/introduction.mdx |
Gradio Demo: hello_world_4
```
!pip install -q gradio
```
```
import gradio as gr
def greet(name, intensity):
return "Hello " * intensity + name + "!"
demo = gr.Interface(
fn=greet,
inputs=["text", "slider"],
outputs=["text"],
)
if __name__ == "__main__":
demo.launch()
```
| gradio-app/gradio/blob/main/demo/hello_world_4/run.ipynb |
--
title: COMET
emoji: 🤗
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false
tags:
- evaluate
- metric
description: >-
Crosslingual Optimized Metric for Evaluation of Translation (COMET) is an open-source framework used to train Machine Translation metrics that achieve high levels of correlation with different types of human judgments (HTER, DA's or MQM).
With the release of the framework the authors also released fully trained models that were used to compete in the WMT20 Metrics Shared Task achieving SOTA in that years competition.
See the [README.md] file at https://unbabel.github.io/COMET/html/models.html for more information.
---
# Metric Card for COMET
## Metric description
Crosslingual Optimized Metric for Evaluation of Translation (COMET) is an open-source framework used to train Machine Translation metrics that achieve high levels of correlation with different types of human judgments.
## How to use
COMET takes 3 lists of strings as input: `sources` (a list of source sentences), `predictions` (a list of candidate translations) and `references` (a list of reference translations).
```python
from evaluate import load
comet_metric = load('comet')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["The fire could be stopped", "Schools and kindergartens were open"]
reference = ["They were able to control the fire.", "Schools and kindergartens opened"]
comet_score = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
```
It has several configurations, named after the COMET model to be used. For versions below 2.0 it will default to `wmt20-comet-da` (previously known as `wmt-large-da-estimator-1719`) and for the latest versions (>= 2.0) it will default to `Unbabel/wmt22-comet-da`.
Alternative models that can be chosen include `wmt20-comet-qe-da`, `wmt21-comet-mqm`, `wmt21-cometinho-da`, `wmt21-comet-qe-mqm` and `emnlp20-comet-rank`. Notably, a distilled model is also available, which is 80% smaller and 2.128x faster while performing close to non-distilled alternatives. You can use it with the identifier `eamt22-cometinho-da`. This version, called Cometinho, was elected as [the best paper](https://aclanthology.org/2022.eamt-1.9) at the annual European conference on Machine Translation.
> NOTE: In `unbabel-comet>=2.0` all models were moved to Hugging Face Hub and you need to add the suffix `Unbabel/` to be able to download and use them. For example for the distilled version replace `eamt22-cometinho-da` with `Unbabel/eamt22-cometinho-da`.
It also has several optional arguments:
`gpus`: optional, an integer (number of GPUs to train on) or a list of integers (which GPUs to train on). Set to 0 to use CPU. The default value is `None` (uses one GPU if possible, else use CPU).
`progress_bar`a boolean -- if set to `True`, progress updates will be printed out. The default value is `False`.
More information about model characteristics can be found on the [COMET website](https://unbabel.github.io/COMET/html/index.html).
## Output values
The COMET metric outputs two lists:
`scores`: a list of COMET scores for each of the input sentences, ranging from 0-1.
`mean_score`: the mean value of COMET scores `scores` over all the input sentences, ranging from 0-1.
### Values from popular papers
The [original COMET paper](https://arxiv.org/pdf/2009.09025.pdf) reported average COMET scores ranging from 0.4 to 0.6, depending on the language pairs used for evaluating translation models. They also illustrate that COMET correlates well with human judgements compared to other metrics such as [BLEU](https://huggingface.co/metrics/bleu) and [CHRF](https://huggingface.co/metrics/chrf).
## Examples
Full match:
```python
from evaluate import load
comet_metric = load('comet')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["They were able to control the fire.", "Schools and kindergartens opened"]
reference = ["They were able to control the fire.", "Schools and kindergartens opened"]
results = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
print([round(v, 1) for v in results["scores"]])
[1.0, 1.0]
```
Partial match:
```python
from evaluate import load
comet_metric = load('comet')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["The fire could be stopped", "Schools and kindergartens were open"]
reference = ["They were able to control the fire", "Schools and kindergartens opened"]
results = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
print([round(v, 2) for v in results["scores"]])
[0.19, 0.92]
```
No match:
```python
from evaluate import load
comet_metric = load('comet')
source = ["Dem Feuer konnte Einhalt geboten werden", "Schulen und Kindergärten wurden eröffnet."]
hypothesis = ["The girl went for a walk", "The boy was sleeping"]
reference = ["They were able to control the fire", "Schools and kindergartens opened"]
results = comet_metric.compute(predictions=hypothesis, references=reference, sources=source)
print([round(v, 2) for v in results["scores"]])
[0.00, 0.00]
```
## Limitations and bias
The models provided for calculating the COMET metric are built on top of XLM-R and cover the following languages:
Afrikaans, Albanian, Amharic, Arabic, Armenian, Assamese, Azerbaijani, Basque, Belarusian, Bengali, Bengali Romanized, Bosnian, Breton, Bulgarian, Burmese, Burmese, Catalan, Chinese (Simplified), Chinese (Traditional), Croatian, Czech, Danish, Dutch, English, Esperanto, Estonian, Filipino, Finnish, French, Galician, Georgian, German, Greek, Gujarati, Hausa, Hebrew, Hindi, Hindi Romanized, Hungarian, Icelandic, Indonesian, Irish, Italian, Japanese, Javanese, Kannada, Kazakh, Khmer, Korean, Kurdish (Kurmanji), Kyrgyz, Lao, Latin, Latvian, Lithuanian, Macedonian, Malagasy, Malay, Malayalam, Marathi, Mongolian, Nepali, Norwegian, Oriya, Oromo, Pashto, Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Sanskri, Scottish, Gaelic, Serbian, Sindhi, Sinhala, Slovak, Slovenian, Somali, Spanish, Sundanese, Swahili, Swedish, Tamil, Tamil Romanized, Telugu, Telugu Romanized, Thai, Turkish, Ukrainian, Urdu, Urdu Romanized, Uyghur, Uzbek, Vietnamese, Welsh, Western, Frisian, Xhosa, Yiddish.
Thus, results for language pairs containing uncovered languages are unreliable, as per the [COMET website](https://github.com/Unbabel/COMET)
Also, calculating the COMET metric involves downloading the model from which features are obtained -- the default model, `wmt22-comet-da`, takes over 2.32GB of storage space and downloading it can take a significant amount of time depending on the speed of your internet connection. If this is an issue, choose a smaller model; for instance `eamt22-cometinho-da` is 344MB.
### Interpreting Scores:
When using COMET to evaluate machine translation, it's important to understand how to interpret the scores it produces.
In general, COMET models are trained to predict quality scores for translations. These scores are typically normalized using a z-score transformation to account for individual differences among annotators. While the raw score itself does not have a direct interpretation, it is useful for ranking translations and systems according to their quality.
However, for the latest COMET models like `Unbabel/wmt22-comet-da`, we have introduced a new training approach that scales the scores between 0 and 1. This makes it easier to interpret the scores: a score close to 1 indicates a high-quality translation, while a score close to 0 indicates a translation that is no better than random chance.
It's worth noting that when using COMET to compare the performance of two different translation systems, it's important to run statistical significance measures to reliably compare scores between systems.
## Citation
```bibtex
@inproceedings{rei-etal-2022-comet,
title = "{COMET}-22: Unbabel-{IST} 2022 Submission for the Metrics Shared Task",
author = "Rei, Ricardo and
C. de Souza, Jos{\'e} G. and
Alves, Duarte and
Zerva, Chrysoula and
Farinha, Ana C and
Glushkova, Taisiya and
Lavie, Alon and
Coheur, Luisa and
Martins, Andr{\'e} F. T.",
booktitle = "Proceedings of the Seventh Conference on Machine Translation (WMT)",
month = dec,
year = "2022",
address = "Abu Dhabi, United Arab Emirates (Hybrid)",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2022.wmt-1.52",
pages = "578--585",
}
```
```bibtex
@inproceedings{rei-EtAl:2020:WMT,
author = {Rei, Ricardo and Stewart, Craig and Farinha, Ana C and Lavie, Alon},
title = {Unbabel's Participation in the WMT20 Metrics Shared Task},
booktitle = {Proceedings of the Fifth Conference on Machine Translation},
month = {November},
year = {2020},
address = {Online},
publisher = {Association for Computational Linguistics},
pages = {909--918},
}
```
```bibtex
@inproceedings{rei-etal-2020-comet,
title = "{COMET}: A Neural Framework for {MT} Evaluation",
author = "Rei, Ricardo and
Stewart, Craig and
Farinha, Ana C and
Lavie, Alon",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.emnlp-main.213",
pages = "2685--2702",
}
```
For the distilled version:
```bibtex
@inproceedings{rei-etal-2022-searching,
title = "Searching for {COMETINHO}: The Little Metric That Could",
author = "Rei, Ricardo and
Farinha, Ana C and
de Souza, Jos{\'e} G.C. and
Ramos, Pedro G. and
Martins, Andr{\'e} F.T. and
Coheur, Luisa and
Lavie, Alon",
booktitle = "Proceedings of the 23rd Annual Conference of the European Association for Machine Translation",
month = jun,
year = "2022",
address = "Ghent, Belgium",
publisher = "European Association for Machine Translation",
url = "https://aclanthology.org/2022.eamt-1.9",
pages = "61--70",
}
```
## Further References
- [COMET website](https://unbabel.github.io/COMET/html/index.html)
- [Hugging Face Tasks - Machine Translation](https://huggingface.co/tasks/translation)
| huggingface/evaluate/blob/main/metrics/comet/README.md |
Using PEFT at Hugging Face
🤗 [Parameter-Efficient Fine-Tuning (PEFT)](https://huggingface.co/docs/peft/index) is a library for efficiently adapting pre-trained language models to various downstream applications without fine-tuning all the model’s parameters.
## Exploring PEFT on the Hub
You can find PEFT models by filtering at the left of the [models page](https://huggingface.co/models?library=peft&sort=trending).
## Installation
To get started, you can check out the [Quick Tour in the PEFT docs](https://huggingface.co/docs/peft/quicktour). To install, follow the [PEFT installation guide](https://huggingface.co/docs/peft/install).
You can also use the following one-line install through pip:
```
$ pip install peft
```
## Using existing models
All PEFT models can be loaded from the Hub. To use a PEFT model you also need to load the base model that was fine-tuned, as shown below. Every fine-tuned model has the base model in it's model card.
```py
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel, PeftConfig
base_model = "mistralai/Mistral-7B-v0.1"
adapter_model = "dfurman/Mistral-7B-Instruct-v0.2"
model = AutoModelForCausalLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, adapter_model)
tokenizer = AutoTokenizer.from_pretrained(base_model)
model = model.to("cuda")
model.eval()
```
Once loaded, you can pass your inputs to the tokenizer to prepare them, and call `model.generate()` in regular `transformers` fashion.
```py
inputs = tokenizer("Tell me the recipe for chocolate chip cookie", return_tensors="pt")
with torch.no_grad():
outputs = model.generate(input_ids=inputs["input_ids"].to("cuda"), max_new_tokens=10)
print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0])
```
It outputs the following:
```text
Tell me the recipe for chocolate chip cookie dough.
1. Preheat oven to 375 degrees F (190 degrees C).
2. In a large bowl, cream together 1/2 cup (1 stick) of butter or margarine, 1/2 cup granulated sugar, and 1/2 cup packed brown sugar.
3. Beat in 1 egg and 1 teaspoon vanilla extract.
4. Mix in 1 1/4 cups all-purpose flour.
5. Stir in 1/2 teaspoon baking soda and 1/2 teaspoon salt.
6. Fold in 3/4 cup semisweet chocolate chips.
7. Drop by
```
If you want to load a specific PEFT model, you can click `Use in PEFT` in the model card and you will be given a working snippet!
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/peft_repo_light_new.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/peft_repo.png"/>
</div>
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/peft_snippet_light.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/peft_snippet.png"/>
</div>
## Additional resources
* PEFT [repository](https://github.com/huggingface/peft)
* PEFT [docs](https://huggingface.co/docs/peft/index)
* PEFT [models](https://huggingface.co/models?library=peft&sort=trending)
| huggingface/hub-docs/blob/main/docs/hub/peft.md |
--
title: "Hugging Face and AWS partner to make AI more accessible"
thumbnail: /blog/assets/131_aws-partnership/aws-partnership-thumbnail.png
authors:
- user: jeffboudier
- user: philschmid
- user: juliensimon
---
# Hugging Face and AWS partner to make AI more accessible
It’s time to make AI open and accessible to all. That’s the goal of this expanded long-term strategic partnership between Hugging Face and Amazon Web Services (AWS). Together, the two leaders aim to accelerate the availability of next-generation machine learning models by making them more accessible to the machine learning community and helping developers achieve the highest performance at the lowest cost.
## A new generation of open, accessible AI
Machine learning is quickly becoming embedded in all applications. As its impact on every sector of the economy comes into focus, it’s more important than ever to ensure every developer can access and assess the latest models. The partnership with AWS paves the way toward this future by making it faster and easier to build, train, and deploy the latest machine learning models in the cloud using purpose-built tools.
There have been significant advances in new Transformer and Diffuser machine learning models that process and generate text, audio, and images. However, most of these popular generative AI models are not publicly available, widening the gap of machine learning capabilities between the largest tech companies and everyone else. To counter this trend, AWS and Hugging Face are partnering to contribute next-generation models to the global AI community and democratize machine learning. Through the strategic partnership, Hugging Face will leverage AWS as a preferred cloud provider so developers in Hugging Face’s community can access AWS’s state-of-the-art tools (e.g., [Amazon SageMaker](https://aws.amazon.com/sagemaker), [AWS Trainium](https://aws.amazon.com/machine-learning/trainium/), [AWS Inferentia](https://aws.amazon.com/machine-learning/inferentia/)) to train, fine-tune, and deploy models on AWS. This will allow developers to further optimize the performance of their models for their specific use cases while lowering costs. Hugging Face will apply the latest in innovative research findings using Amazon SageMaker to build next-generation AI models. Together, Hugging Face and AWS are bridging the gap so the global AI community can benefit from the latest advancements in machine learning to accelerate the creation of generative AI applications.
“The future of AI is here, but it’s not evenly distributed,” said Clement Delangue, CEO of Hugging Face. “Accessibility and transparency are the keys to sharing progress and creating tools to use these new capabilities wisely and responsibly. Amazon SageMaker and AWS-designed chips will enable our team and the larger machine learning community to convert the latest research into openly reproducible models that anyone can build on.”
## Collaborating to scale AI in the cloud
This expanded strategic partnership enables Hugging Face and AWS to accelerate machine learning adoption using the latest models hosted on Hugging Face with the industry-leading capabilities of Amazon SageMaker. Customers can now easily fine-tune and deploy state-of-the-art Hugging Face models in just a few clicks on Amazon SageMaker and Amazon Elastic Computing Cloud (EC2), taking advantage of purpose-built machine learning accelerators including AWS Trainium and AWS Inferentia.
“Generative AI has the potential to transform entire industries, but its cost and the required expertise puts the technology out of reach for all but a select few companies,” said Adam Selipsky, CEO of AWS. “Hugging Face and AWS are making it easier for customers to access popular machine learning models to create their own generative AI applications with the highest performance and lowest costs. This partnership demonstrates how generative AI companies and AWS can work together to put this innovative technology into the hands of more customers.”
Hugging Face has become the central hub for machine learning, with more than [100,000 free and accessible machine learning models](https://huggingface.co/models) downloaded more than 1 million times daily by researchers, data scientists, and machine learning engineers. AWS is by far the most popular place to run models from the Hugging Face Hub. Since the [start of our collaboration](https://huggingface.co/blog/the-partnership-amazon-sagemaker-and-hugging-face), [Hugging Face on Amazon SageMaker](https://aws.amazon.com/machine-learning/hugging-face/) has grown exponentially. We are experiencing an exciting renaissance with generative AI, and we're just getting started. We look forward to what the future holds for Hugging Face, AWS, and the AI community.
| huggingface/blog/blob/main/aws-partnership.md |
Gradio Demo: fake_gan_no_input
```
!pip install -q gradio
```
```
import time
import gradio as gr
def fake_gan():
time.sleep(1)
images = [
"https://images.unsplash.com/photo-1507003211169-0a1dd7228f2d?ixlib=rb-1.2.1&ixid=MnwxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8&auto=format&fit=crop&w=387&q=80",
"https://images.unsplash.com/photo-1554151228-14d9def656e4?ixlib=rb-1.2.1&ixid=MnwxMjA3fDB8MHxwaG90by1wYWdlfHx8fGVufDB8fHx8&auto=format&fit=crop&w=386&q=80",
"https://images.unsplash.com/photo-1542909168-82c3e7fdca5c?ixlib=rb-1.2.1&ixid=MnwxMjA3fDB8MHxzZWFyY2h8MXx8aHVtYW4lMjBmYWNlfGVufDB8fDB8fA%3D%3D&w=1000&q=80",
]
return images
demo = gr.Interface(
fn=fake_gan,
inputs=None,
outputs=gr.Gallery(label="Generated Images", columns=[2]),
title="FD-GAN",
description="This is a fake demo of a GAN. In reality, the images are randomly chosen from Unsplash.",
)
if __name__ == "__main__":
demo.launch()
```
| gradio-app/gradio/blob/main/demo/fake_gan_no_input/run.ipynb |
Organization cards
You can create an organization card to help users learn more about what your organization is working on and how users can use your libraries, models, datasets, and Spaces.
An organization card is displayed on an organization's profile:
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/org-card.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/org-card-dark.png"/>
</div>
If you're a member of an organization, you'll see a button to create or edit your organization card on the organization's main page. Organization cards are a `README.md` static file inside a Space repo named `README`. The card can be as simple as Markdown text, or you can create a more customized appearance with HTML.
The card for the [Hugging Face Course organization](https://huggingface.co/huggingface-course), shown above, [contains the following HTML](https://huggingface.co/spaces/huggingface-course/README/blob/main/README.md):
```html
<p>
This is the organization grouping all the models and datasets used in the <a href="https://huggingface.co/course/chapter1" class="underline">Hugging Face course</a>.
</p>
```
For more examples, take a look at:
* [Amazon's](https://huggingface.co/spaces/amazon/README/blob/main/README.md) organization card source code
* [spaCy's](https://huggingface.co/spaces/spacy/README/blob/main/README.md) organization card source code.
| huggingface/hub-docs/blob/main/docs/hub/organizations-cards.md |
!---
Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
# Generating the documentation
> The following guide is adapted from [🤗 Transformers](https://github.com/huggingface/transformers/tree/main/docs).
To generate the documentation for 🤗 Optimum, simply run the following command
from the root of the `optimum` repository:
```bash
make doc BUILD_DIR=optimum-doc-build VERSION=main
```
This command will generate the HTML files that will be rendered as the
documentation on the [Hugging Face
website](https://huggingface.co/docs/optimum/index). You can inspect them in
your favorite browser. You can also adapt the `BUILD_DIR` and `VERSION`
arguments to any temporary folder or version that you prefer.
To generate the documentation for one of the hardware partner integrations, you
first need to clone the corresponding repository and run the `make doc` command
to build the docs. For example, the following commands generate the
documentation for `optimum-habana`:
```
git clone https://github.com/huggingface/optimum-habana.git
cd optimum-habana
make doc BUILD_DIR=habana-doc-build
```
---
**NOTE**
You only need to generate the documentation to inspect it locally, e.g. if you're
planning changes and want to check how they look like before committing. You
don't have to commit the built documentation.
---
# Writing documentation - specification
The 🤗 Optimum documentation follows the [Google
documentation](https://sphinxcontrib-napoleon.readthedocs.io/en/latest/example_google.html)
style for docstrings, although we can write them directly in Markdown.
## Adding a new element to the navigation bar
Under the hood, the documentation is generated by the
[`hf-doc-builder`](https://github.com/huggingface/doc-builder) library. Here we
summarize the main syntax needed to write the documentation -- consult
`hf-doc-builder` for more details.
Accepted files are Markdown (.md or .mdx).
Create a file with its extension and put it in the `docs/source` directory. You
can then link it to the table of contents by putting the filename _without the
extension_ in the
[`_toctree.yml`](https://github.com/huggingface/optimum/blob/main/docs/source/_toctree.yml)
file.
## Renaming section headers and moving sections
It helps to keep the old links working when renaming section header and/or
moving sections from one document to another. This is because the old links are
likely to be used in Issues, Forums and social media and it makes for a much
more superior user experience if users reading those months later could still
easily navigate to the originally intended information.
Therefore we simply keep a little map of moved sections at the end of the
document where the original section was. The key is to preserve the original
anchor.
So if you renamed a section from: "Section A" to "Section B", then you can add
at the end of the file:
```
Sections that were moved:
[ <a href="#section-b">Section A</a><a id="section-a"></a> ]
```
and of course if you moved it to another file, then:
```
Sections that were moved:
[ <a href="../new-file#section-b">Section A</a><a id="section-a"></a> ]
```
Use the relative style to link to the new file so that the versioned docs
continue to work.
For an example of a rich moved sections set please see the very end of [the
`Trainer`
doc](https://github.com/huggingface/transformers/blob/main/docs/source/main_classes/trainer.mdx)
in `transformers`.
## Adding a new tutorial
Adding a new tutorial or section is done in two steps:
- Add a new file under `docs/source`. This file should be in Markdown (.md)
format.
- Link that file in `docs/source/_toctree.yml` on the correct toc-tree.
Make sure to put your new file under the proper section. It's unlikely to go in
the first section (*Get Started*), so depending on the intended targets
(beginners, more advanced users or researchers) it should go in a later section.
## Writing source documentation
Values that should be put in `code` should either be surrounded by backticks:
\`like so\`. Note that argument names and objects like True, None or any strings
should usually be put in `code`.
When mentioning a class, function or method, it is recommended to use our syntax
for internal links so that our tool adds a link to its documentation with this
syntax: \[\`XXXClass\`\] or \[\`function\`\]. This requires the class or
function to be in the main package.
If you want to create a link to some internal class or function, you need to
provide its path. For instance: \[\`utils.ModelOutput\`\]. This will be
converted into a link with `utils.ModelOutput` in the description. To get rid of
the path and only keep the name of the object you are linking to in the
description, add a ~: \[\`~utils.ModelOutput\`\] will generate a link with
`ModelOutput` in the description.
The same works for methods so you can either use \[\`XXXClass.method\`\] or
\[~\`XXXClass.method\`\].
### Defining arguments in a method
Arguments should be defined with the `Args:` (or `Arguments:` or `Parameters:`)
prefix, followed by a line return and an indentation. The argument should be
followed by its type, with its shape if it is a tensor, a colon and its
description:
```
Args:
n_layers (`int`): The number of layers of the model.
```
If the description is too long to fit in one line, another indentation is
necessary before writing the description after the argument.
Here's an example showcasing everything so far:
```
Args:
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
Indices of input sequence tokens in the vocabulary.
Indices can be obtained using [`AlbertTokenizer`]. See [`~PreTrainedTokenizer.encode`] and
[`~PreTrainedTokenizer.__call__`] for details.
[What are input IDs?](../glossary#input-ids)
```
For optional arguments or arguments with defaults we follow the following
syntax: imagine we have a function with the following signature:
```
def my_function(x: str = None, a: float = 1):
```
then its documentation should look like this:
```
Args:
x (`str`, *optional*):
This argument controls ...
a (`float`, *optional*, defaults to 1):
This argument is used to ...
```
Note that we always omit the "defaults to \`None\`" when None is the default for
any argument. Also note that even if the first line describing your argument
type and its default gets long, you can't break it on several lines. You can
however write as many lines as you want in the indented description (see the
example above with `input_ids`).
### Writing a multi-line code block
Multi-line code blocks can be useful for displaying examples. They are done
between two lines of three backticks as usual in Markdown:
````
```
# first line of code
# second line
# etc
```
````
We follow the [doctest](https://docs.python.org/3/library/doctest.html) syntax
for the examples to automatically test the results stay consistent with the
library.
### Writing a return block
The return block should be introduced with the `Returns:` prefix, followed by a
line return and an indentation. The first line should be the type of the return,
followed by a line return. No need to indent further for the elements building
the return.
Here's an example for a single value return:
```
Returns:
`List[int]`: A list of integers in the range [0, 1] --- 1 for a special token, 0 for a sequence token.
```
Here's an example for tuple return, comprising several objects:
```
Returns:
`tuple(torch.FloatTensor)` comprising various elements depending on the configuration ([`BertConfig`]) and inputs:
- ** loss** (*optional*, returned when `masked_lm_labels` is provided) `torch.FloatTensor` of shape `(1,)` --
Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
- **prediction_scores** (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`) --
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
```
## Adding an image
Due to the rapidly growing repository, it is important to make sure that no
files that would significantly weigh down the repository are added. This
includes images, videos and other non-text files. We prefer to leverage a hf.co
hosted `dataset` like the ones hosted on
[`hf-internal-testing`](https://huggingface.co/hf-internal-testing) in which to
place these files and reference them by URL. We recommend putting them in the
following dataset:
[huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images).
If an external contribution, feel free to add the images to your PR and ask a
Hugging Face member to migrate your images to this dataset.
## Styling the docstring
We have an automatic script running with the `make style` comment that will make
sure that:
- the docstrings fully take advantage of the line width
- all code examples are formatted using black, like the code of the 🤗 Optimum
library
This script may have some weird failures if you made a syntax mistake or if you
uncover a bug. Therefore, it's recommended to commit your changes before running
`make style`, so you can revert the changes done by that script easily.
## Testing documentation examples
Good documentation often comes with an example of how a specific function or
class should be used. Each model class should contain at least one example
showcasing how to use this model class in inference. *E.g.* the class
[Wav2Vec2ForCTC](https://huggingface.co/docs/transformers/model_doc/wav2vec2#transformers.Wav2Vec2ForCTC)
includes an example of how to transcribe speech to text in the [docstring of its
forward
function](https://huggingface.co/docs/transformers/model_doc/wav2vec2#transformers.Wav2Vec2ForCTC.forward).
Reference: https://github.com/huggingface/transformers/blob/main/docs/README.md#writing-doctests
## Writing documentation examples
The syntax for Example docstrings can look as follows:
```
Example:
```python
>>> from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
>>> from datasets import load_dataset
>>> import torch
>>> dataset = load_dataset("hf-internal-testing/librispeech_asr_demo", "clean", split="validation")
>>> dataset = dataset.sort("id")
>>> sampling_rate = dataset.features["audio"].sampling_rate
>>> processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
>>> model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")
>>> # audio file is decoded on the fly
>>> inputs = processor(dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt")
>>> with torch.no_grad():
... logits = model(**inputs).logits
>>> predicted_ids = torch.argmax(logits, dim=-1)
>>> # transcribe speech
>>> transcription = processor.batch_decode(predicted_ids)
>>> transcription[0]
'MISTER QUILTER IS THE APOSTLE OF THE MIDDLE CLASSES AND WE ARE GLAD TO WELCOME HIS GOSPEL'
```
```
The docstring should give a minimal, clear example of how the respective model
is to be used in inference and also include the expected (ideally sensible)
output.
Often, readers will try out the example before even going through the function
or class definitions. Therefore it is of utmost importance that the example
works as expected.
# Adding documentation support for an 🤗 Optimum subpackage
🤗 Optimum is distributed as a [namespace
package](https://packaging.python.org/en/latest/guides/packaging-namespace-packages/),
where each hardware integration _subpackage_ such as `optimum-graphcore` or
`optimum-intel` is bundled together as a single package. For every pull request
or release of 🤗 Optimum, we use GitHub Actions to combine the documentation for
each subpackage with the base documentation of the `optimum` repository.
Including the documentation for a subpackage involves four main steps:
1. Adding a `docs/source` folder to your `optimum-*` repo with content and a
`_toctree.yml` file that follows the same specification as 🤗 Optimum (see
the _Writing documentation_ section above)
2. Creating a Dockerfile in `docs` that installs all necessary dependencies
3. Adding a `make doc` target to the Makefile of the subpackage that generates
the HTML files of the documentation
4. Updating the GitHub Actions in `build_pr_documentation.yml` and
`build_main_documentation.yml` to render the subpackage documentation on the
Hugging Face website
Let's walk through an example with `optimum-habana` to see how steps 2-4 work in
detail. The Docker file for this subpackage looks as follows:
```dockerfile
# Define base image for Habana
FROM vault.habana.ai/gaudi-docker/1.4.0/ubuntu20.04/habanalabs/pytorch-installer-1.10.2:1.4.0-442
# Need node to build doc HTML. Taken from https://stackoverflow.com/a/67491580
RUN apt-get update && apt-get install -y \
software-properties-common \
npm
RUN npm install npm@latest -g && \
npm install n -g && \
n latest
# Clone repo and install basic dependencies
RUN python3 -m pip install --no-cache-dir --upgrade pip
RUN git clone https://github.com/huggingface/optimum-habana.git
RUN python3 -m pip install --no-cache-dir ./optimum-habana[quality]
```
The main thing to note here is the need to install Node in the Docker image -
that's because we need Node to generate the HTML files with the `hf-doc-builder`
library. Once you have the Dockerfile, the next step is to define a `doc` target
in the Makefile:
```
SHELL := /bin/bash
CURRENT_DIR = $(shell pwd)
...
build_doc_docker_image:
docker build -t doc_maker ./docs
doc: build_doc_docker_image
@test -n "$(BUILD_DIR)" || (echo "BUILD_DIR is empty." ; exit 1)
@test -n "$(VERSION)" || (echo "VERSION is empty." ; exit 1)
docker run -v $(CURRENT_DIR):/doc_folder --workdir=/doc_folder doc_maker \
doc-builder build optimum.habana /optimum-habana/docs/source/ \
--build_dir $(BUILD_DIR) \
--version $(VERSION) \
--version_tag_suffix "" \
--html \
--clean
```
Once you've added the `doc` target to the Makefile, you can generate the
documentation by running the following command from the root of the subpackage
repository:
```
make doc BUILD_DIR=habana-doc-build VERSION=main
```
The final step is to include the subpackage in the GitHub Actions of the
`optimum` repo, e.g. add/edit these steps to `build_pr_documentation.yml` and
`build_main_documentation.yml`:
```
# Add this
- uses: actions/checkout@v2
with:
repository: 'huggingface/optimum-habana'
path: optimum-habana
# Add this
- name: Make Habana documentation
run: |
cd optimum-habana
make doc BUILD_DIR=habana-doc-build VERSION=pr_$PR_NUMBER # Make sure BUILD_DIR={subpackage_name}-doc-build
sudo mv habana-doc-build ../optimum
cd ..
# Tweak this to include your subpackage
- name: Combine subpackage documentation
run: |
cd optimum
sudo python docs/combine_docs.py --subpackages habana --version pr_$PR_NUMBER # Make sure the subpackage is listed here!
sudo mv optimum-doc-build ../
cd ..
```
---
**NOTE**
Since the `optimum` documentation depends on the documentation of each
subpackage, it is good practice to ensure the subpackage documentation will
always build successfully. To ensure this, add a GitHub Action to your
subpackage that tests the documentation builds with every pull request / push to
`main`. Check out the `optimum-habana` repo for an example.
---
| huggingface/optimum/blob/main/docs/README.md |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Using Diffusers with other modalities
Diffusers is in the process of expanding to modalities other than images.
Example type | Colab | Pipeline |
:-------------------------:|:-------------------------:|:-------------------------:|
[Molecule conformation](https://www.nature.com/subjects/molecular-conformation#:~:text=Definition,to%20changes%20in%20their%20environment.) generation | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/geodiff_molecule_conformation.ipynb) | ❌
More coming soon! | huggingface/diffusers/blob/main/docs/source/en/using-diffusers/other-modalities.md |
gradio_test
## 0.3.3
### Patch Changes
- Updated dependencies [[`828fb9e`](https://github.com/gradio-app/gradio/commit/828fb9e6ce15b6ea08318675a2361117596a1b5d), [`73268ee`](https://github.com/gradio-app/gradio/commit/73268ee2e39f23ebdd1e927cb49b8d79c4b9a144)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.3.2
### Patch Changes
- Updated dependencies [[`4d1cbbc`](https://github.com/gradio-app/gradio/commit/4d1cbbcf30833ef1de2d2d2710c7492a379a9a00)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.3.1
### Patch Changes
- Updated dependencies []:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.3.0
### Features
- [#6532](https://github.com/gradio-app/gradio/pull/6532) [`96290d304`](https://github.com/gradio-app/gradio/commit/96290d304a61064b52c10a54b2feeb09ca007542) - tweak deps. Thanks [@pngwn](https://github.com/pngwn)!
## 0.2.3
### Patch Changes
- Updated dependencies [[`9caddc17b`](https://github.com/gradio-app/gradio/commit/9caddc17b1dea8da1af8ba724c6a5eab04ce0ed8)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.2.2
### Patch Changes
- Updated dependencies [[`f816136a0`](https://github.com/gradio-app/gradio/commit/f816136a039fa6011be9c4fb14f573e4050a681a)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.2.1
### Patch Changes
- Updated dependencies [[`3cdeabc68`](https://github.com/gradio-app/gradio/commit/3cdeabc6843000310e1a9e1d17190ecbf3bbc780), [`fad92c29d`](https://github.com/gradio-app/gradio/commit/fad92c29dc1f5cd84341aae417c495b33e01245f)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.2.0
### Features
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - fix cc build. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Swap websockets for SSE. Thanks [@pngwn](https://github.com/pngwn)!
## 0.2.0-beta.8
### Features
- [#6136](https://github.com/gradio-app/gradio/pull/6136) [`667802a6c`](https://github.com/gradio-app/gradio/commit/667802a6cdbfb2ce454a3be5a78e0990b194548a) - JS Component Documentation. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#6069](https://github.com/gradio-app/gradio/pull/6069) [`bf127e124`](https://github.com/gradio-app/gradio/commit/bf127e1241a41401e144874ea468dff8474eb505) - Swap websockets for SSE. Thanks [@aliabid94](https://github.com/aliabid94)!
## 0.2.0-beta.7
### Features
- [#6079](https://github.com/gradio-app/gradio/pull/6079) [`3b2d9eaa3`](https://github.com/gradio-app/gradio/commit/3b2d9eaa3e84de3e4a0799e4585a94510d665f26) - fix cc build. Thanks [@pngwn](https://github.com/pngwn)!
| gradio-app/gradio/blob/main/js/preview/test/test/frontend/CHANGELOG.md |
--
title: "Deploy Livebook notebooks as apps to Hugging Face Spaces"
thumbnail: /blog/assets/120_elixir-bumblebee/thumbnail.png
authors:
- user: josevalim
guest: true
---
# Deploy Livebook notebooks as apps to Hugging Face Spaces
The [Elixir](https://elixir-lang.org/) community has been making great strides towards Machine Learning and Hugging Face is playing an important role on making it possible. To showcase what you can already achieve with Elixir and Machine Learning today, we use [Livebook](https://livebook.dev/) to build a Whisper-based chat app and then deploy it to Hugging Face Spaces. All under 15 minutes, check it out:
<iframe width="100%" style="aspect-ratio: 16 / 9;"src="https://www.youtube.com/embed/uyVRPEXOqzw" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>
In this chat app, users can communicate only by sending audio messages, which are then automatically converted to text by the Whisper Machine Learning model.
This app showcases a few interesting features from Livebook and the Machine Learning ecosystem in Elixir:
- integration with Hugging Face Models
- multiplayer Machine Learning apps
- concurrent Machine Learning model serving (bonus point: [you can also distribute model servings over a cluster just as easily](https://news.livebook.dev/distributed2-machine-learning-notebooks-with-elixir-and-livebook---launch-week-1---day-2-1aIlaw))
If you don't know Livebook yet, it is an open-source tool for writing interactive code notebooks in Elixir, and it's part of the [growing collection of Elixir tools](https://github.com/elixir-nx) for numerical computing, data science, and Machine Learning.
## Hugging Face and Elixir
The Elixir community leverages the Hugging Face platform and its open source projects throughout its machine learning landscape. Here are some examples.
The first positive impact Hugging Face had was in the [Bumblebee library](https://github.com/elixir-nx/bumblebee), which brought pre-trained neural network models from Hugging Face to the Elixir community and was inspired by [Hugging Face Transformers](https://huggingface.co/docs/transformers/index). Besides the inspiration, Bumblebee also uses the Hugging Face Hub to download parameters for its models.
Another example is the [tokenizers library](https://github.com/elixir-nx/tokenizers), which is an Elixir binding for [Hugging Face Tokenizers](https://github.com/huggingface/tokenizers).
And last but not least, [Livebook can run inside Hugging Face Spaces](https://huggingface.co/docs/hub/spaces-sdks-docker-livebook) with just a few clicks as one of their Space Docker templates. So, not only can you deploy Livebook apps to Hugging Face, but you can also use it to run Livebook for free to write and experiment with your own notebooks.
## Your turn
We hope this new integration between Livebook and Hugging Face empowers even more people to use Machine Learning and show their work to the world.
Go ahead and [install Livebook on Hugging Face Spaces](https://huggingface.co/docs/hub/spaces-sdks-docker-livebook), and [follow our video tutorial](https://www.youtube.com/watch?v=uyVRPEXOqzw) to build and deploy your first Livebook ML app to Hugging Face. | huggingface/blog/blob/main/livebook-app-deployment.md |
--
title: "Getting Started with Sentiment Analysis using Python"
thumbnail: /blog/assets/50_sentiment_python/thumbnail.png
authors:
- user: federicopascual
---
# Getting Started with Sentiment Analysis using Python
<script async defer src="https://unpkg.com/medium-zoom-element@0/dist/medium-zoom-element.min.js"></script>
Sentiment analysis is the automated process of tagging data according to their sentiment, such as positive, negative and neutral. Sentiment analysis allows companies to analyze data at scale, detect insights and automate processes.
In the past, sentiment analysis used to be limited to researchers, machine learning engineers or data scientists with experience in natural language processing. However, the AI community has built awesome tools to democratize access to machine learning in recent years. Nowadays, you can use sentiment analysis with a few lines of code and no machine learning experience at all! 🤯
In this guide, you'll learn everything to get started with sentiment analysis using Python, including:
1. [What is sentiment analysis?](#1-what-is-sentiment-analysis)
2. [How to use pre-trained sentiment analysis models with Python](#2-how-to-use-pre-trained-sentiment-analysis-models-with-python)
3. [How to build your own sentiment analysis model](#3-building-your-own-sentiment-analysis-model)
4. [How to analyze tweets with sentiment analysis](#4-analyzing-tweets-with-sentiment-analysis-and-python)
Let's get started! 🚀
## 1. What is Sentiment Analysis?
Sentiment analysis is a [natural language processing](https://en.wikipedia.org/wiki/Natural_language_processing) technique that identifies the polarity of a given text. There are different flavors of sentiment analysis, but one of the most widely used techniques labels data into positive, negative and neutral. For example, let's take a look at these tweets mentioning [@VerizonSupport](https://twitter.com/VerizonSupport):
- *"dear @verizonsupport your service is straight 💩 in dallas.. been with y’all over a decade and this is all time low for y’all. i’m talking no internet at all."* → Would be tagged as "Negative".
- *"@verizonsupport ive sent you a dm"* → would be tagged as "Neutral".
- *"thanks to michelle et al at @verizonsupport who helped push my no-show-phone problem along. order canceled successfully and ordered this for pickup today at the apple store in the mall."* → would be tagged as "Positive".
Sentiment analysis allows processing data at scale and in real-time. For example, do you want to analyze thousands of tweets, product reviews or support tickets? Instead of sorting through this data manually, you can use sentiment analysis to automatically understand how people are talking about a specific topic, get insights for data-driven decisions and automate business processes.
Sentiment analysis is used in a wide variety of applications, for example:
- Analyze social media mentions to understand how people are talking about your brand vs your competitors.
- Analyze feedback from surveys and product reviews to quickly get insights into what your customers like and dislike about your product.
- Analyze incoming support tickets in real-time to detect angry customers and act accordingly to prevent churn.
## 2. How to Use Pre-trained Sentiment Analysis Models with Python
Now that we have covered what sentiment analysis is, we are ready to play with some sentiment analysis models! 🎉
On the [Hugging Face Hub](https://huggingface.co/models), we are building the largest collection of models and datasets publicly available in order to democratize machine learning 🚀. In the Hub, you can find more than 27,000 models shared by the AI community with state-of-the-art performances on tasks such as sentiment analysis, object detection, text generation, speech recognition and more. The Hub is free to use and most models have a widget that allows to test them directly on your browser!
There are more than [215 sentiment analysis models](https://huggingface.co/models?pipeline_tag=text-classification&sort=downloads&search=sentiment) publicly available on the Hub and integrating them with Python just takes 5 lines of code:
```python
pip install -q transformers
from transformers import pipeline
sentiment_pipeline = pipeline("sentiment-analysis")
data = ["I love you", "I hate you"]
sentiment_pipeline(data)
```
This code snippet uses the [pipeline class](https://huggingface.co/docs/transformers/main_classes/pipelines) to make predictions from models available in the Hub. It uses the [default model for sentiment analysis](https://huggingface.co/distilbert-base-uncased-finetuned-sst-2-english?text=I+like+you.+I+love+you) to analyze the list of texts `data` and it outputs the following results:
```python
[{'label': 'POSITIVE', 'score': 0.9998},
{'label': 'NEGATIVE', 'score': 0.9991}]
```
You can use a specific sentiment analysis model that is better suited to your language or use case by providing the name of the model. For example, if you want a sentiment analysis model for tweets, you can specify the [model id](https://huggingface.co/finiteautomata/bertweet-base-sentiment-analysis):
```python
specific_model = pipeline(model="finiteautomata/bertweet-base-sentiment-analysis")
specific_model(data)
```
You can test these models with your own data using this [Colab notebook](https://colab.research.google.com/drive/1G4nvWf6NtytiEyiIkYxs03nno5ZupIJn?usp=sharing):
<!-- <div class="flex text-center items-center"> -->
<figure class="flex justify-center w-full">
<iframe width="560" height="315" src="https://www.youtube.com/embed/eN-mbWOKJ7Q" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>
</figure>
The following are some popular models for sentiment analysis models available on the Hub that we recommend checking out:
- [Twitter-roberta-base-sentiment](https://huggingface.co/cardiffnlp/twitter-roberta-base-sentiment) is a roBERTa model trained on ~58M tweets and fine-tuned for sentiment analysis. Fine-tuning is the process of taking a pre-trained large language model (e.g. roBERTa in this case) and then tweaking it with additional training data to make it perform a second similar task (e.g. sentiment analysis).
- [Bert-base-multilingual-uncased-sentiment](https://huggingface.co/nlptown/bert-base-multilingual-uncased-sentiment) is a model fine-tuned for sentiment analysis on product reviews in six languages: English, Dutch, German, French, Spanish and Italian.
- [Distilbert-base-uncased-emotion](https://huggingface.co/bhadresh-savani/distilbert-base-uncased-emotion?text=I+feel+a+bit+let+down) is a model fine-tuned for detecting emotions in texts, including sadness, joy, love, anger, fear and surprise.
Are you interested in doing sentiment analysis in languages such as Spanish, French, Italian or German? On the Hub, you will find many models fine-tuned for different use cases and ~28 languages. You can check out the complete list of sentiment analysis models [here](https://huggingface.co/models?pipeline_tag=text-classification&sort=downloads&search=sentiment) and filter at the left according to the language of your interest.
## 3. Building Your Own Sentiment Analysis Model
Using pre-trained models publicly available on the Hub is a great way to get started right away with sentiment analysis. These models use deep learning architectures such as transformers that achieve state-of-the-art performance on sentiment analysis and other machine learning tasks. However, you can fine-tune a model with your own data to further improve the sentiment analysis results and get an extra boost of accuracy in your particular use case.
In this section, we'll go over two approaches on how to fine-tune a model for sentiment analysis with your own data and criteria. The first approach uses the Trainer API from the [🤗Transformers](https://github.com/huggingface/transformers), an open source library with 50K stars and 1K+ contributors and requires a bit more coding and experience. The second approach is a bit easier and more straightforward, it uses [AutoNLP](https://huggingface.co/autonlp), a tool to automatically train, evaluate and deploy state-of-the-art NLP models without code or ML experience.
Let's dive in!
### a. Fine-tuning model with Python
In this tutorial, you'll use the IMDB dataset to fine-tune a DistilBERT model for sentiment analysis.
The [IMDB dataset](https://huggingface.co/datasets/imdb) contains 25,000 movie reviews labeled by sentiment for training a model and 25,000 movie reviews for testing it. [DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert) is a smaller, faster and cheaper version of [BERT](https://huggingface.co/docs/transformers/model_doc/bert). It has 40% smaller than BERT and runs 60% faster while preserving over 95% of BERT’s performance. You'll use the IMDB dataset to fine-tune a DistilBERT model that is able to classify whether a movie review is positive or negative. Once you train the model, you will use it to analyze new data! ⚡️
We have [created this notebook](https://colab.research.google.com/drive/1t-NJadXsPTDT6EWIR0PRzpn5o8oMHzp3?usp=sharing) so you can use it through this tutorial in Google Colab.
#### 1. Activate GPU and Install Dependencies
As a first step, let's set up Google Colab to use a GPU (instead of CPU) to train the model much faster. You can do this by going to the menu, clicking on 'Runtime' > 'Change runtime type', and selecting 'GPU' as the Hardware accelerator. Once you do this, you should check if GPU is available on our notebook by running the following code:
```python
import torch
torch.cuda.is_available()
```
Then, install the libraries you will be using in this tutorial:
```python
!pip install datasets transformers huggingface_hub
```
You should also install `git-lfs` to use git in our model repository:
```python
!apt-get install git-lfs
```
#### 2. Preprocess data
You need data to fine-tune DistilBERT for sentiment analysis. So, let's use [🤗Datasets](https://github.com/huggingface/datasets/) library to download and preprocess the IMDB dataset so you can then use this data for training your model:
```python
from datasets import load_dataset
imdb = load_dataset("imdb")
```
IMDB is a huge dataset, so let's create smaller datasets to enable faster training and testing:
```python
small_train_dataset = imdb["train"].shuffle(seed=42).select([i for i in list(range(3000))])
small_test_dataset = imdb["test"].shuffle(seed=42).select([i for i in list(range(300))])
```
To preprocess our data, you will use [DistilBERT tokenizer](https://huggingface.co/docs/transformers/v4.15.0/en/model_doc/distilbert#transformers.DistilBertTokenizer):
```python
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
```
Next, you will prepare the text inputs for the model for both splits of our dataset (training and test) by using the [map method](https://huggingface.co/docs/datasets/about_map_batch.html):
```python
def preprocess_function(examples):
return tokenizer(examples["text"], truncation=True)
tokenized_train = small_train_dataset.map(preprocess_function, batched=True)
tokenized_test = small_test_dataset.map(preprocess_function, batched=True)
```
To speed up training, let's use a data_collator to convert your training samples to PyTorch tensors and concatenate them with the correct amount of [padding](https://huggingface.co/docs/transformers/preprocessing#everything-you-always-wanted-to-know-about-padding-and-truncation):
```python
from transformers import DataCollatorWithPadding
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
```
#### 3. Training the model
Now that the preprocessing is done, you can go ahead and train your model 🚀
You will be throwing away the pretraining head of the DistilBERT model and replacing it with a classification head fine-tuned for sentiment analysis. This enables you to transfer the knowledge from DistilBERT to your custom model 🔥
For training, you will be using the [Trainer API](https://huggingface.co/docs/transformers/v4.15.0/en/main_classes/trainer#transformers.Trainer), which is optimized for fine-tuning [Transformers](https://github.com/huggingface/transformers)🤗 models such as DistilBERT, BERT and RoBERTa.
First, let's define DistilBERT as your base model:
```python
from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased", num_labels=2)
```
Then, let's define the metrics you will be using to evaluate how good is your fine-tuned model ([accuracy and f1 score](https://huggingface.co/metrics)):
```python
import numpy as np
from datasets import load_metric
def compute_metrics(eval_pred):
load_accuracy = load_metric("accuracy")
load_f1 = load_metric("f1")
logits, labels = eval_pred
predictions = np.argmax(logits, axis=-1)
accuracy = load_accuracy.compute(predictions=predictions, references=labels)["accuracy"]
f1 = load_f1.compute(predictions=predictions, references=labels)["f1"]
return {"accuracy": accuracy, "f1": f1}
```
Next, let's login to your [Hugging Face account](https://huggingface.co/join) so you can manage your model repositories. `notebook_login` will launch a widget in your notebook where you'll need to add your [Hugging Face token](https://huggingface.co/settings/token):
```python
from huggingface_hub import notebook_login
notebook_login()
```
You are almost there! Before training our model, you need to define the training arguments and define a Trainer with all the objects you constructed up to this point:
```python
from transformers import TrainingArguments, Trainer
repo_name = "finetuning-sentiment-model-3000-samples"
training_args = TrainingArguments(
output_dir=repo_name,
learning_rate=2e-5,
per_device_train_batch_size=16,
per_device_eval_batch_size=16,
num_train_epochs=2,
weight_decay=0.01,
save_strategy="epoch",
push_to_hub=True,
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=tokenized_train,
eval_dataset=tokenized_test,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
)
```
Now, it's time to fine-tune the model on the sentiment analysis dataset! 🙌 You just have to call the `train()` method of your Trainer:
```python
trainer.train()
```
And voila! You fine-tuned a DistilBERT model for sentiment analysis! 🎉
Training time depends on the hardware you use and the number of samples in the dataset. In our case, it took almost 10 minutes using a GPU and fine-tuning the model with 3,000 samples. The more samples you use for training your model, the more accurate it will be but training could be significantly slower.
Next, let's compute the evaluation metrics to see how good your model is:
```python
trainer.evaluate()
```
In our case, we got 88% accuracy and 89% f1 score. Quite good for a sentiment analysis model just trained with 3,000 samples!
#### 4. Analyzing new data with the model
Now that you have trained a model for sentiment analysis, let's use it to analyze new data and get 🤖 predictions! This unlocks the power of machine learning; using a model to automatically analyze data at scale, in real-time ⚡️
First, let's upload the model to the Hub:
```python
trainer.push_to_hub()
```
Now that you have pushed the model to the Hub, you can use it [pipeline class](https://huggingface.co/docs/transformers/main_classes/pipelines) to analyze two new movie reviews and see how your model predicts its sentiment with just two lines of code 🤯:
```python
from transformers import pipeline
sentiment_model = pipeline(model="federicopascual/finetuning-sentiment-model-3000-samples")
sentiment_model(["I love this move", "This movie sucks!"])
```
These are the predictions from our model:
```python
[{'label': 'LABEL_1', 'score': 0.9558},
{'label': 'LABEL_0', 'score': 0.9413}]
```
In the IMDB dataset, `Label 1` means positive and `Label 0` is negative. Quite good! 🔥
### b. Training a sentiment model with AutoNLP
[AutoNLP](https://huggingface.co/autonlp) is a tool to train state-of-the-art machine learning models without code. It provides a friendly and easy-to-use user interface, where you can train custom models by simply uploading your data. AutoNLP will automatically fine-tune various pre-trained models with your data, take care of the hyperparameter tuning and find the best model for your use case. All models trained with AutoNLP are deployed and ready for production.
Training a sentiment analysis model using AutoNLP is super easy and it just takes a few clicks 🤯. Let's give it a try!
As a first step, let's get some data! You'll use [Sentiment140](https://huggingface.co/datasets/sentiment140), a popular sentiment analysis dataset that consists of Twitter messages labeled with 3 sentiments: 0 (negative), 2 (neutral), and 4 (positive). The dataset is quite big; it contains 1,600,000 tweets. As you don't need this amount of data to get your feet wet with AutoNLP and train your first models, we have prepared a smaller version of the Sentiment140 dataset with 3,000 samples that you can download from [here](https://cdn-media.huggingface.co/marketing/content/sentiment%20analysis/sentiment-analysis-python/sentiment140-3000samples.csv). This is how the dataset looks like:
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="Sentiment 140 dataset" src="assets/50_sentiment_python/sentiment140-dataset.png"></medium-zoom>
<figcaption>Sentiment 140 dataset</figcaption>
</figure>
Next, let's create a [new project on AutoNLP](https://ui.autonlp.huggingface.co/new) to train 5 candidate models:
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="Creating a new project on AutoNLP" src="assets/50_sentiment_python/new-project.png"></medium-zoom>
<figcaption>Creating a new project on AutoNLP</figcaption>
</figure>
Then, upload the dataset and map the text column and target columns:
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="Adding a dataset to AutoNLP" src="assets/50_sentiment_python/add-dataset.png"></medium-zoom>
<figcaption>Adding a dataset to AutoNLP</figcaption>
</figure>
Once you add your dataset, go to the "Trainings" tab and accept the pricing to start training your models. AutoNLP pricing can be as low as $10 per model:
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="Adding a dataset to AutoNLP" src="assets/50_sentiment_python/trainings.png"></medium-zoom>
<figcaption>Adding a dataset to AutoNLP</figcaption>
</figure>
After a few minutes, AutoNLP has trained all models, showing the performance metrics for all of them:
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="Adding a dataset to AutoNLP" src="assets/50_sentiment_python/training-success.png"></medium-zoom>
<figcaption>Trained sentiment analysis models by AutoNLP</figcaption>
</figure>
The best model has 77.87% accuracy 🔥 Pretty good for a sentiment analysis model for tweets trained with just 3,000 samples!
All these models are automatically uploaded to the Hub and deployed for production. You can use any of these models to start analyzing new data right away by using the [pipeline class](https://huggingface.co/docs/transformers/main_classes/pipelines) as shown in previous sections of this post.
## 4. Analyzing Tweets with Sentiment Analysis and Python
In this last section, you'll take what you have learned so far in this post and put it into practice with a fun little project: analyzing tweets about NFTs with sentiment analysis!
First, you'll use [Tweepy](https://www.tweepy.org/), an easy-to-use Python library for getting tweets mentioning #NFTs using the [Twitter API](https://developer.twitter.com/en/docs/twitter-api). Then, you will use a sentiment analysis model from the 🤗Hub to analyze these tweets. Finally, you will create some visualizations to explore the results and find some interesting insights.
You can use [this notebook](https://colab.research.google.com/drive/182UbzmSeAFgOiow7WNMxvnz-yO-SJQ0W?usp=sharing) to follow this tutorial. Let’s jump into it!
### 1. Install dependencies
First, let's install all the libraries you will use in this tutorial:
```
!pip install -q transformers tweepy wordcloud matplotlib
```
### 2. Set up Twitter API credentials
Next, you will set up the credentials for interacting with the Twitter API. First, you'll need to sign up for a [developer account on Twitter](https://developer.twitter.com/en/docs/twitter-api/getting-started/getting-access-to-the-twitter-api). Then, you have to create a new project and connect an app to get an API key and token. You can follow this [step-by-step guide](https://developer.twitter.com/en/docs/tutorials/step-by-step-guide-to-making-your-first-request-to-the-twitter-api-v2) to get your credentials.
Once you have the API key and token, let's create a wrapper with Tweepy for interacting with the Twitter API:
```python
import tweepy
# Add Twitter API key and secret
consumer_key = "XXXXXX"
consumer_secret = "XXXXXX"
# Handling authentication with Twitter
auth = tweepy.AppAuthHandler(consumer_key, consumer_secret)
# Create a wrapper for the Twitter API
api = tweepy.API(auth, wait_on_rate_limit=True, wait_on_rate_limit_notify=True)
```
### 3. Search for tweets using Tweepy
At this point, you are ready to start using the Twitter API to collect tweets 🎉. You will use [Tweepy Cursor](https://docs.tweepy.org/en/v3.5.0/cursor_tutorial.html) to extract 1,000 tweets mentioning #NFTs:
```python
# Helper function for handling pagination in our search and handle rate limits
def limit_handled(cursor):
while True:
try:
yield cursor.next()
except tweepy.RateLimitError:
print('Reached rate limite. Sleeping for >15 minutes')
time.sleep(15 * 61)
except StopIteration:
break
# Define the term you will be using for searching tweets
query = '#NFTs'
query = query + ' -filter:retweets'
# Define how many tweets to get from the Twitter API
count = 1000
# Let's search for tweets using Tweepy
search = limit_handled(tweepy.Cursor(api.search,
q=query,
tweet_mode='extended',
lang='en',
result_type="recent").items(count))
```
### 4. Run sentiment analysis on the tweets
Now you can put our new skills to work and run sentiment analysis on your data! 🎉
You will use one of the models available on the Hub fine-tuned for [sentiment analysis of tweets](https://huggingface.co/finiteautomata/bertweet-base-sentiment-analysis). Like in other sections of this post, you will use the [pipeline class](https://huggingface.co/docs/transformers/main_classes/pipelines) to make the predictions with this model:
```python
from transformers import pipeline
# Set up the inference pipeline using a model from the 🤗 Hub
sentiment_analysis = pipeline(model="finiteautomata/bertweet-base-sentiment-analysis")
# Let's run the sentiment analysis on each tweet
tweets = []
for tweet in search:
try:
content = tweet.full_text
sentiment = sentiment_analysis(content)
tweets.append({'tweet': content, 'sentiment': sentiment[0]['label']})
except:
pass
```
### 5. Explore the results of sentiment analysis
How are people talking about NFTs on Twitter? Are they talking mostly positively or negatively? Let's explore the results of the sentiment analysis to find out!
First, let's load the results on a dataframe and see examples of tweets that were labeled for each sentiment:
```python
import pandas as pd
# Load the data in a dataframe
df = pd.DataFrame(tweets)
pd.set_option('display.max_colwidth', None)
# Show a tweet for each sentiment
display(df[df["sentiment"] == 'POS'].head(1))
display(df[df["sentiment"] == 'NEU'].head(1))
display(df[df["sentiment"] == 'NEG'].head(1))
```
Output:
```
Tweet: @NFTGalIery Warm, exquisite and elegant palette of charming beauty Its price is 2401 ETH. \nhttps://t.co/Ej3BfVOAqc\n#NFTs #NFTartists #art #Bitcoin #Crypto #OpenSeaNFT #Ethereum #BTC Sentiment: POS
Tweet: How much our followers made on #Crypto in December:\n#DAPPRadar airdrop — $200\nFree #VPAD tokens — $800\n#GasDAO airdrop — up to $1000\nStarSharks_SSS IDO — $3500\nCeloLaunch IDO — $3000\n12 Binance XMas #NFTs — $360 \nTOTAL PROFIT: $8500+\n\nJoin and earn with us https://t.co/fS30uj6SYx Sentiment: NEU
Tweet: Stupid guy #2\nhttps://t.co/8yKzYjCYIl\n\n#NFT #NFTs #nftcollector #rarible https://t.co/O4V19gMmVk Sentiment: NEG
```
Then, let's see how many tweets you got for each sentiment and visualize these results:
```python
# Let's count the number of tweets by sentiments
sentiment_counts = df.groupby(['sentiment']).size()
print(sentiment_counts)
# Let's visualize the sentiments
fig = plt.figure(figsize=(6,6), dpi=100)
ax = plt.subplot(111)
sentiment_counts.plot.pie(ax=ax, autopct='%1.1f%%', startangle=270, fontsize=12, label="")
```
Interestingly, most of the tweets about NFTs are positive (56.1%) and almost none are negative
(2.0%):
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="Sentiment analysis result of NFTs tweets" src="assets/50_sentiment_python/sentiment-result.png"></medium-zoom>
<figcaption>Sentiment analysis result of NFTs tweets</figcaption>
</figure>
Finally, let's see what words stand out for each sentiment by creating a word cloud:
```python
from wordcloud import WordCloud
from wordcloud import STOPWORDS
# Wordcloud with positive tweets
positive_tweets = df['tweet'][df["sentiment"] == 'POS']
stop_words = ["https", "co", "RT"] + list(STOPWORDS)
positive_wordcloud = WordCloud(max_font_size=50, max_words=100, background_color="white", stopwords = stop_words).generate(str(positive_tweets))
plt.figure()
plt.title("Positive Tweets - Wordcloud")
plt.imshow(positive_wordcloud, interpolation="bilinear")
plt.axis("off")
plt.show()
# Wordcloud with negative tweets
negative_tweets = df['tweet'][df["sentiment"] == 'NEG']
stop_words = ["https", "co", "RT"] + list(STOPWORDS)
negative_wordcloud = WordCloud(max_font_size=50, max_words=100, background_color="white", stopwords = stop_words).generate(str(negative_tweets))
plt.figure()
plt.title("Negative Tweets - Wordcloud")
plt.imshow(negative_wordcloud, interpolation="bilinear")
plt.axis("off")
plt.show()
```
Some of the words associated with positive tweets include Discord, Ethereum, Join, Mars4 and Shroom:
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="Word cloud for positive tweets" src="assets/50_sentiment_python/positive-tweets-wordcloud.png"></medium-zoom>
<figcaption>Word cloud for positive tweets</figcaption>
</figure>
In contrast, words associated with negative tweets include: cookies chaos, Solana, and OpenseaNFT:
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="Word cloud for negative tweets" src="assets/50_sentiment_python/negative-tweets-wordcloud.png"></medium-zoom>
<figcaption>Word cloud for negative tweets</figcaption>
</figure>
And that is it! With just a few lines of python code, you were able to collect tweets, analyze them with sentiment analysis and create some cool visualizations to analyze the results! Pretty cool, huh?
## 5. Wrapping up
Sentiment analysis with Python has never been easier! Tools such as [🤗Transformers](https://github.com/huggingface/transformers) and the [🤗Hub](https://huggingface.co/models) makes sentiment analysis accessible to all developers. You can use open source, pre-trained models for sentiment analysis in just a few lines of code 🔥
Do you want to train a custom model for sentiment analysis with your own data? Easy peasy! You can fine-tune a model using [Trainer API](https://huggingface.co/docs/transformers/v4.15.0/en/main_classes/trainer#transformers.Trainer) to build on top of large language models and get state-of-the-art results. If you want something even easier, you can use [AutoNLP](https://huggingface.co/autonlp) to train custom machine learning models by simply uploading data.
If you have questions, the Hugging Face community can help answer and/or benefit from, please ask them in the [Hugging Face forum](https://discuss.huggingface.co/). Also, join our [discord server](https://discord.gg/YRAq8fMnUG) to talk with us and with the Hugging Face community.
| huggingface/blog/blob/main/sentiment-analysis-python.md |
!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Create and share Model Cards
The `huggingface_hub` library provides a Python interface to create, share, and update Model Cards.
Visit [the dedicated documentation page](https://huggingface.co/docs/hub/models-cards)
for a deeper view of what Model Cards on the Hub are, and how they work under the hood.
<Tip>
[New (beta)! Try our experimental Model Card Creator App](https://huggingface.co/spaces/huggingface/Model_Cards_Writing_Tool)
</Tip>
## Load a Model Card from the Hub
To load an existing card from the Hub, you can use the [`ModelCard.load`] function. Here, we'll load the card from [`nateraw/vit-base-beans`](https://huggingface.co/nateraw/vit-base-beans).
```python
from huggingface_hub import ModelCard
card = ModelCard.load('nateraw/vit-base-beans')
```
This card has some helpful attributes that you may want to access/leverage:
- `card.data`: Returns a [`ModelCardData`] instance with the model card's metadata. Call `.to_dict()` on this instance to get the representation as a dictionary.
- `card.text`: Returns the text of the card, *excluding the metadata header*.
- `card.content`: Returns the text content of the card, *including the metadata header*.
## Create Model Cards
### From Text
To initialize a Model Card from text, just pass the text content of the card to the `ModelCard` on init.
```python
content = """
---
language: en
license: mit
---
# My Model Card
"""
card = ModelCard(content)
card.data.to_dict() == {'language': 'en', 'license': 'mit'} # True
```
Another way you might want to do this is with f-strings. In the following example, we:
- Use [`ModelCardData.to_yaml`] to convert metadata we defined to YAML so we can use it to insert the YAML block in the model card.
- Show how you might use a template variable via Python f-strings.
```python
card_data = ModelCardData(language='en', license='mit', library='timm')
example_template_var = 'nateraw'
content = f"""
---
{ card_data.to_yaml() }
---
# My Model Card
This model created by [@{example_template_var}](https://github.com/{example_template_var})
"""
card = ModelCard(content)
print(card)
```
The above example would leave us with a card that looks like this:
```
---
language: en
license: mit
library: timm
---
# My Model Card
This model created by [@nateraw](https://github.com/nateraw)
```
### From a Jinja Template
If you have `Jinja2` installed, you can create Model Cards from a jinja template file. Let's see a basic example:
```python
from pathlib import Path
from huggingface_hub import ModelCard, ModelCardData
# Define your jinja template
template_text = """
---
{{ card_data }}
---
# Model Card for MyCoolModel
This model does this and that.
This model was created by [@{{ author }}](https://hf.co/{{author}}).
""".strip()
# Write the template to a file
Path('custom_template.md').write_text(template_text)
# Define card metadata
card_data = ModelCardData(language='en', license='mit', library_name='keras')
# Create card from template, passing it any jinja template variables you want.
# In our case, we'll pass author
card = ModelCard.from_template(card_data, template_path='custom_template.md', author='nateraw')
card.save('my_model_card_1.md')
print(card)
```
The resulting card's markdown looks like this:
```
---
language: en
license: mit
library_name: keras
---
# Model Card for MyCoolModel
This model does this and that.
This model was created by [@nateraw](https://hf.co/nateraw).
```
If you update any card.data, it'll reflect in the card itself.
```
card.data.library_name = 'timm'
card.data.language = 'fr'
card.data.license = 'apache-2.0'
print(card)
```
Now, as you can see, the metadata header has been updated:
```
---
language: fr
license: apache-2.0
library_name: timm
---
# Model Card for MyCoolModel
This model does this and that.
This model was created by [@nateraw](https://hf.co/nateraw).
```
As you update the card data, you can validate the card is still valid against the Hub by calling [`ModelCard.validate`]. This ensures that the card passes any validation rules set up on the Hugging Face Hub.
### From the Default Template
Instead of using your own template, you can also use the [default template](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/templates/modelcard_template.md), which is a fully featured model card with tons of sections you may want to fill out. Under the hood, it uses [Jinja2](https://jinja.palletsprojects.com/en/3.1.x/) to fill out a template file.
<Tip>
Note that you will have to have Jinja2 installed to use `from_template`. You can do so with `pip install Jinja2`.
</Tip>
```python
card_data = ModelCardData(language='en', license='mit', library_name='keras')
card = ModelCard.from_template(
card_data,
model_id='my-cool-model',
model_description="this model does this and that",
developers="Nate Raw",
repo="https://github.com/huggingface/huggingface_hub",
)
card.save('my_model_card_2.md')
print(card)
```
## Share Model Cards
If you're authenticated with the Hugging Face Hub (either by using `huggingface-cli login` or [`login`]), you can push cards to the Hub by simply calling [`ModelCard.push_to_hub`]. Let's take a look at how to do that...
First, we'll create a new repo called 'hf-hub-modelcards-pr-test' under the authenticated user's namespace:
```python
from huggingface_hub import whoami, create_repo
user = whoami()['name']
repo_id = f'{user}/hf-hub-modelcards-pr-test'
url = create_repo(repo_id, exist_ok=True)
```
Then, we'll create a card from the default template (same as the one defined in the section above):
```python
card_data = ModelCardData(language='en', license='mit', library_name='keras')
card = ModelCard.from_template(
card_data,
model_id='my-cool-model',
model_description="this model does this and that",
developers="Nate Raw",
repo="https://github.com/huggingface/huggingface_hub",
)
```
Finally, we'll push that up to the hub
```python
card.push_to_hub(repo_id)
```
You can check out the resulting card [here](https://huggingface.co/nateraw/hf-hub-modelcards-pr-test/blob/main/README.md).
If you instead wanted to push a card as a pull request, you can just say `create_pr=True` when calling `push_to_hub`:
```python
card.push_to_hub(repo_id, create_pr=True)
```
A resulting PR created from this command can be seen [here](https://huggingface.co/nateraw/hf-hub-modelcards-pr-test/discussions/3).
## Update metadata
In this section we will see what metadata are in repo cards and how to update them.
`metadata` refers to a hash map (or key value) context that provides some high-level information about a model, dataset or Space. That information can include details such as the model's `pipeline type`, `model_id` or `model_description`. For more detail you can take a look to these guides: [Model Card](https://huggingface.co/docs/hub/model-cards#model-card-metadata), [Dataset Card](https://huggingface.co/docs/hub/datasets-cards#dataset-card-metadata) and [Spaces Settings](https://huggingface.co/docs/hub/spaces-settings#spaces-settings).
Now lets see some examples on how to update those metadata.
Let's start with a first example:
```python
>>> from huggingface_hub import metadata_update
>>> metadata_update("username/my-cool-model", {"pipeline_tag": "image-classification"})
```
With these two lines of code you will update the metadata to set a new `pipeline_tag`.
By default, you cannot update a key that is already existing on the card. If you want to do so, you must pass
`overwrite=True` explicitly:
```python
>>> from huggingface_hub import metadata_update
>>> metadata_update("username/my-cool-model", {"pipeline_tag": "text-generation"}, overwrite=True)
```
It often happen that you want to suggest some changes to a repository
on which you don't have write permission. You can do that by creating a PR on that repo which will allow the owners to
review and merge your suggestions.
```python
>>> from huggingface_hub import metadata_update
>>> metadata_update("someone/model", {"pipeline_tag": "text-classification"}, create_pr=True)
```
## Include Evaluation Results
To include evaluation results in the metadata `model-index`, you can pass an [`EvalResult`] or a list of `EvalResult` with your associated evaluation results. Under the hood it'll create the `model-index` when you call `card.data.to_dict()`. For more information on how this works, you can check out [this section of the Hub docs](https://huggingface.co/docs/hub/models-cards#evaluation-results).
<Tip>
Note that using this function requires you to include the `model_name` attribute in [`ModelCardData`].
</Tip>
```python
card_data = ModelCardData(
language='en',
license='mit',
model_name='my-cool-model',
eval_results = EvalResult(
task_type='image-classification',
dataset_type='beans',
dataset_name='Beans',
metric_type='accuracy',
metric_value=0.7
)
)
card = ModelCard.from_template(card_data)
print(card.data)
```
The resulting `card.data` should look like this:
```
language: en
license: mit
model-index:
- name: my-cool-model
results:
- task:
type: image-classification
dataset:
name: Beans
type: beans
metrics:
- type: accuracy
value: 0.7
```
If you have more than one evaluation result you'd like to share, just pass a list of `EvalResult`:
```python
card_data = ModelCardData(
language='en',
license='mit',
model_name='my-cool-model',
eval_results = [
EvalResult(
task_type='image-classification',
dataset_type='beans',
dataset_name='Beans',
metric_type='accuracy',
metric_value=0.7
),
EvalResult(
task_type='image-classification',
dataset_type='beans',
dataset_name='Beans',
metric_type='f1',
metric_value=0.65
)
]
)
card = ModelCard.from_template(card_data)
card.data
```
Which should leave you with the following `card.data`:
```
language: en
license: mit
model-index:
- name: my-cool-model
results:
- task:
type: image-classification
dataset:
name: Beans
type: beans
metrics:
- type: accuracy
value: 0.7
- type: f1
value: 0.65
``` | huggingface/huggingface_hub/blob/main/docs/source/en/guides/model-cards.md |
!--Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# GPT-J
## Overview
The GPT-J model was released in the [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax) repository by Ben Wang and Aran Komatsuzaki. It is a GPT-2-like
causal language model trained on [the Pile](https://pile.eleuther.ai/) dataset.
This model was contributed by [Stella Biderman](https://huggingface.co/stellaathena).
## Usage tips
- To load [GPT-J](https://huggingface.co/EleutherAI/gpt-j-6B) in float32 one would need at least 2x model size
RAM: 1x for initial weights and another 1x to load the checkpoint. So for GPT-J it would take at least 48GB
RAM to just load the model. To reduce the RAM usage there are a few options. The `torch_dtype` argument can be
used to initialize the model in half-precision on a CUDA device only. There is also a fp16 branch which stores the fp16 weights,
which could be used to further minimize the RAM usage:
```python
>>> from transformers import GPTJForCausalLM
>>> import torch
>>> device = "cuda"
>>> model = GPTJForCausalLM.from_pretrained(
... "EleutherAI/gpt-j-6B",
... revision="float16",
... torch_dtype=torch.float16,
... ).to(device)
```
- The model should fit on 16GB GPU for inference. For training/fine-tuning it would take much more GPU RAM. Adam
optimizer for example makes four copies of the model: model, gradients, average and squared average of the gradients.
So it would need at least 4x model size GPU memory, even with mixed precision as gradient updates are in fp32. This
is not including the activations and data batches, which would again require some more GPU RAM. So one should explore
solutions such as DeepSpeed, to train/fine-tune the model. Another option is to use the original codebase to
train/fine-tune the model on TPU and then convert the model to Transformers format for inference. Instructions for
that could be found [here](https://github.com/kingoflolz/mesh-transformer-jax/blob/master/howto_finetune.md)
- Although the embedding matrix has a size of 50400, only 50257 entries are used by the GPT-2 tokenizer. These extra
tokens are added for the sake of efficiency on TPUs. To avoid the mismatch between embedding matrix size and vocab
size, the tokenizer for [GPT-J](https://huggingface.co/EleutherAI/gpt-j-6B) contains 143 extra tokens
`<|extratoken_1|>... <|extratoken_143|>`, so the `vocab_size` of tokenizer also becomes 50400.
## Usage examples
The [`~generation.GenerationMixin.generate`] method can be used to generate text using GPT-J
model.
```python
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
>>> model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-j-6B")
>>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B")
>>> prompt = (
... "In a shocking finding, scientists discovered a herd of unicorns living in a remote, "
... "previously unexplored valley, in the Andes Mountains. Even more surprising to the "
... "researchers was the fact that the unicorns spoke perfect English."
... )
>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids
>>> gen_tokens = model.generate(
... input_ids,
... do_sample=True,
... temperature=0.9,
... max_length=100,
... )
>>> gen_text = tokenizer.batch_decode(gen_tokens)[0]
```
...or in float16 precision:
```python
>>> from transformers import GPTJForCausalLM, AutoTokenizer
>>> import torch
>>> device = "cuda"
>>> model = GPTJForCausalLM.from_pretrained("EleutherAI/gpt-j-6B", torch_dtype=torch.float16).to(device)
>>> tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-j-6B")
>>> prompt = (
... "In a shocking finding, scientists discovered a herd of unicorns living in a remote, "
... "previously unexplored valley, in the Andes Mountains. Even more surprising to the "
... "researchers was the fact that the unicorns spoke perfect English."
... )
>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
>>> gen_tokens = model.generate(
... input_ids,
... do_sample=True,
... temperature=0.9,
... max_length=100,
... )
>>> gen_text = tokenizer.batch_decode(gen_tokens)[0]
```
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with GPT-J. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
<PipelineTag pipeline="text-generation"/>
- Description of [GPT-J](https://huggingface.co/EleutherAI/gpt-j-6B).
- A blog on how to [Deploy GPT-J 6B for inference using Hugging Face Transformers and Amazon SageMaker](https://huggingface.co/blog/gptj-sagemaker).
- A blog on how to [Accelerate GPT-J inference with DeepSpeed-Inference on GPUs](https://www.philschmid.de/gptj-deepspeed-inference).
- A blog post introducing [GPT-J-6B: 6B JAX-Based Transformer](https://arankomatsuzaki.wordpress.com/2021/06/04/gpt-j/). 🌎
- A notebook for [GPT-J-6B Inference Demo](https://colab.research.google.com/github/kingoflolz/mesh-transformer-jax/blob/master/colab_demo.ipynb). 🌎
- Another notebook demonstrating [Inference with GPT-J-6B](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/GPT-J-6B/Inference_with_GPT_J_6B.ipynb).
- [Causal language modeling](https://huggingface.co/course/en/chapter7/6?fw=pt#training-a-causal-language-model-from-scratch) chapter of the 🤗 Hugging Face Course.
- [`GPTJForCausalLM`] is supported by this [causal language modeling example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling#gpt-2gpt-and-causal-language-modeling), [text generation example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-generation), and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb).
- [`TFGPTJForCausalLM`] is supported by this [causal language modeling example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/language-modeling#run_clmpy) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling-tf.ipynb).
- [`FlaxGPTJForCausalLM`] is supported by this [causal language modeling example script](https://github.com/huggingface/transformers/tree/main/examples/flax/language-modeling#causal-language-modeling) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/causal_language_modeling_flax.ipynb).
**Documentation resources**
- [Text classification task guide](../tasks/sequence_classification)
- [Question answering task guide](../tasks/question_answering)
- [Causal language modeling task guide](../tasks/language_modeling)
## GPTJConfig
[[autodoc]] GPTJConfig
- all
<frameworkcontent>
<pt>
## GPTJModel
[[autodoc]] GPTJModel
- forward
## GPTJForCausalLM
[[autodoc]] GPTJForCausalLM
- forward
## GPTJForSequenceClassification
[[autodoc]] GPTJForSequenceClassification
- forward
## GPTJForQuestionAnswering
[[autodoc]] GPTJForQuestionAnswering
- forward
</pt>
<tf>
## TFGPTJModel
[[autodoc]] TFGPTJModel
- call
## TFGPTJForCausalLM
[[autodoc]] TFGPTJForCausalLM
- call
## TFGPTJForSequenceClassification
[[autodoc]] TFGPTJForSequenceClassification
- call
## TFGPTJForQuestionAnswering
[[autodoc]] TFGPTJForQuestionAnswering
- call
</tf>
<jax>
## FlaxGPTJModel
[[autodoc]] FlaxGPTJModel
- __call__
## FlaxGPTJForCausalLM
[[autodoc]] FlaxGPTJForCausalLM
- __call__
</jax>
</frameworkcontent>
| huggingface/transformers/blob/main/docs/source/en/model_doc/gptj.md |
!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Model outputs
All models have outputs that are instances of subclasses of [`~utils.ModelOutput`]. Those are
data structures containing all the information returned by the model, but that can also be used as tuples or
dictionaries.
Let's see how this looks in an example:
```python
from transformers import BertTokenizer, BertForSequenceClassification
import torch
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
model = BertForSequenceClassification.from_pretrained("bert-base-uncased")
inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
labels = torch.tensor([1]).unsqueeze(0) # Batch size 1
outputs = model(**inputs, labels=labels)
```
The `outputs` object is a [`~modeling_outputs.SequenceClassifierOutput`], as we can see in the
documentation of that class below, it means it has an optional `loss`, a `logits`, an optional `hidden_states` and
an optional `attentions` attribute. Here we have the `loss` since we passed along `labels`, but we don't have
`hidden_states` and `attentions` because we didn't pass `output_hidden_states=True` or
`output_attentions=True`.
<Tip>
When passing `output_hidden_states=True` you may expect the `outputs.hidden_states[-1]` to match `outputs.last_hidden_states` exactly.
However, this is not always the case. Some models apply normalization or subsequent process to the last hidden state when it's returned.
</Tip>
You can access each attribute as you would usually do, and if that attribute has not been returned by the model, you
will get `None`. Here for instance `outputs.loss` is the loss computed by the model, and `outputs.attentions` is
`None`.
When considering our `outputs` object as tuple, it only considers the attributes that don't have `None` values.
Here for instance, it has two elements, `loss` then `logits`, so
```python
outputs[:2]
```
will return the tuple `(outputs.loss, outputs.logits)` for instance.
When considering our `outputs` object as dictionary, it only considers the attributes that don't have `None`
values. Here for instance, it has two keys that are `loss` and `logits`.
We document here the generic model outputs that are used by more than one model type. Specific output types are
documented on their corresponding model page.
## ModelOutput
[[autodoc]] utils.ModelOutput
- to_tuple
## BaseModelOutput
[[autodoc]] modeling_outputs.BaseModelOutput
## BaseModelOutputWithPooling
[[autodoc]] modeling_outputs.BaseModelOutputWithPooling
## BaseModelOutputWithCrossAttentions
[[autodoc]] modeling_outputs.BaseModelOutputWithCrossAttentions
## BaseModelOutputWithPoolingAndCrossAttentions
[[autodoc]] modeling_outputs.BaseModelOutputWithPoolingAndCrossAttentions
## BaseModelOutputWithPast
[[autodoc]] modeling_outputs.BaseModelOutputWithPast
## BaseModelOutputWithPastAndCrossAttentions
[[autodoc]] modeling_outputs.BaseModelOutputWithPastAndCrossAttentions
## Seq2SeqModelOutput
[[autodoc]] modeling_outputs.Seq2SeqModelOutput
## CausalLMOutput
[[autodoc]] modeling_outputs.CausalLMOutput
## CausalLMOutputWithCrossAttentions
[[autodoc]] modeling_outputs.CausalLMOutputWithCrossAttentions
## CausalLMOutputWithPast
[[autodoc]] modeling_outputs.CausalLMOutputWithPast
## MaskedLMOutput
[[autodoc]] modeling_outputs.MaskedLMOutput
## Seq2SeqLMOutput
[[autodoc]] modeling_outputs.Seq2SeqLMOutput
## NextSentencePredictorOutput
[[autodoc]] modeling_outputs.NextSentencePredictorOutput
## SequenceClassifierOutput
[[autodoc]] modeling_outputs.SequenceClassifierOutput
## Seq2SeqSequenceClassifierOutput
[[autodoc]] modeling_outputs.Seq2SeqSequenceClassifierOutput
## MultipleChoiceModelOutput
[[autodoc]] modeling_outputs.MultipleChoiceModelOutput
## TokenClassifierOutput
[[autodoc]] modeling_outputs.TokenClassifierOutput
## QuestionAnsweringModelOutput
[[autodoc]] modeling_outputs.QuestionAnsweringModelOutput
## Seq2SeqQuestionAnsweringModelOutput
[[autodoc]] modeling_outputs.Seq2SeqQuestionAnsweringModelOutput
## Seq2SeqSpectrogramOutput
[[autodoc]] modeling_outputs.Seq2SeqSpectrogramOutput
## SemanticSegmenterOutput
[[autodoc]] modeling_outputs.SemanticSegmenterOutput
## ImageClassifierOutput
[[autodoc]] modeling_outputs.ImageClassifierOutput
## ImageClassifierOutputWithNoAttention
[[autodoc]] modeling_outputs.ImageClassifierOutputWithNoAttention
## DepthEstimatorOutput
[[autodoc]] modeling_outputs.DepthEstimatorOutput
## Wav2Vec2BaseModelOutput
[[autodoc]] modeling_outputs.Wav2Vec2BaseModelOutput
## XVectorOutput
[[autodoc]] modeling_outputs.XVectorOutput
## Seq2SeqTSModelOutput
[[autodoc]] modeling_outputs.Seq2SeqTSModelOutput
## Seq2SeqTSPredictionOutput
[[autodoc]] modeling_outputs.Seq2SeqTSPredictionOutput
## SampleTSPredictionOutput
[[autodoc]] modeling_outputs.SampleTSPredictionOutput
## TFBaseModelOutput
[[autodoc]] modeling_tf_outputs.TFBaseModelOutput
## TFBaseModelOutputWithPooling
[[autodoc]] modeling_tf_outputs.TFBaseModelOutputWithPooling
## TFBaseModelOutputWithPoolingAndCrossAttentions
[[autodoc]] modeling_tf_outputs.TFBaseModelOutputWithPoolingAndCrossAttentions
## TFBaseModelOutputWithPast
[[autodoc]] modeling_tf_outputs.TFBaseModelOutputWithPast
## TFBaseModelOutputWithPastAndCrossAttentions
[[autodoc]] modeling_tf_outputs.TFBaseModelOutputWithPastAndCrossAttentions
## TFSeq2SeqModelOutput
[[autodoc]] modeling_tf_outputs.TFSeq2SeqModelOutput
## TFCausalLMOutput
[[autodoc]] modeling_tf_outputs.TFCausalLMOutput
## TFCausalLMOutputWithCrossAttentions
[[autodoc]] modeling_tf_outputs.TFCausalLMOutputWithCrossAttentions
## TFCausalLMOutputWithPast
[[autodoc]] modeling_tf_outputs.TFCausalLMOutputWithPast
## TFMaskedLMOutput
[[autodoc]] modeling_tf_outputs.TFMaskedLMOutput
## TFSeq2SeqLMOutput
[[autodoc]] modeling_tf_outputs.TFSeq2SeqLMOutput
## TFNextSentencePredictorOutput
[[autodoc]] modeling_tf_outputs.TFNextSentencePredictorOutput
## TFSequenceClassifierOutput
[[autodoc]] modeling_tf_outputs.TFSequenceClassifierOutput
## TFSeq2SeqSequenceClassifierOutput
[[autodoc]] modeling_tf_outputs.TFSeq2SeqSequenceClassifierOutput
## TFMultipleChoiceModelOutput
[[autodoc]] modeling_tf_outputs.TFMultipleChoiceModelOutput
## TFTokenClassifierOutput
[[autodoc]] modeling_tf_outputs.TFTokenClassifierOutput
## TFQuestionAnsweringModelOutput
[[autodoc]] modeling_tf_outputs.TFQuestionAnsweringModelOutput
## TFSeq2SeqQuestionAnsweringModelOutput
[[autodoc]] modeling_tf_outputs.TFSeq2SeqQuestionAnsweringModelOutput
## FlaxBaseModelOutput
[[autodoc]] modeling_flax_outputs.FlaxBaseModelOutput
## FlaxBaseModelOutputWithPast
[[autodoc]] modeling_flax_outputs.FlaxBaseModelOutputWithPast
## FlaxBaseModelOutputWithPooling
[[autodoc]] modeling_flax_outputs.FlaxBaseModelOutputWithPooling
## FlaxBaseModelOutputWithPastAndCrossAttentions
[[autodoc]] modeling_flax_outputs.FlaxBaseModelOutputWithPastAndCrossAttentions
## FlaxSeq2SeqModelOutput
[[autodoc]] modeling_flax_outputs.FlaxSeq2SeqModelOutput
## FlaxCausalLMOutputWithCrossAttentions
[[autodoc]] modeling_flax_outputs.FlaxCausalLMOutputWithCrossAttentions
## FlaxMaskedLMOutput
[[autodoc]] modeling_flax_outputs.FlaxMaskedLMOutput
## FlaxSeq2SeqLMOutput
[[autodoc]] modeling_flax_outputs.FlaxSeq2SeqLMOutput
## FlaxNextSentencePredictorOutput
[[autodoc]] modeling_flax_outputs.FlaxNextSentencePredictorOutput
## FlaxSequenceClassifierOutput
[[autodoc]] modeling_flax_outputs.FlaxSequenceClassifierOutput
## FlaxSeq2SeqSequenceClassifierOutput
[[autodoc]] modeling_flax_outputs.FlaxSeq2SeqSequenceClassifierOutput
## FlaxMultipleChoiceModelOutput
[[autodoc]] modeling_flax_outputs.FlaxMultipleChoiceModelOutput
## FlaxTokenClassifierOutput
[[autodoc]] modeling_flax_outputs.FlaxTokenClassifierOutput
## FlaxQuestionAnsweringModelOutput
[[autodoc]] modeling_flax_outputs.FlaxQuestionAnsweringModelOutput
## FlaxSeq2SeqQuestionAnsweringModelOutput
[[autodoc]] modeling_flax_outputs.FlaxSeq2SeqQuestionAnsweringModelOutput
| huggingface/transformers/blob/main/docs/source/en/main_classes/output.md |
!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# DeBERTa
## Overview
The DeBERTa model was proposed in [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen It is based on Google's
BERT model released in 2018 and Facebook's RoBERTa model released in 2019.
It builds on RoBERTa with disentangled attention and enhanced mask decoder training with half of the data used in
RoBERTa.
The abstract from the paper is the following:
*Recent progress in pre-trained neural language models has significantly improved the performance of many natural
language processing (NLP) tasks. In this paper we propose a new model architecture DeBERTa (Decoding-enhanced BERT with
disentangled attention) that improves the BERT and RoBERTa models using two novel techniques. The first is the
disentangled attention mechanism, where each word is represented using two vectors that encode its content and
position, respectively, and the attention weights among words are computed using disentangled matrices on their
contents and relative positions. Second, an enhanced mask decoder is used to replace the output softmax layer to
predict the masked tokens for model pretraining. We show that these two techniques significantly improve the efficiency
of model pretraining and performance of downstream tasks. Compared to RoBERTa-Large, a DeBERTa model trained on half of
the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9%
(90.2% vs. 91.1%), on SQuAD v2.0 by +2.3% (88.4% vs. 90.7%) and RACE by +3.6% (83.2% vs. 86.8%). The DeBERTa code and
pre-trained models will be made publicly available at https://github.com/microsoft/DeBERTa.*
This model was contributed by [DeBERTa](https://huggingface.co/DeBERTa). This model TF 2.0 implementation was
contributed by [kamalkraj](https://huggingface.co/kamalkraj) . The original code can be found [here](https://github.com/microsoft/DeBERTa).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with DeBERTa. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
<PipelineTag pipeline="text-classification"/>
- A blog post on how to [Accelerate Large Model Training using DeepSpeed](https://huggingface.co/blog/accelerate-deepspeed) with DeBERTa.
- A blog post on [Supercharged Customer Service with Machine Learning](https://huggingface.co/blog/supercharge-customer-service-with-machine-learning) with DeBERTa.
- [`DebertaForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification.ipynb).
- [`TFDebertaForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/text-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/text_classification-tf.ipynb).
- [Text classification task guide](../tasks/sequence_classification)
<PipelineTag pipeline="token-classification" />
- [`DebertaForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/token-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/token_classification.ipynb).
- [`TFDebertaForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/token-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/token_classification-tf.ipynb).
- [Token classification](https://huggingface.co/course/chapter7/2?fw=pt) chapter of the 🤗 Hugging Face Course.
- [Byte-Pair Encoding tokenization](https://huggingface.co/course/chapter6/5?fw=pt) chapter of the 🤗 Hugging Face Course.
- [Token classification task guide](../tasks/token_classification)
<PipelineTag pipeline="fill-mask"/>
- [`DebertaForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling#robertabertdistilbert-and-masked-language-modeling) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb).
- [`TFDebertaForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/language-modeling#run_mlmpy) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling-tf.ipynb).
- [Masked language modeling](https://huggingface.co/course/chapter7/3?fw=pt) chapter of the 🤗 Hugging Face Course.
- [Masked language modeling task guide](../tasks/masked_language_modeling)
<PipelineTag pipeline="question-answering"/>
- [`DebertaForQuestionAnswering`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering.ipynb).
- [`TFDebertaForQuestionAnswering`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/tensorflow/question-answering) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering-tf.ipynb).
- [Question answering](https://huggingface.co/course/chapter7/7?fw=pt) chapter of the 🤗 Hugging Face Course.
- [Question answering task guide](../tasks/question_answering)
## DebertaConfig
[[autodoc]] DebertaConfig
## DebertaTokenizer
[[autodoc]] DebertaTokenizer
- build_inputs_with_special_tokens
- get_special_tokens_mask
- create_token_type_ids_from_sequences
- save_vocabulary
## DebertaTokenizerFast
[[autodoc]] DebertaTokenizerFast
- build_inputs_with_special_tokens
- create_token_type_ids_from_sequences
<frameworkcontent>
<pt>
## DebertaModel
[[autodoc]] DebertaModel
- forward
## DebertaPreTrainedModel
[[autodoc]] DebertaPreTrainedModel
## DebertaForMaskedLM
[[autodoc]] DebertaForMaskedLM
- forward
## DebertaForSequenceClassification
[[autodoc]] DebertaForSequenceClassification
- forward
## DebertaForTokenClassification
[[autodoc]] DebertaForTokenClassification
- forward
## DebertaForQuestionAnswering
[[autodoc]] DebertaForQuestionAnswering
- forward
</pt>
<tf>
## TFDebertaModel
[[autodoc]] TFDebertaModel
- call
## TFDebertaPreTrainedModel
[[autodoc]] TFDebertaPreTrainedModel
- call
## TFDebertaForMaskedLM
[[autodoc]] TFDebertaForMaskedLM
- call
## TFDebertaForSequenceClassification
[[autodoc]] TFDebertaForSequenceClassification
- call
## TFDebertaForTokenClassification
[[autodoc]] TFDebertaForTokenClassification
- call
## TFDebertaForQuestionAnswering
[[autodoc]] TFDebertaForQuestionAnswering
- call
</tf>
</frameworkcontent>
| huggingface/transformers/blob/main/docs/source/en/model_doc/deberta.md |
@gradio/preview
## 0.6.0
### Features
- [#6738](https://github.com/gradio-app/gradio/pull/6738) [`f3c4d78`](https://github.com/gradio-app/gradio/commit/f3c4d78b710854b94d9a15db78178e504a02c680) - reload on css changes + fix css specificity. Thanks [@pngwn](https://github.com/pngwn)!
- [#6654](https://github.com/gradio-app/gradio/pull/6654) [`95827bb`](https://github.com/gradio-app/gradio/commit/95827bbe6e5e766c44d8e357cd513c3330534f75) - Update dependency @sveltejs/vite-plugin-svelte to v3. Thanks [@renovate](https://github.com/apps/renovate)!
## 0.5.0
### Features
- [#6537](https://github.com/gradio-app/gradio/pull/6537) [`6d3fecfa4`](https://github.com/gradio-app/gradio/commit/6d3fecfa42dde1c70a60c397434c88db77289be6) - chore(deps): update all non-major dependencies. Thanks [@renovate](https://github.com/apps/renovate)!
## 0.4.0
### Features
- [#6532](https://github.com/gradio-app/gradio/pull/6532) [`96290d304`](https://github.com/gradio-app/gradio/commit/96290d304a61064b52c10a54b2feeb09ca007542) - tweak deps. Thanks [@pngwn](https://github.com/pngwn)!
- [#6296](https://github.com/gradio-app/gradio/pull/6296) [`46f13f496`](https://github.com/gradio-app/gradio/commit/46f13f4968c8177e318c9d75f2eed1ed55c2c042) - chore(deps): update all non-major dependencies. Thanks [@renovate](https://github.com/apps/renovate)!
## 0.3.0
### Highlights
#### New `ImageEditor` component ([#6169](https://github.com/gradio-app/gradio/pull/6169) [`9caddc17b`](https://github.com/gradio-app/gradio/commit/9caddc17b1dea8da1af8ba724c6a5eab04ce0ed8))
A brand new component, completely separate from `Image` that provides simple editing capabilities.
- Set background images from file uploads, webcam, or just paste!
- Crop images with an improved cropping UI. App authors can event set specific crop size, or crop ratios (`1:1`, etc)
- Paint on top of any image (or no image) and erase any mistakes!
- The ImageEditor supports layers, confining draw and erase actions to that layer.
- More flexible access to data. The image component returns a composite image representing the final state of the canvas as well as providing the background and all layers as individual images.
- Fully customisable. All features can be enabled and disabled. Even the brush color swatches can be customised.
<video src="https://user-images.githubusercontent.com/12937446/284027169-31188926-fd16-4a1c-8718-998e7aae4695.mp4" autoplay muted></video>
```py
def fn(im):
im["composite"] # the full canvas
im["background"] # the background image
im["layers"] # a list of individual layers
im = gr.ImageEditor(
# decide which sources you'd like to accept
sources=["upload", "webcam", "clipboard"],
# set a cropsize constraint, can either be a ratio or a concrete [width, height]
crop_size="1:1",
# enable crop (or disable it)
transforms=["crop"],
# customise the brush
brush=Brush(
default_size="25", # or leave it as 'auto'
color_mode="fixed", # 'fixed' hides the user swatches and colorpicker, 'defaults' shows it
default_color="hotpink", # html names are supported
colors=[
"rgba(0, 150, 150, 1)", # rgb(a)
"#fff", # hex rgb
"hsl(360, 120, 120)" # in fact any valid colorstring
]
),
brush=Eraser(default_size="25")
)
```
Thanks [@pngwn](https://github.com/pngwn)!
## 0.2.2
### Features
- [#6467](https://github.com/gradio-app/gradio/pull/6467) [`739e3a5a0`](https://github.com/gradio-app/gradio/commit/739e3a5a09771a4a386cab0c6605156cf9fda7f6) - Fix dev mode. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
## 0.2.1
### Fixes
- [#6457](https://github.com/gradio-app/gradio/pull/6457) [`d00fcf89d`](https://github.com/gradio-app/gradio/commit/d00fcf89d1c3ecbc910e81bb1311479ec2b73e4e) - Gradio custom component dev mode now detects changes to Example.svelte file. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
## 0.2.0
### Features
- [#6261](https://github.com/gradio-app/gradio/pull/6261) [`8bbeca0e7`](https://github.com/gradio-app/gradio/commit/8bbeca0e772a5a2853d02a058b35abb2c15ffaf1) - Improve Embed and CDN handling and fix a couple of related bugs. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.1
### Fixes
- [#6191](https://github.com/gradio-app/gradio/pull/6191) [`b555bc09f`](https://github.com/gradio-app/gradio/commit/b555bc09ffe8e58b10da6227e2f11a0c084aa71d) - fix cdn build. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.0
### Features
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Adds the ability to build the frontend and backend of custom components in preparation for publishing to pypi using `gradio_component build`. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Image v4. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Publish all components to npm. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Custom components. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - fix cc build. Thanks [@pngwn](https://github.com/pngwn)!
- [#6171](https://github.com/gradio-app/gradio/pull/6171) [`28322422c`](https://github.com/gradio-app/gradio/commit/28322422cb9d8d3e471e439ad602959662e79312) - strip dangling svelte imports. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Strip vite import warning. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.0-beta.8
### Features
- [#6094](https://github.com/gradio-app/gradio/pull/6094) [`c476bd5a5`](https://github.com/gradio-app/gradio/commit/c476bd5a5b70836163b9c69bf4bfe068b17fbe13) - Image v4. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.0-beta.7
### Features
- [#6016](https://github.com/gradio-app/gradio/pull/6016) [`83e947676`](https://github.com/gradio-app/gradio/commit/83e947676d327ca2ab6ae2a2d710c78961c771a0) - Format js in v4 branch. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#6079](https://github.com/gradio-app/gradio/pull/6079) [`3b2d9eaa3`](https://github.com/gradio-app/gradio/commit/3b2d9eaa3e84de3e4a0799e4585a94510d665f26) - fix cc build. Thanks [@pngwn](https://github.com/pngwn)!
- [#6112](https://github.com/gradio-app/gradio/pull/6112) [`e402bf07a`](https://github.com/gradio-app/gradio/commit/e402bf07af637b0763291f6936583afc305f1e31) - fix build. Thanks [@pngwn](https://github.com/pngwn)!
### Fixes
- [#6046](https://github.com/gradio-app/gradio/pull/6046) [`dbb7de5e0`](https://github.com/gradio-app/gradio/commit/dbb7de5e02c53fee05889d696d764d212cb96c74) - fix tests. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.0-beta.6
### Features
- [#5960](https://github.com/gradio-app/gradio/pull/5960) [`319c30f3f`](https://github.com/gradio-app/gradio/commit/319c30f3fccf23bfe1da6c9b132a6a99d59652f7) - rererefactor frontend files. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`85ba6de13`](https://github.com/gradio-app/gradio/commit/85ba6de136a45b3e92c74e410bb27e3cbe7138d7) - Add host to dev mode for vite. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`85ba6de13`](https://github.com/gradio-app/gradio/commit/85ba6de136a45b3e92c74e410bb27e3cbe7138d7) - Use tags to identify custom component dirs and ignore uninstalled components. Thanks [@pngwn](https://github.com/pngwn)!
- [#5938](https://github.com/gradio-app/gradio/pull/5938) [`13ed8a485`](https://github.com/gradio-app/gradio/commit/13ed8a485d5e31d7d75af87fe8654b661edcca93) - V4: Use beta release versions for '@gradio' packages. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`85ba6de13`](https://github.com/gradio-app/gradio/commit/85ba6de136a45b3e92c74e410bb27e3cbe7138d7) - Adds the ability to build the frontend and backend of custom components in preparation for publishing to pypi using `gradio_component build`. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`85ba6de13`](https://github.com/gradio-app/gradio/commit/85ba6de136a45b3e92c74e410bb27e3cbe7138d7) - In dev/build use full path to python/gradio executables. Thanks [@pngwn](https://github.com/pngwn)!
- [#5962](https://github.com/gradio-app/gradio/pull/5962) [`d298e7695`](https://github.com/gradio-app/gradio/commit/d298e76952289f87213e243e813dbce3cf09a5b3) - Strip vite import warning. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
### Fixes
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`85ba6de13`](https://github.com/gradio-app/gradio/commit/85ba6de136a45b3e92c74e410bb27e3cbe7138d7) - Better logs in dev mode. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.0-beta.5
### Features
- [#5745](https://github.com/gradio-app/gradio/pull/5745) [`f2154eb7d`](https://github.com/gradio-app/gradio/commit/f2154eb7d871162bdf01e5f6bd903bed3a969f05) - Fix windows paths. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
## 0.1.0-beta.4
### Features
- [#5649](https://github.com/gradio-app/gradio/pull/5649) [`d56b355c1`](https://github.com/gradio-app/gradio/commit/d56b355c12ccdeeb8406a3520fecc15ae69d9141) - Fix front-end imports + other misc fixes. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
## 0.1.0-beta.3
### Features
- [#5648](https://github.com/gradio-app/gradio/pull/5648) [`c573e2339`](https://github.com/gradio-app/gradio/commit/c573e2339b86c85b378dc349de5e9223a3c3b04a) - Publish all components to npm. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
## 0.1.0-beta.2
### Features
- [#5630](https://github.com/gradio-app/gradio/pull/5630) [`0b4fd5b6d`](https://github.com/gradio-app/gradio/commit/0b4fd5b6db96fc95a155e5e935e17e1ab11d1161) - Fix esbuild. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.0-beta.1
### Features
- [#5624](https://github.com/gradio-app/gradio/pull/5624) [`14fc612d8`](https://github.com/gradio-app/gradio/commit/14fc612d84bf6b1408eccd3a40fab41f25477571) - Fix esbuild. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.0-beta.0
### Features
- [#5507](https://github.com/gradio-app/gradio/pull/5507) [`1385dc688`](https://github.com/gradio-app/gradio/commit/1385dc6881f2d8ae7a41106ec21d33e2ef04d6a9) - Custom components. Thanks [@pngwn](https://github.com/pngwn)! | gradio-app/gradio/blob/main/js/preview/CHANGELOG.md |
Gradio Demo: blocks_js_methods
```
!pip install -q gradio
```
```
import gradio as gr
blocks = gr.Blocks()
with blocks as demo:
subject = gr.Textbox(placeholder="subject")
verb = gr.Radio(["ate", "loved", "hated"])
object = gr.Textbox(placeholder="object")
with gr.Row():
btn = gr.Button("Create sentence.")
reverse_btn = gr.Button("Reverse sentence.")
foo_bar_btn = gr.Button("Append foo")
reverse_then_to_the_server_btn = gr.Button(
"Reverse sentence and send to server."
)
def sentence_maker(w1, w2, w3):
return f"{w1} {w2} {w3}"
output1 = gr.Textbox(label="output 1")
output2 = gr.Textbox(label="verb")
output3 = gr.Textbox(label="verb reversed")
output4 = gr.Textbox(label="front end process and then send to backend")
btn.click(sentence_maker, [subject, verb, object], output1)
reverse_btn.click(
None, [subject, verb, object], output2, js="(s, v, o) => o + ' ' + v + ' ' + s"
)
verb.change(lambda x: x, verb, output3, js="(x) => [...x].reverse().join('')")
foo_bar_btn.click(None, [], subject, js="(x) => x + ' foo'")
reverse_then_to_the_server_btn.click(
sentence_maker,
[subject, verb, object],
output4,
js="(s, v, o) => [s, v, o].map(x => [...x].reverse().join(''))",
)
if __name__ == "__main__":
demo.launch()
```
| gradio-app/gradio/blob/main/demo/blocks_js_methods/run.ipynb |
a href="https://colab.research.google.com/github/sanchit-gandhi/notebooks/blob/main/fine_tune_whisper.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/></a>
# Fine-Tune Whisper For Multilingual ASR with 🤗 Transformers
In this Colab, we present a step-by-step guide on how to fine-tune Whisper
for any multilingual ASR dataset using Hugging Face 🤗 Transformers. This is a
more "hands-on" version of the accompanying [blog post](https://huggingface.co/blog/fine-tune-whisper).
For a more in-depth explanation of Whisper, the Common Voice dataset and the theory behind fine-tuning, the reader is advised to refer to the blog post.
## Introduction
Whisper is a pre-trained model for automatic speech recognition (ASR)
published in [September 2022](https://openai.com/blog/whisper/) by the authors
Alec Radford et al. from OpenAI. Unlike many of its predecessors, such as
[Wav2Vec 2.0](https://arxiv.org/abs/2006.11477), which are pre-trained
on un-labelled audio data, Whisper is pre-trained on a vast quantity of
**labelled** audio-transcription data, 680,000 hours to be precise.
This is an order of magnitude more data than the un-labelled audio data used
to train Wav2Vec 2.0 (60,000 hours). What is more, 117,000 hours of this
pre-training data is multilingual ASR data. This results in checkpoints
that can be applied to over 96 languages, many of which are considered
_low-resource_.
When scaled to 680,000 hours of labelled pre-training data, Whisper models
demonstrate a strong ability to generalise to many datasets and domains.
The pre-trained checkpoints achieve competitive results to state-of-the-art
ASR systems, with near 3% word error rate (WER) on the test-clean subset of
LibriSpeech ASR and a new state-of-the-art on TED-LIUM with 4.7% WER (_c.f._
Table 8 of the [Whisper paper](https://cdn.openai.com/papers/whisper.pdf)).
The extensive multilingual ASR knowledge acquired by Whisper during pre-training
can be leveraged for other low-resource languages; through fine-tuning, the
pre-trained checkpoints can be adapted for specific datasets and languages
to further improve upon these results. We'll show just how Whisper can be fine-tuned
for low-resource languages in this Colab.
<figure>
<img src="https://raw.githubusercontent.com/sanchit-gandhi/notebooks/main/whisper_architecture.svg" alt="Trulli" style="width:100%">
<figcaption align = "center"><b>Figure 1:</b> Whisper model. The architecture
follows the standard Transformer-based encoder-decoder model. A
log-Mel spectrogram is input to the encoder. The last encoder
hidden states are input to the decoder via cross-attention mechanisms. The
decoder autoregressively predicts text tokens, jointly conditional on the
encoder hidden states and previously predicted tokens. Figure source:
<a href="https://openai.com/blog/whisper/">OpenAI Whisper Blog</a>.</figcaption>
</figure>
The Whisper checkpoints come in five configurations of varying model sizes.
The smallest four are trained on either English-only or multilingual data.
The largest checkpoint is multilingual only. All nine of the pre-trained checkpoints
are available on the [Hugging Face Hub](https://huggingface.co/models?search=openai/whisper). The
checkpoints are summarised in the following table with links to the models on the Hub:
| Size | Layers | Width | Heads | Parameters | English-only | Multilingual |
|--------|--------|-------|-------|------------|------------------------------------------------------|---------------------------------------------------|
| tiny | 4 | 384 | 6 | 39 M | [✓](https://huggingface.co/openai/whisper-tiny.en) | [✓](https://huggingface.co/openai/whisper-tiny.) |
| base | 6 | 512 | 8 | 74 M | [✓](https://huggingface.co/openai/whisper-base.en) | [✓](https://huggingface.co/openai/whisper-base) |
| small | 12 | 768 | 12 | 244 M | [✓](https://huggingface.co/openai/whisper-small.en) | [✓](https://huggingface.co/openai/whisper-small) |
| medium | 24 | 1024 | 16 | 769 M | [✓](https://huggingface.co/openai/whisper-medium.en) | [✓](https://huggingface.co/openai/whisper-medium) |
| large | 32 | 1280 | 20 | 1550 M | x | [✓](https://huggingface.co/openai/whisper-large) |
For demonstration purposes, we'll fine-tune the multilingual version of the
[`"small"`](https://huggingface.co/openai/whisper-small) checkpoint with 244M params (~= 1GB).
As for our data, we'll train and evaluate our system on a low-resource language
taken from the [Common Voice](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0)
dataset. We'll show that with as little as 8 hours of fine-tuning data, we can achieve
strong performance in this language.
------------------------------------------------------------------------
\\({}^1\\) The name Whisper follows from the acronym “WSPSR”, which stands for “Web-scale Supervised Pre-training for Speech Recognition”.
## Prepare Environment
First of all, let's try to secure a decent GPU for our Colab! Unfortunately, it's becoming much harder to get access to a good GPU with the free version of Google Colab. However, with Google Colab Pro one should have no issues in being allocated a V100 or P100 GPU.
To get a GPU, click _Runtime_ -> _Change runtime type_, then change _Hardware accelerator_ from _None_ to _GPU_.
We can verify that we've been assigned a GPU and view its specifications:
```python
gpu_info = !nvidia-smi
gpu_info = '\n'.join(gpu_info)
if gpu_info.find('failed') >= 0:
print('Not connected to a GPU')
else:
print(gpu_info)
```
Next, we need to update the Unix package `ffmpeg` to version 4:
```python
!add-apt-repository -y ppa:jonathonf/ffmpeg-4
!apt update
!apt install -y ffmpeg
```
We'll employ several popular Python packages to fine-tune the Whisper model.
We'll use `datasets` to download and prepare our training data and
`transformers` to load and train our Whisper model. We'll also require
the `soundfile` package to pre-process audio files, `evaluate` and `jiwer` to
assess the performance of our model. Finally, we'll
use `gradio` to build a flashy demo of our fine-tuned model.
```python
!pip install datasets>=2.6.1
!pip install git+https://github.com/huggingface/transformers
!pip install librosa
!pip install evaluate>=0.30
!pip install jiwer
!pip install gradio
```
We strongly advise you to upload model checkpoints directly the [Hugging Face Hub](https://huggingface.co/)
whilst training. The Hub provides:
- Integrated version control: you can be sure that no model checkpoint is lost during training.
- Tensorboard logs: track important metrics over the course of training.
- Model cards: document what a model does and its intended use cases.
- Community: an easy way to share and collaborate with the community!
Linking the notebook to the Hub is straightforward - it simply requires entering your
Hub authentication token when prompted. Find your Hub authentication token [here](https://huggingface.co/settings/tokens):
```python
from huggingface_hub import notebook_login
notebook_login()
```
## Load Dataset
Using 🤗 Datasets, downloading and preparing data is extremely simple.
We can download and prepare the Common Voice splits in just one line of code.
First, ensure you have accepted the terms of use on the Hugging Face Hub: [mozilla-foundation/common_voice_11_0](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0). Once you have accepted the terms, you will have full access to the dataset and be able to download the data locally.
Since Hindi is very low-resource, we'll combine the `train` and `validation`
splits to give approximately 8 hours of training data. We'll use the 4 hours
of `test` data as our held-out test set:
```python
from datasets import load_dataset, DatasetDict
common_voice = DatasetDict()
common_voice["train"] = load_dataset("mozilla-foundation/common_voice_11_0", "hi", split="train+validation", use_auth_token=True)
common_voice["test"] = load_dataset("mozilla-foundation/common_voice_11_0", "hi", split="test", use_auth_token=True)
print(common_voice)
```
Most ASR datasets only provide input audio samples (`audio`) and the
corresponding transcribed text (`sentence`). Common Voice contains additional
metadata information, such as `accent` and `locale`, which we can disregard for ASR.
Keeping the notebook as general as possible, we only consider the input audio and
transcribed text for fine-tuning, discarding the additional metadata information:
```python
common_voice = common_voice.remove_columns(["accent", "age", "client_id", "down_votes", "gender", "locale", "path", "segment", "up_votes"])
print(common_voice)
```
## Prepare Feature Extractor, Tokenizer and Data
The ASR pipeline can be de-composed into three stages:
1) A feature extractor which pre-processes the raw audio-inputs
2) The model which performs the sequence-to-sequence mapping
3) A tokenizer which post-processes the model outputs to text format
In 🤗 Transformers, the Whisper model has an associated feature extractor and tokenizer,
called [WhisperFeatureExtractor](https://huggingface.co/docs/transformers/main/model_doc/whisper#transformers.WhisperFeatureExtractor)
and [WhisperTokenizer](https://huggingface.co/docs/transformers/main/model_doc/whisper#transformers.WhisperTokenizer)
respectively.
We'll go through details for setting-up the feature extractor and tokenizer one-by-one!
### Load WhisperFeatureExtractor
The Whisper feature extractor performs two operations:
1. Pads / truncates the audio inputs to 30s: any audio inputs shorter than 30s are padded to 30s with silence (zeros), and those longer that 30s are truncated to 30s
2. Converts the audio inputs to _log-Mel spectrogram_ input features, a visual representation of the audio and the form of the input expected by the Whisper model
<figure>
<img src="https://raw.githubusercontent.com/sanchit-gandhi/notebooks/main/spectrogram.jpg" alt="Trulli" style="width:100%">
<figcaption align = "center"><b>Figure 2:</b> Conversion of sampled audio array to log-Mel spectrogram.
Left: sampled 1-dimensional audio signal. Right: corresponding log-Mel spectrogram. Figure source:
<a href="https://ai.googleblog.com/2019/04/specaugment-new-data-augmentation.html">Google SpecAugment Blog</a>.
</figcaption>
We'll load the feature extractor from the pre-trained checkpoint with the default values:
```python
from transformers import WhisperFeatureExtractor
feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-small")
```
### Load WhisperTokenizer
The Whisper model outputs a sequence of _token ids_. The tokenizer maps each of these token ids to their corresponding text string. For Hindi, we can load the pre-trained tokenizer and use it for fine-tuning without any further modifications. We simply have to
specify the target language and the task. These arguments inform the
tokenizer to prefix the language and task tokens to the start of encoded
label sequences:
```python
from transformers import WhisperTokenizer
tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-small", language="Hindi", task="transcribe")
```
### Combine To Create A WhisperProcessor
To simplify using the feature extractor and tokenizer, we can _wrap_
both into a single `WhisperProcessor` class. This processor object
inherits from the `WhisperFeatureExtractor` and `WhisperProcessor`,
and can be used on the audio inputs and model predictions as required.
In doing so, we only need to keep track of two objects during training:
the `processor` and the `model`:
```python
from transformers import WhisperProcessor
processor = WhisperProcessor.from_pretrained("openai/whisper-small", language="Hindi", task="transcribe")
```
### Prepare Data
Let's print the first example of the Common Voice dataset to see
what form the data is in:
```python
print(common_voice["train"][0])
```
Since
our input audio is sampled at 48kHz, we need to _downsample_ it to
16kHz prior to passing it to the Whisper feature extractor, 16kHz being the sampling rate expected by the Whisper model.
We'll set the audio inputs to the correct sampling rate using dataset's
[`cast_column`](https://huggingface.co/docs/datasets/package_reference/main_classes.html?highlight=cast_column#datasets.DatasetDict.cast_column)
method. This operation does not change the audio in-place,
but rather signals to `datasets` to resample audio samples _on the fly_ the
first time that they are loaded:
```python
from datasets import Audio
common_voice = common_voice.cast_column("audio", Audio(sampling_rate=16000))
```
Re-loading the first audio sample in the Common Voice dataset will resample
it to the desired sampling rate:
```python
print(common_voice["train"][0])
```
Now we can write a function to prepare our data ready for the model:
1. We load and resample the audio data by calling `batch["audio"]`. As explained above, 🤗 Datasets performs any necessary resampling operations on the fly.
2. We use the feature extractor to compute the log-Mel spectrogram input features from our 1-dimensional audio array.
3. We encode the transcriptions to label ids through the use of the tokenizer.
```python
def prepare_dataset(batch):
# load and resample audio data from 48 to 16kHz
audio = batch["audio"]
# compute log-Mel input features from input audio array
batch["input_features"] = feature_extractor(audio["array"], sampling_rate=audio["sampling_rate"]).input_features[0]
# encode target text to label ids
batch["labels"] = tokenizer(batch["sentence"]).input_ids
return batch
```
We can apply the data preparation function to all of our training examples using dataset's `.map` method:
```python
common_voice = common_voice.map(prepare_dataset, remove_columns=common_voice.column_names["train"], num_proc=4)
```
## Training and Evaluation
Now that we've prepared our data, we're ready to dive into the training pipeline.
The [🤗 Trainer](https://huggingface.co/transformers/master/main_classes/trainer.html?highlight=trainer)
will do much of the heavy lifting for us. All we have to do is:
- Define a data collator: the data collator takes our pre-processed data and prepares PyTorch tensors ready for the model.
- Evaluation metrics: during evaluation, we want to evaluate the model using the [word error rate (WER)](https://huggingface.co/metrics/wer) metric. We need to define a `compute_metrics` function that handles this computation.
- Load a pre-trained checkpoint: we need to load a pre-trained checkpoint and configure it correctly for training.
- Define the training configuration: this will be used by the 🤗 Trainer to define the training schedule.
Once we've fine-tuned the model, we will evaluate it on the test data to verify that we have correctly trained it
to transcribe speech in Hindi.
### Define a Data Collator
The data collator for a sequence-to-sequence speech model is unique in the sense that it
treats the `input_features` and `labels` independently: the `input_features` must be
handled by the feature extractor and the `labels` by the tokenizer.
The `input_features` are already padded to 30s and converted to a log-Mel spectrogram
of fixed dimension by action of the feature extractor, so all we have to do is convert the `input_features`
to batched PyTorch tensors. We do this using the feature extractor's `.pad` method with `return_tensors=pt`.
The `labels` on the other hand are un-padded. We first pad the sequences
to the maximum length in the batch using the tokenizer's `.pad` method. The padding tokens
are then replaced by `-100` so that these tokens are **not** taken into account when
computing the loss. We then cut the BOS token from the start of the label sequence as we
append it later during training.
We can leverage the `WhisperProcessor` we defined earlier to perform both the
feature extractor and the tokenizer operations:
```python
import torch
from dataclasses import dataclass
from typing import Any, Dict, List, Union
@dataclass
class DataCollatorSpeechSeq2SeqWithPadding:
processor: Any
def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
# split inputs and labels since they have to be of different lengths and need different padding methods
# first treat the audio inputs by simply returning torch tensors
input_features = [{"input_features": feature["input_features"]} for feature in features]
batch = self.processor.feature_extractor.pad(input_features, return_tensors="pt")
# get the tokenized label sequences
label_features = [{"input_ids": feature["labels"]} for feature in features]
# pad the labels to max length
labels_batch = self.processor.tokenizer.pad(label_features, return_tensors="pt")
# replace padding with -100 to ignore loss correctly
labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
# if bos token is appended in previous tokenization step,
# cut bos token here as it's append later anyways
if (labels[:, 0] == self.processor.tokenizer.bos_token_id).all().cpu().item():
labels = labels[:, 1:]
batch["labels"] = labels
return batch
```
Let's initialise the data collator we've just defined:
```python
data_collator = DataCollatorSpeechSeq2SeqWithPadding(processor=processor)
```
### Evaluation Metrics
We'll use the word error rate (WER) metric, the 'de-facto' metric for assessing
ASR systems. For more information, refer to the WER [docs](https://huggingface.co/metrics/wer). We'll load the WER metric from 🤗 Evaluate:
```python
import evaluate
metric = evaluate.load("wer")
```
We then simply have to define a function that takes our model
predictions and returns the WER metric. This function, called
`compute_metrics`, first replaces `-100` with the `pad_token_id`
in the `label_ids` (undoing the step we applied in the
data collator to ignore padded tokens correctly in the loss).
It then decodes the predicted and label ids to strings. Finally,
it computes the WER between the predictions and reference labels:
```python
def compute_metrics(pred):
pred_ids = pred.predictions
label_ids = pred.label_ids
# replace -100 with the pad_token_id
label_ids[label_ids == -100] = tokenizer.pad_token_id
# we do not want to group tokens when computing the metrics
pred_str = tokenizer.batch_decode(pred_ids, skip_special_tokens=True)
label_str = tokenizer.batch_decode(label_ids, skip_special_tokens=True)
wer = 100 * metric.compute(predictions=pred_str, references=label_str)
return {"wer": wer}
```
### Load a Pre-Trained Checkpoint
Now let's load the pre-trained Whisper `small` checkpoint. Again, this
is trivial through use of 🤗 Transformers!
```python
from transformers import WhisperForConditionalGeneration
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
```
Override generation arguments - no tokens are forced as decoder outputs (see [`forced_decoder_ids`](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.generation_utils.GenerationMixin.generate.forced_decoder_ids)), no tokens are suppressed during generation (see [`suppress_tokens`](https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.generation_utils.GenerationMixin.generate.suppress_tokens)):
```python
model.config.forced_decoder_ids = None
model.config.suppress_tokens = []
```
### Define the Training Configuration
In the final step, we define all the parameters related to training. For more detail on the training arguments, refer to the Seq2SeqTrainingArguments [docs](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.Seq2SeqTrainingArguments).
```python
from transformers import Seq2SeqTrainingArguments
training_args = Seq2SeqTrainingArguments(
output_dir="./whisper-small-hi", # change to a repo name of your choice
per_device_train_batch_size=16,
gradient_accumulation_steps=1, # increase by 2x for every 2x decrease in batch size
learning_rate=1e-5,
warmup_steps=500,
max_steps=4000,
gradient_checkpointing=True,
fp16=True,
group_by_length=True,
evaluation_strategy="steps",
per_device_eval_batch_size=8,
predict_with_generate=True,
generation_max_length=225,
save_steps=1000,
eval_steps=1000,
logging_steps=25,
report_to=["tensorboard"],
load_best_model_at_end=True,
metric_for_best_model="wer",
greater_is_better=False,
push_to_hub=True,
)
```
**Note**: if one does not want to upload the model checkpoints to the Hub,
set `push_to_hub=False`.
We can forward the training arguments to the 🤗 Trainer along with our model,
dataset, data collator and `compute_metrics` function:
```python
from transformers import Seq2SeqTrainer
trainer = Seq2SeqTrainer(
args=training_args,
model=model,
train_dataset=common_voice["train"],
eval_dataset=common_voice["test"],
data_collator=data_collator,
compute_metrics=compute_metrics,
tokenizer=processor.feature_extractor,
)
```
### Training
Training will take approximately 5-10 hours depending on your GPU or the one
allocated to this Google Colab. If using this Google Colab directly to
fine-tune a Whisper model, you should make sure that training isn't
interrupted due to inactivity. A simple workaround to prevent this is
to paste the following code into the console of this tab (_right mouse click_
-> _inspect_ -> _Console tab_ -> _insert code_).
```javascript
function ConnectButton(){
console.log("Connect pushed");
document.querySelector("#top-toolbar > colab-connect-button").shadowRoot.querySelector("#connect").click()
}
setInterval(ConnectButton, 60000);
```
The peak GPU memory for the given training configuration is approximately 15.8GB.
Depending on the GPU allocated to the Google Colab, it is possible that you will encounter a CUDA `"out-of-memory"` error when you launch training.
In this case, you can reduce the `per_device_train_batch_size` incrementally by factors of 2
and employ [`gradient_accumulation_steps`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.Seq2SeqTrainingArguments.gradient_accumulation_steps)
to compensate.
To launch training, simply execute:
```python
trainer.train()
```
Our best WER is 32.0% - not bad for 8h of training data! We can submit our checkpoint to the [`hf-speech-bench`](https://huggingface.co/spaces/huggingface/hf-speech-bench) on push by setting the appropriate key-word arguments (kwargs):
```python
kwargs = {
"dataset_tags": "mozilla-foundation/common_voice_11_0",
"dataset": "Common Voice 11.0", # a 'pretty' name for the training dataset
"language": "hi",
"model_name": "Whisper Small Hi - Sanchit Gandhi", # a 'pretty' name for our model
"finetuned_from": "openai/whisper-small",
"tasks": "automatic-speech-recognition",
"tags": "hf-asr-leaderboard",
}
```
The training results can now be uploaded to the Hub. To do so, execute the `push_to_hub` command:
```python
trainer.push_to_hub(**kwargs)
```
## Building a Demo
Now that we've fine-tuned our model we can build a demo to show
off its ASR capabilities! We'll make use of 🤗 Transformers
`pipeline`, which will take care of the entire ASR pipeline,
right from pre-processing the audio inputs to decoding the
model predictions.
Running the example below will generate a Gradio demo where we
can record speech through the microphone of our computer and input it to
our fine-tuned Whisper model to transcribe the corresponding text:
```python
from transformers import pipeline
import gradio as gr
pipe = pipeline(model="sanchit-gandhi/whisper-small-hi") # change to "your-username/the-name-you-picked"
def transcribe(audio):
text = pipe(audio)["text"]
return text
iface = gr.Interface(
fn=transcribe,
inputs=gr.Audio(source="microphone", type="filepath"),
outputs="text",
title="Whisper Small Hindi",
description="Realtime demo for Hindi speech recognition using a fine-tuned Whisper small model.",
)
iface.launch()
```
## Closing Remarks
In this blog, we covered a step-by-step guide on fine-tuning Whisper for multilingual ASR
using 🤗 Datasets, Transformers and the Hugging Face Hub. For more details on the Whisper model, the Common Voice dataset and the theory behind fine-tuning, refere to the accompanying [blog post](https://huggingface.co/blog/fine-tune-whisper). If you're interested in fine-tuning other
Transformers models, both for English and multilingual ASR, be sure to check out the
examples scripts at [examples/pytorch/speech-recognition](https://github.com/huggingface/transformers/tree/main/examples/pytorch/speech-recognition).
| huggingface/blog/blob/main/notebooks/111_fine_tune_whisper.ipynb |
Xception
**Xception** is a convolutional neural network architecture that relies solely on [depthwise separable convolution layers](https://paperswithcode.com/method/depthwise-separable-convolution).
The weights from this model were ported from [Tensorflow/Models](https://github.com/tensorflow/models).
## How do I use this model on an image?
To load a pretrained model:
```py
>>> import timm
>>> model = timm.create_model('xception', pretrained=True)
>>> model.eval()
```
To load and preprocess the image:
```py
>>> import urllib
>>> from PIL import Image
>>> from timm.data import resolve_data_config
>>> from timm.data.transforms_factory import create_transform
>>> config = resolve_data_config({}, model=model)
>>> transform = create_transform(**config)
>>> url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
>>> urllib.request.urlretrieve(url, filename)
>>> img = Image.open(filename).convert('RGB')
>>> tensor = transform(img).unsqueeze(0) # transform and add batch dimension
```
To get the model predictions:
```py
>>> import torch
>>> with torch.no_grad():
... out = model(tensor)
>>> probabilities = torch.nn.functional.softmax(out[0], dim=0)
>>> print(probabilities.shape)
>>> # prints: torch.Size([1000])
```
To get the top-5 predictions class names:
```py
>>> # Get imagenet class mappings
>>> url, filename = ("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt", "imagenet_classes.txt")
>>> urllib.request.urlretrieve(url, filename)
>>> with open("imagenet_classes.txt", "r") as f:
... categories = [s.strip() for s in f.readlines()]
>>> # Print top categories per image
>>> top5_prob, top5_catid = torch.topk(probabilities, 5)
>>> for i in range(top5_prob.size(0)):
... print(categories[top5_catid[i]], top5_prob[i].item())
>>> # prints class names and probabilities like:
>>> # [('Samoyed', 0.6425196528434753), ('Pomeranian', 0.04062102362513542), ('keeshond', 0.03186424449086189), ('white wolf', 0.01739676296710968), ('Eskimo dog', 0.011717947199940681)]
```
Replace the model name with the variant you want to use, e.g. `xception`. You can find the IDs in the model summaries at the top of this page.
To extract image features with this model, follow the [timm feature extraction examples](../feature_extraction), just change the name of the model you want to use.
## How do I finetune this model?
You can finetune any of the pre-trained models just by changing the classifier (the last layer).
```py
>>> model = timm.create_model('xception', pretrained=True, num_classes=NUM_FINETUNE_CLASSES)
```
To finetune on your own dataset, you have to write a training loop or adapt [timm's training
script](https://github.com/rwightman/pytorch-image-models/blob/master/train.py) to use your dataset.
## How do I train this model?
You can follow the [timm recipe scripts](../scripts) for training a new model afresh.
## Citation
```BibTeX
@article{DBLP:journals/corr/ZagoruykoK16,
@misc{chollet2017xception,
title={Xception: Deep Learning with Depthwise Separable Convolutions},
author={François Chollet},
year={2017},
eprint={1610.02357},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
<!--
Type: model-index
Collections:
- Name: Xception
Paper:
Title: 'Xception: Deep Learning with Depthwise Separable Convolutions'
URL: https://paperswithcode.com/paper/xception-deep-learning-with-depthwise
Models:
- Name: xception
In Collection: Xception
Metadata:
FLOPs: 10600506792
Parameters: 22860000
File Size: 91675053
Architecture:
- 1x1 Convolution
- Convolution
- Dense Connections
- Depthwise Separable Convolution
- Global Average Pooling
- Max Pooling
- ReLU
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: xception
Crop Pct: '0.897'
Image Size: '299'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/xception.py#L229
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-cadene/xception-43020ad28.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.05%
Top 5 Accuracy: 94.4%
- Name: xception41
In Collection: Xception
Metadata:
FLOPs: 11681983232
Parameters: 26970000
File Size: 108422028
Architecture:
- 1x1 Convolution
- Convolution
- Dense Connections
- Depthwise Separable Convolution
- Global Average Pooling
- Max Pooling
- ReLU
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: xception41
Crop Pct: '0.903'
Image Size: '299'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/xception_aligned.py#L181
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_xception_41-e6439c97.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 78.54%
Top 5 Accuracy: 94.28%
- Name: xception65
In Collection: Xception
Metadata:
FLOPs: 17585702144
Parameters: 39920000
File Size: 160536780
Architecture:
- 1x1 Convolution
- Convolution
- Dense Connections
- Depthwise Separable Convolution
- Global Average Pooling
- Max Pooling
- ReLU
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: xception65
Crop Pct: '0.903'
Image Size: '299'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/xception_aligned.py#L200
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_xception_65-c9ae96e8.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.55%
Top 5 Accuracy: 94.66%
- Name: xception71
In Collection: Xception
Metadata:
FLOPs: 22817346560
Parameters: 42340000
File Size: 170295556
Architecture:
- 1x1 Convolution
- Convolution
- Dense Connections
- Depthwise Separable Convolution
- Global Average Pooling
- Max Pooling
- ReLU
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: xception71
Crop Pct: '0.903'
Image Size: '299'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/xception_aligned.py#L219
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/tf_xception_71-8eec7df1.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.88%
Top 5 Accuracy: 94.93%
--> | huggingface/pytorch-image-models/blob/main/hfdocs/source/models/xception.mdx |
--
title: How to train a Language Model with Megatron-LM
thumbnail: /blog/assets/100_megatron_training/thumbnail.png
authors:
- user: loubnabnl
---
# How to train a Language Model with Megatron-LM
Training large language models in Pytorch requires more than a simple training loop. It is usually distributed across multiple devices, with many optimization techniques for a stable and efficient training. Hugging Face 🤗 [Accelerate](https://huggingface.co/docs/accelerate/index) library was created to support distributed training across GPUs and TPUs with very easy integration into the training loops. 🤗 [Transformers](https://huggingface.co/docs/transformers/index) also support distributed training through the [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.Trainer) API, which provides feature-complete training in PyTorch, without even needing to implement a training loop.
Another popular tool among researchers to pre-train large transformer models is [Megatron-LM](https://github.com/NVIDIA/Megatron-LM), a powerful framework developed by the Applied Deep Learning Research team at NVIDIA. Unlike `accelerate` and the `Trainer`, using Megatron-LM is not straightforward and can be a little overwhelming for beginners. But it is highly optimized for the training on GPUs and can give some speedups. In this blogpost, you will learn how to train a language model on NVIDIA GPUs in Megatron-LM, and use it with `transformers`.
We will try to break down the different steps for training a GPT2 model in this framework, this includes:
* Environment setup
* Data preprocessing
* Training
* Model conversion to 🤗 Transformers
## Why Megatron-LM?
Before getting into the training details, let’s first understand what makes this framework more efficient than others. This section is inspired by this great [blog](https://huggingface.co/blog/bloom-megatron-deepspeed) about BLOOM training with [Megatron-DeepSpeed](https://github.com/bigscience-workshop/Megatron-DeepSpeed), please refer to it for more details as this blog is intended to give a gentle introduction to Megatron-LM.
### DataLoader
Megatron-LM comes with an efficient DataLoader where the data is tokenized and shuffled before the training. It also splits the data into numbered sequences with indexes that are stored such that they need to be computed only once. To build the index, the number of epochs is computed based on the training parameters and an ordering is created and then shuffled. This is unlike most cases where we iterate through the entire dataset until it is exhausted and then repeat for the second epoch. This smoothes the learning curve and saves time during the training.
### Fused CUDA Kernels
When a computation is run on the GPU, the necessary data is fetched from memory, then the computation is run and the result is saved back into memory. In simple terms, the idea of fused kernels is that similar operations, usually performed separately by Pytorch, are combined into a single hardware operation. So they reduce the number of memory movements done in multiple discrete computations by merging them into one. The figure below illustrates the idea of Kernel Fusion. It is inspired from this [paper](https://www.arxiv-vanity.com/papers/1305.1183/), which discusses the concept in detail.
<p align="center">
<img src="assets/100_megatron_training/kernel_fusion.png" width="600" />
</p>
When f, g and h are fused in one kernel, the intermediary results x’ and y’ of f and g are stored in the GPU registers and immediately used by h. But without fusion, x’ and y’ would need to be copied to the memory and then loaded by h. Therefore, Kernel Fusion gives a significant speed up to the computations.
Megatron-LM also uses a Fused implementation of AdamW from [Apex](https://github.com/NVIDIA/apex) which is faster than the Pytorch implementation.
While one can customize the DataLoader like Megatron-LM and use Apex’s Fused optimizer with `transformers`, it is not a beginner friendly undertaking to build custom Fused CUDA Kernels.
Now that you are familiar with the framework and what makes it advantageous, let’s get into the training details!
## How to train with Megatron-LM ?
### Setup
The easiest way to setup the environment is to pull an NVIDIA PyTorch Container that comes with all the required installations from [NGC](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch). See [documentation](https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/index.html) for more details. If you don't want to use this container you will need to install the latest pytorch, cuda, nccl, and NVIDIA [APEX](https://github.com/NVIDIA/apex#quick-start) releases and the `nltk` library.
So after having installed Docker, you can run the container with the following command (`xx.xx` denotes your Docker version), and then clone [Megatron-LM repository](https://github.com/NVIDIA/Megatron-LM) inside it:
```bash
docker run --gpus all -it --rm nvcr.io/nvidia/pytorch:xx.xx-py3
git clone https://github.com/NVIDIA/Megatron-LM
```
You also need to add the vocabulary file `vocab.json` and merges table `merges.txt` of your tokenizer inside Megatron-LM folder of your container. These files can be found in the model’s repository with the weights, see this [repository](https://huggingface.co/gpt2/tree/main) for GPT2. You can also train your own tokenizer using `transformers`. You can checkout the [CodeParrot project](https://github.com/huggingface/transformers/tree/main/examples/research_projects/codeparrot) for a practical example.
Now if you want to copy this data from outside the container you can use the following commands:
```bash
sudo docker cp vocab.json CONTAINER_ID:/workspace/Megatron-LM
sudo docker cp merges.txt CONTAINER_ID:/workspace/Megatron-LM
```
### Data preprocessing
In the rest of this tutorial we will be using [CodeParrot](https://huggingface.co/codeparrot/codeparrot-small) model and data as an example.
The training data requires some preprocessing. First, you need to convert it into a loose json format, with one json containing a text sample per line. If you're using 🤗 [Datasets](https://huggingface.co/docs/datasets/index), here is an example on how to do that (always inside Megatron-LM folder):
```python
from datasets import load_dataset
train_data = load_dataset('codeparrot/codeparrot-clean-train', split='train')
train_data.to_json("codeparrot_data.json", lines=True)
```
The data is then tokenized, shuffled and processed into a binary format for training using the following command:
```bash
#if nltk isn't installed
pip install nltk
python tools/preprocess_data.py \
--input codeparrot_data.json \
--output-prefix codeparrot \
--vocab vocab.json \
--dataset-impl mmap \
--tokenizer-type GPT2BPETokenizer \
--merge-file merges.txt \
--json-keys content \
--workers 32 \
--chunk-size 25 \
--append-eod
```
The `workers` and `chunk_size` options refer to the number of workers used in the preprocessing and the chunk size of data assigned to each one. `dataset-impl` refers to the implementation mode of the indexed datasets from ['lazy', 'cached', 'mmap'].
This outputs two files `codeparrot_content_document.idx` and `codeparrot_content_document.bin` which are used in the training.
### Training
You can configure the model architecture and training parameters as shown below, or put it in a bash script that you will run. This command runs the pretraining on 8 GPUs for a 110M parameter CodeParrot model. Note that the data is partitioned by default into a 969:30:1 ratio for training/validation/test sets.
```bash
GPUS_PER_NODE=8
MASTER_ADDR=localhost
MASTER_PORT=6001
NNODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
DISTRIBUTED_ARGS="--nproc_per_node $GPUS_PER_NODE --nnodes $NNODES --node_rank $NODE_RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
CHECKPOINT_PATH=/workspace/Megatron-LM/experiments/codeparrot-small
VOCAB_FILE=vocab.json
MERGE_FILE=merges.txt
DATA_PATH=codeparrot_content_document
GPT_ARGS="--num-layers 12
--hidden-size 768
--num-attention-heads 12
--seq-length 1024
--max-position-embeddings 1024
--micro-batch-size 12
--global-batch-size 192
--lr 0.0005
--train-iters 150000
--lr-decay-iters 150000
--lr-decay-style cosine
--lr-warmup-iters 2000
--weight-decay .1
--adam-beta2 .999
--fp16
--log-interval 10
--save-interval 2000
--eval-interval 200
--eval-iters 10
"
TENSORBOARD_ARGS="--tensorboard-dir experiments/tensorboard"
python3 -m torch.distributed.launch $DISTRIBUTED_ARGS \
pretrain_gpt.py \
--tensor-model-parallel-size 1 \
--pipeline-model-parallel-size 1 \
$GPT_ARGS \
--vocab-file $VOCAB_FILE \
--merge-file $MERGE_FILE \
--save $CHECKPOINT_PATH \
--load $CHECKPOINT_PATH \
--data-path $DATA_PATH \
$TENSORBOARD_ARGS
```
With this setting, the training takes roughly 12 hours.
This setup uses Data Parallelism, but it is also possible to use Model Parallelism for very large models that don't fit in one GPU. The first option consists of Tensor Parallelism that splits the execution of a single transformer module over multiple GPUs, you will need to change `tensor-model-parallel-size` parameter to the desired number of GPUs. The second option is Pipeline Parallelism where the transformer modules are split into equally sized stages. The parameter `pipeline-model-parallel-size` determines the number of stages to split the model into. For more details please refer to this [blog](https://huggingface.co/blog/bloom-megatron-deepspeed)
### Converting the model to 🤗 Transformers
After training we want to use the model in `transformers` e.g. for evaluation or to deploy it to production. You can convert it to a `transformers` model following this [tutorial](https://huggingface.co/nvidia/megatron-gpt2-345m). For instance, after the training is finished you can copy the weights of the last iteration 150k and convert the `model_optim_rng.pt` file to a `pytorch_model.bin` file that is supported by `transformers` with the following commands:
```bash
# to execute outside the container:
mkdir -p nvidia/megatron-codeparrot-small
# copy the weights from the container
sudo docker cp CONTAINER_ID:/workspace/Megatron-LM/experiments/codeparrot-small/iter_0150000/mp_rank_00/model_optim_rng.pt nvidia/megatron-codeparrot-small
git clone https://github.com/huggingface/transformers.git
git clone https://github.com/NVIDIA/Megatron-LM.git
export PYTHONPATH=Megatron-LM
python transformers/src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py nvidia/megatron-codeparrot-small/model_optim_rng.pt
```
Be careful, you will need to replace the generated vocabulary file and merges table after the conversion, with the original ones we introduced earlier if you plan to load the tokenizer from there.
Don't forget to push your model to the hub and share it with the community, it only takes three lines of code 🤗:
```python
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("nvidia/megatron-codeparrot-small")
# this creates a repository under your username with the model name codeparrot-small
model.push_to_hub("codeparrot-small")
```
You can also easily use it to generate text:
```python
from transformers import pipeline
pipe = pipeline("text-generation", model="your_username/codeparrot-small")
outputs = pipe("def hello_world():")
print(outputs[0]["generated_text"])
```
```
def hello_world():
print("Hello World!")
```
Tranfsormers also handle big model inference efficiently. In case you trained a very large model (e.g. using Model Parallelism), you can easily use it for inference with the following command:
```python
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("your_username/codeparrot-large", device_map="auto")
```
This will use [accelerate](https://huggingface.co/docs/accelerate/index) library behind the scenes to automatically dispatch the model weights across the devices you have available (GPUs, CPU RAM).
Disclaimer: We have shown that anyone can use Megatron-LM to train language models. The question is when to use it. This framework obviously adds some time overhead because of the extra preprocessing and conversion steps. So it is important that you decide which framework is more appropriate for your case and model size. We recommend trying it for pre-training models or extended fine-tuning, but probably not for shorter fine-tuning of medium-sized models. The `Trainer` API and `accelerate` library are also very handy for model training, they are device-agnostic and give significant flexibility to the users.
Congratulations 🎉 now you know how to train a GPT2 model in Megatron-LM and make it supported by `transformers`!
| huggingface/blog/blob/main/megatron-training.md |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Load adapters
[[open-in-colab]]
There are several [training](../training/overview) techniques for personalizing diffusion models to generate images of a specific subject or images in certain styles. Each of these training methods produces a different type of adapter. Some of the adapters generate an entirely new model, while other adapters only modify a smaller set of embeddings or weights. This means the loading process for each adapter is also different.
This guide will show you how to load DreamBooth, textual inversion, and LoRA weights.
<Tip>
Feel free to browse the [Stable Diffusion Conceptualizer](https://huggingface.co/spaces/sd-concepts-library/stable-diffusion-conceptualizer), [LoRA the Explorer](https://huggingface.co/spaces/multimodalart/LoraTheExplorer), and the [Diffusers Models Gallery](https://huggingface.co/spaces/huggingface-projects/diffusers-gallery) for checkpoints and embeddings to use.
</Tip>
## DreamBooth
[DreamBooth](https://dreambooth.github.io/) finetunes an *entire diffusion model* on just several images of a subject to generate images of that subject in new styles and settings. This method works by using a special word in the prompt that the model learns to associate with the subject image. Of all the training methods, DreamBooth produces the largest file size (usually a few GBs) because it is a full checkpoint model.
Let's load the [herge_style](https://huggingface.co/sd-dreambooth-library/herge-style) checkpoint, which is trained on just 10 images drawn by Hergé, to generate images in that style. For it to work, you need to include the special word `herge_style` in your prompt to trigger the checkpoint:
```py
from diffusers import AutoPipelineForText2Image
import torch
pipeline = AutoPipelineForText2Image.from_pretrained("sd-dreambooth-library/herge-style", torch_dtype=torch.float16).to("cuda")
prompt = "A cute herge_style brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration"
image = pipeline(prompt).images[0]
image
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_dreambooth.png" />
</div>
## Textual inversion
[Textual inversion](https://textual-inversion.github.io/) is very similar to DreamBooth and it can also personalize a diffusion model to generate certain concepts (styles, objects) from just a few images. This method works by training and finding new embeddings that represent the images you provide with a special word in the prompt. As a result, the diffusion model weights stay the same and the training process produces a relatively tiny (a few KBs) file.
Because textual inversion creates embeddings, it cannot be used on its own like DreamBooth and requires another model.
```py
from diffusers import AutoPipelineForText2Image
import torch
pipeline = AutoPipelineForText2Image.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
```
Now you can load the textual inversion embeddings with the [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] method and generate some images. Let's load the [sd-concepts-library/gta5-artwork](https://huggingface.co/sd-concepts-library/gta5-artwork) embeddings and you'll need to include the special word `<gta5-artwork>` in your prompt to trigger it:
```py
pipeline.load_textual_inversion("sd-concepts-library/gta5-artwork")
prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration, <gta5-artwork> style"
image = pipeline(prompt).images[0]
image
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_txt_embed.png" />
</div>
Textual inversion can also be trained on undesirable things to create *negative embeddings* to discourage a model from generating images with those undesirable things like blurry images or extra fingers on a hand. This can be an easy way to quickly improve your prompt. You'll also load the embeddings with [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`], but this time, you'll need two more parameters:
- `weight_name`: specifies the weight file to load if the file was saved in the 🤗 Diffusers format with a specific name or if the file is stored in the A1111 format
- `token`: specifies the special word to use in the prompt to trigger the embeddings
Let's load the [sayakpaul/EasyNegative-test](https://huggingface.co/sayakpaul/EasyNegative-test) embeddings:
```py
pipeline.load_textual_inversion(
"sayakpaul/EasyNegative-test", weight_name="EasyNegative.safetensors", token="EasyNegative"
)
```
Now you can use the `token` to generate an image with the negative embeddings:
```py
prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration, EasyNegative"
negative_prompt = "EasyNegative"
image = pipeline(prompt, negative_prompt=negative_prompt, num_inference_steps=50).images[0]
image
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_neg_embed.png" />
</div>
## LoRA
[Low-Rank Adaptation (LoRA)](https://huggingface.co/papers/2106.09685) is a popular training technique because it is fast and generates smaller file sizes (a couple hundred MBs). Like the other methods in this guide, LoRA can train a model to learn new styles from just a few images. It works by inserting new weights into the diffusion model and then only the new weights are trained instead of the entire model. This makes LoRAs faster to train and easier to store.
<Tip>
LoRA is a very general training technique that can be used with other training methods. For example, it is common to train a model with DreamBooth and LoRA.
</Tip>
LoRAs also need to be used with another model:
```py
from diffusers import AutoPipelineForText2Image
import torch
pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
```
Then use the [`~loaders.LoraLoaderMixin.load_lora_weights`] method to load the [ostris/super-cereal-sdxl-lora](https://huggingface.co/ostris/super-cereal-sdxl-lora) weights and specify the weights filename from the repository:
```py
pipeline.load_lora_weights("ostris/super-cereal-sdxl-lora", weight_name="cereal_box_sdxl_v1.safetensors")
prompt = "bears, pizza bites"
image = pipeline(prompt).images[0]
image
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_lora.png" />
</div>
The [`~loaders.LoraLoaderMixin.load_lora_weights`] method loads LoRA weights into both the UNet and text encoder. It is the preferred way for loading LoRAs because it can handle cases where:
- the LoRA weights don't have separate identifiers for the UNet and text encoder
- the LoRA weights have separate identifiers for the UNet and text encoder
But if you only need to load LoRA weights into the UNet, then you can use the [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] method. Let's load the [jbilcke-hf/sdxl-cinematic-1](https://huggingface.co/jbilcke-hf/sdxl-cinematic-1) LoRA:
```py
from diffusers import AutoPipelineForText2Image
import torch
pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
pipeline.unet.load_attn_procs("jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors")
# use cnmt in the prompt to trigger the LoRA
prompt = "A cute cnmt eating a slice of pizza, stunning color scheme, masterpiece, illustration"
image = pipeline(prompt).images[0]
image
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_attn_proc.png" />
</div>
<Tip>
For both [`~loaders.LoraLoaderMixin.load_lora_weights`] and [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`], you can pass the `cross_attention_kwargs={"scale": 0.5}` parameter to adjust how much of the LoRA weights to use. A value of `0` is the same as only using the base model weights, and a value of `1` is equivalent to using the fully finetuned LoRA.
</Tip>
To unload the LoRA weights, use the [`~loaders.LoraLoaderMixin.unload_lora_weights`] method to discard the LoRA weights and restore the model to its original weights:
```py
pipeline.unload_lora_weights()
```
### Load multiple LoRAs
It can be fun to use multiple LoRAs together to create something entirely new and unique. The [`~loaders.LoraLoaderMixin.fuse_lora`] method allows you to fuse the LoRA weights with the original weights of the underlying model.
<Tip>
Fusing the weights can lead to a speedup in inference latency because you don't need to separately load the base model and LoRA! You can save your fused pipeline with [`~DiffusionPipeline.save_pretrained`] to avoid loading and fusing the weights every time you want to use the model.
</Tip>
Load an initial model:
```py
from diffusers import StableDiffusionXLPipeline, AutoencoderKL
import torch
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
pipeline = StableDiffusionXLPipeline.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
vae=vae,
torch_dtype=torch.float16,
).to("cuda")
```
Next, load the LoRA checkpoint and fuse it with the original weights. The `lora_scale` parameter controls how much to scale the output by with the LoRA weights. It is important to make the `lora_scale` adjustments in the [`~loaders.LoraLoaderMixin.fuse_lora`] method because it won't work if you try to pass `scale` to the `cross_attention_kwargs` in the pipeline.
If you need to reset the original model weights for any reason (use a different `lora_scale`), you should use the [`~loaders.LoraLoaderMixin.unfuse_lora`] method.
```py
pipeline.load_lora_weights("ostris/ikea-instructions-lora-sdxl")
pipeline.fuse_lora(lora_scale=0.7)
# to unfuse the LoRA weights
pipeline.unfuse_lora()
```
Then fuse this pipeline with the next set of LoRA weights:
```py
pipeline.load_lora_weights("ostris/super-cereal-sdxl-lora")
pipeline.fuse_lora(lora_scale=0.7)
```
<Tip warning={true}>
You can't unfuse multiple LoRA checkpoints, so if you need to reset the model to its original weights, you'll need to reload it.
</Tip>
Now you can generate an image that uses the weights from both LoRAs:
```py
prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration"
image = pipeline(prompt).images[0]
image
```
### 🤗 PEFT
<Tip>
Read the [Inference with 🤗 PEFT](../tutorials/using_peft_for_inference) tutorial to learn more about its integration with 🤗 Diffusers and how you can easily work with and juggle multiple adapters. You'll need to install 🤗 Diffusers and PEFT from source to run the example in this section.
</Tip>
Another way you can load and use multiple LoRAs is to specify the `adapter_name` parameter in [`~loaders.LoraLoaderMixin.load_lora_weights`]. This method takes advantage of the 🤗 PEFT integration. For example, load and name both LoRA weights:
```py
from diffusers import DiffusionPipeline
import torch
pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
pipeline.load_lora_weights("ostris/ikea-instructions-lora-sdxl", weight_name="ikea_instructions_xl_v1_5.safetensors", adapter_name="ikea")
pipeline.load_lora_weights("ostris/super-cereal-sdxl-lora", weight_name="cereal_box_sdxl_v1.safetensors", adapter_name="cereal")
```
Now use the [`~loaders.UNet2DConditionLoadersMixin.set_adapters`] to activate both LoRAs, and you can configure how much weight each LoRA should have on the output:
```py
pipeline.set_adapters(["ikea", "cereal"], adapter_weights=[0.7, 0.5])
```
Then, generate an image:
```py
prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration"
image = pipeline(prompt, num_inference_steps=30, cross_attention_kwargs={"scale": 1.0}).images[0]
image
```
### Kohya and TheLastBen
Other popular LoRA trainers from the community include those by [Kohya](https://github.com/kohya-ss/sd-scripts/) and [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion). These trainers create different LoRA checkpoints than those trained by 🤗 Diffusers, but they can still be loaded in the same way.
Let's download the [Blueprintify SD XL 1.0](https://civitai.com/models/150986/blueprintify-sd-xl-10) checkpoint from [Civitai](https://civitai.com/):
```sh
!wget https://civitai.com/api/download/models/168776 -O blueprintify-sd-xl-10.safetensors
```
Load the LoRA checkpoint with the [`~loaders.LoraLoaderMixin.load_lora_weights`] method, and specify the filename in the `weight_name` parameter:
```py
from diffusers import AutoPipelineForText2Image
import torch
pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
pipeline.load_lora_weights("path/to/weights", weight_name="blueprintify-sd-xl-10.safetensors")
```
Generate an image:
```py
# use bl3uprint in the prompt to trigger the LoRA
prompt = "bl3uprint, a highly detailed blueprint of the eiffel tower, explaining how to build all parts, many txt, blueprint grid backdrop"
image = pipeline(prompt).images[0]
image
```
<Tip warning={true}>
Some limitations of using Kohya LoRAs with 🤗 Diffusers include:
- Images may not look like those generated by UIs - like ComfyUI - for multiple reasons, which are explained [here](https://github.com/huggingface/diffusers/pull/4287/#issuecomment-1655110736).
- [LyCORIS checkpoints](https://github.com/KohakuBlueleaf/LyCORIS) aren't fully supported. The [`~loaders.LoraLoaderMixin.load_lora_weights`] method loads LyCORIS checkpoints with LoRA and LoCon modules, but Hada and LoKR are not supported.
</Tip>
Loading a checkpoint from TheLastBen is very similar. For example, to load the [TheLastBen/William_Eggleston_Style_SDXL](https://huggingface.co/TheLastBen/William_Eggleston_Style_SDXL) checkpoint:
```py
from diffusers import AutoPipelineForText2Image
import torch
pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
pipeline.load_lora_weights("TheLastBen/William_Eggleston_Style_SDXL", weight_name="wegg.safetensors")
# use by william eggleston in the prompt to trigger the LoRA
prompt = "a house by william eggleston, sunrays, beautiful, sunlight, sunrays, beautiful"
image = pipeline(prompt=prompt).images[0]
image
```
## IP-Adapter
[IP-Adapter](https://ip-adapter.github.io/) is an effective and lightweight adapter that adds image prompting capabilities to a diffusion model. This adapter works by decoupling the cross-attention layers of the image and text features. All the other model components are frozen and only the embedded image features in the UNet are trained. As a result, IP-Adapter files are typically only ~100MBs.
IP-Adapter works with most of our pipelines, including Stable Diffusion, Stable Diffusion XL (SDXL), ControlNet, T2I-Adapter, AnimateDiff. And you can use any custom models finetuned from the same base models. It also works with LCM-Lora out of box.
<Tip>
You can find official IP-Adapter checkpoints in [h94/IP-Adapter](https://huggingface.co/h94/IP-Adapter).
IP-Adapter was contributed by [okotaku](https://github.com/okotaku).
</Tip>
Let's first create a Stable Diffusion Pipeline.
```py
from diffusers import AutoPipelineForText2Image
import torch
from diffusers.utils import load_image
pipeline = AutoPipelineForText2Image.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
```
Now load the [h94/IP-Adapter](https://huggingface.co/h94/IP-Adapter) weights with the [`~loaders.IPAdapterMixin.load_ip_adapter`] method.
```py
pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
```
<Tip>
IP-Adapter relies on an image encoder to generate the image features, if your IP-Adapter weights folder contains a "image_encoder" subfolder, the image encoder will be automatically loaded and registered to the pipeline. Otherwise you can so load a [`~transformers.CLIPVisionModelWithProjection`] model and pass it to a Stable Diffusion pipeline when you create it.
```py
from diffusers import AutoPipelineForText2Image, CLIPVisionModelWithProjection
import torch
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
"h94/IP-Adapter",
subfolder="models/image_encoder",
torch_dtype=torch.float16,
).to("cuda")
pipeline = AutoPipelineForText2Image.from_pretrained("runwayml/stable-diffusion-v1-5", image_encoder=image_encoder, torch_dtype=torch.float16).to("cuda")
```
</Tip>
IP-Adapter allows you to use both image and text to condition the image generation process. For example, let's use the bear image from the [Textual Inversion](#textual-inversion) section as the image prompt (`ip_adapter_image`) along with a text prompt to add "sunglasses". 😎
```py
pipeline.set_ip_adapter_scale(0.6)
image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_neg_embed.png")
generator = torch.Generator(device="cpu").manual_seed(33)
images = pipeline(
prompt='best quality, high quality, wearing sunglasses',
ip_adapter_image=image,
negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
num_inference_steps=50,
generator=generator,
).images
images[0]
```
<div class="flex justify-center">
<img src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ip-bear.png" />
</div>
<Tip>
You can use the [`~loaders.IPAdapterMixin.set_ip_adapter_scale`] method to adjust the text prompt and image prompt condition ratio. If you're only using the image prompt, you should set the scale to `1.0`. You can lower the scale to get more generation diversity, but it'll be less aligned with the prompt.
`scale=0.5` can achieve good results in most cases when you use both text and image prompts.
</Tip>
IP-Adapter also works great with Image-to-Image and Inpainting pipelines. See below examples of how you can use it with Image-to-Image and Inpaint.
<hfoptions id="tasks">
<hfoption id="image-to-image">
```py
from diffusers import AutoPipelineForImage2Image
import torch
from diffusers.utils import load_image
pipeline = AutoPipelineForImage2Image.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/vermeer.jpg")
ip_image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/river.png")
pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
generator = torch.Generator(device="cpu").manual_seed(33)
images = pipeline(
prompt='best quality, high quality',
image = image,
ip_adapter_image=ip_image,
num_inference_steps=50,
generator=generator,
strength=0.6,
).images
images[0]
```
</hfoption>
<hfoption id="inpaint">
```py
from diffusers import AutoPipelineForInpaint
import torch
from diffusers.utils import load_image
pipeline = AutoPipelineForInpaint.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float).to("cuda")
image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/inpaint_image.png")
mask = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/mask.png")
ip_image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/girl.png")
image = image.resize((512, 768))
mask = mask.resize((512, 768))
pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
generator = torch.Generator(device="cpu").manual_seed(33)
images = pipeline(
prompt='best quality, high quality',
image = image,
mask_image = mask,
ip_adapter_image=ip_image,
negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
num_inference_steps=50,
generator=generator,
strength=0.5,
).images
images[0]
```
</hfoption>
</hfoptions>
IP-Adapters can also be used with [SDXL](../api/pipelines/stable_diffusion/stable_diffusion_xl.md)
```python
from diffusers import AutoPipelineForText2Image
from diffusers.utils import load_image
import torch
pipeline = AutoPipelineForText2Image.from_pretrained(
"stabilityai/stable-diffusion-xl-base-1.0",
torch_dtype=torch.float16
).to("cuda")
image = load_image("https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/watercolor_painting.jpeg")
pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name="ip-adapter_sdxl.bin")
generator = torch.Generator(device="cpu").manual_seed(33)
image = pipeline(
prompt="best quality, high quality",
ip_adapter_image=image,
negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
num_inference_steps=25,
generator=generator,
).images[0]
image.save("sdxl_t2i.png")
```
<div class="flex flex-row gap-4">
<div class="flex-1">
<img class="rounded-xl" src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/watercolor_painting.jpeg"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">input image</figcaption>
</div>
<div class="flex-1">
<img class="rounded-xl" src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/sdxl_t2i.png"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">adapted image</figcaption>
</div>
</div>
You can use the IP-Adapter face model to apply specific faces to your images. It is an effective way to maintain consistent characters in your image generations.
Weights are loaded with the same method used for the other IP-Adapters.
```python
# Load ip-adapter-full-face_sd15.bin
pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter-full-face_sd15.bin")
```
<Tip>
It is recommended to use `DDIMScheduler` and `EulerDiscreteScheduler` for face model.
</Tip>
```python
import torch
from diffusers import StableDiffusionPipeline, DDIMScheduler
from diffusers.utils import load_image
noise_scheduler = DDIMScheduler(
num_train_timesteps=1000,
beta_start=0.00085,
beta_end=0.012,
beta_schedule="scaled_linear",
clip_sample=False,
set_alpha_to_one=False,
steps_offset=1
)
pipeline = StableDiffusionPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5",
torch_dtype=torch.float16,
scheduler=noise_scheduler,
).to("cuda")
pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter-full-face_sd15.bin")
pipeline.set_ip_adapter_scale(0.7)
image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ai_face2.png")
generator = torch.Generator(device="cpu").manual_seed(33)
image = pipeline(
prompt="A photo of a girl wearing a black dress, holding red roses in hand, upper body, behind is the Eiffel Tower",
ip_adapter_image=image,
negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
num_inference_steps=50, num_images_per_prompt=1, width=512, height=704,
generator=generator,
).images[0]
```
<div class="flex flex-row gap-4">
<div class="flex-1">
<img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ai_face2.png"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">input image</figcaption>
</div>
<div class="flex-1">
<img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ipadapter_full_face_output.png"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">output image</figcaption>
</div>
</div>
### LCM-Lora
You can use IP-Adapter with LCM-Lora to achieve "instant fine-tune" with custom images. Note that you need to load IP-Adapter weights before loading the LCM-Lora weights.
```py
from diffusers import DiffusionPipeline, LCMScheduler
import torch
from diffusers.utils import load_image
model_id = "sd-dreambooth-library/herge-style"
lcm_lora_id = "latent-consistency/lcm-lora-sdv1-5"
pipe = DiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
pipe.load_lora_weights(lcm_lora_id)
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
pipe.enable_model_cpu_offload()
prompt = "best quality, high quality"
image = load_image("https://user-images.githubusercontent.com/24734142/266492875-2d50d223-8475-44f0-a7c6-08b51cb53572.png")
images = pipe(
prompt=prompt,
ip_adapter_image=image,
num_inference_steps=4,
guidance_scale=1,
).images[0]
```
### Other pipelines
IP-Adapter is compatible with any pipeline that (1) uses a text prompt and (2) uses Stable Diffusion or Stable Diffusion XL checkpoint. To use IP-Adapter with a different pipeline, all you need to do is to run `load_ip_adapter()` method after you create the pipeline, and then pass your image to the pipeline as `ip_adapter_image`
<Tip>
🤗 Diffusers currently only supports using IP-Adapter with some of the most popular pipelines, feel free to open a [feature request](https://github.com/huggingface/diffusers/issues/new/choose) if you have a cool use-case and require integrating IP-adapters with a pipeline that does not support it yet!
</Tip>
You can find below examples on how to use IP-Adapter with ControlNet and AnimateDiff.
<hfoptions id="model">
<hfoption id="ControlNet">
```
from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
import torch
from diffusers.utils import load_image
controlnet_model_path = "lllyasviel/control_v11f1p_sd15_depth"
controlnet = ControlNetModel.from_pretrained(controlnet_model_path, torch_dtype=torch.float16)
pipeline = StableDiffusionControlNetPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16)
pipeline.to("cuda")
image = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/statue.png")
depth_map = load_image("https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/depth.png")
pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
generator = torch.Generator(device="cpu").manual_seed(33)
images = pipeline(
prompt='best quality, high quality',
image=depth_map,
ip_adapter_image=image,
negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
num_inference_steps=50,
generator=generator,
).images
images[0]
```
<div class="flex flex-row gap-4">
<div class="flex-1">
<img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/statue.png"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">input image</figcaption>
</div>
<div class="flex-1">
<img class="rounded-xl" src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ipa-controlnet-out.png"/>
<figcaption class="mt-2 text-center text-sm text-gray-500">adapted image</figcaption>
</div>
</div>
</hfoption>
<hfoption id="AnimateDiff">
```py
# animate diff + ip adapter
import torch
from diffusers import MotionAdapter, AnimateDiffPipeline, DDIMScheduler
from diffusers.utils import export_to_gif, load_image
# Load the motion adapter
adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-2", torch_dtype=torch.float16)
# load SD 1.5 based finetuned model
model_id = "Lykon/DreamShaper"
pipe = AnimateDiffPipeline.from_pretrained(model_id, motion_adapter=adapter, torch_dtype=torch.float16)
# scheduler
scheduler = DDIMScheduler(
clip_sample=False,
beta_start=0.00085,
beta_end=0.012,
beta_schedule="linear",
timestep_spacing="trailing",
steps_offset=1
)
pipe.scheduler = scheduler
# enable memory savings
pipe.enable_vae_slicing()
pipe.enable_model_cpu_offload()
# load ip_adapter
pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
# load motion adapters
pipe.load_lora_weights("guoyww/animatediff-motion-lora-zoom-out", adapter_name="zoom-out")
pipe.load_lora_weights("guoyww/animatediff-motion-lora-tilt-up", adapter_name="tilt-up")
pipe.load_lora_weights("guoyww/animatediff-motion-lora-pan-left", adapter_name="pan-left")
seed = 42
image = load_image("https://user-images.githubusercontent.com/24734142/266492875-2d50d223-8475-44f0-a7c6-08b51cb53572.png")
images = [image] * 3
prompts = ["best quality, high quality"] * 3
negative_prompt = "bad quality, worst quality"
adapter_weights = [[0.75, 0.0, 0.0], [0.0, 0.0, 0.75], [0.0, 0.75, 0.75]]
# generate
output_frames = []
for prompt, image, adapter_weight in zip(prompts, images, adapter_weights):
pipe.set_adapters(["zoom-out", "tilt-up", "pan-left"], adapter_weights=adapter_weight)
output = pipe(
prompt= prompt,
num_frames=16,
guidance_scale=7.5,
num_inference_steps=30,
ip_adapter_image = image,
generator=torch.Generator("cpu").manual_seed(seed),
)
frames = output.frames[0]
output_frames.extend(frames)
export_to_gif(output_frames, "test_out_animation.gif")
```
</hfoption>
</hfoptions>
| huggingface/diffusers/blob/main/docs/source/en/using-diffusers/loading_adapters.md |
Gradio Demo: theme_extended_step_4
```
!pip install -q gradio
```
```
import gradio as gr
import time
theme = gr.themes.Default(primary_hue="blue").set(
loader_color="#FF0000",
slider_color="#FF0000",
)
with gr.Blocks(
theme=theme
) as demo:
textbox = gr.Textbox(label="Name")
slider = gr.Slider(label="Count", minimum=0, maximum=100, step=1)
with gr.Row():
button = gr.Button("Submit", variant="primary")
clear = gr.Button("Clear")
output = gr.Textbox(label="Output")
def repeat(name, count):
time.sleep(3)
return name * count
button.click(repeat, [textbox, slider], output)
if __name__ == "__main__":
demo.launch()
```
| gradio-app/gradio/blob/main/demo/theme_extended_step_4/run.ipynb |
Metric Card for Spearman Correlation Coefficient Metric (spearmanr)
## Metric Description
The Spearman rank-order correlation coefficient is a measure of the
relationship between two datasets. Like other correlation coefficients,
this one varies between -1 and +1 with 0 implying no correlation.
Positive correlations imply that as data in dataset x increases, so
does data in dataset y. Negative correlations imply that as x increases,
y decreases. Correlations of -1 or +1 imply an exact monotonic relationship.
Unlike the Pearson correlation, the Spearman correlation does not
assume that both datasets are normally distributed.
The p-value roughly indicates the probability of an uncorrelated system
producing datasets that have a Spearman correlation at least as extreme
as the one computed from these datasets. The p-values are not entirely
reliable but are probably reasonable for datasets larger than 500 or so.
## How to Use
At minimum, this metric only requires a `list` of predictions and a `list` of references:
```python
>>> spearmanr_metric = datasets.load_metric("spearmanr")
>>> results = spearmanr_metric.compute(references=[1, 2, 3, 4, 5], predictions=[10, 9, 2.5, 6, 4])
>>> print(results)
{'spearmanr': -0.7}
```
### Inputs
- **`predictions`** (`list` of `float`): Predicted labels, as returned by a model.
- **`references`** (`list` of `float`): Ground truth labels.
- **`return_pvalue`** (`bool`): If `True`, returns the p-value. If `False`, returns
only the spearmanr score. Defaults to `False`.
### Output Values
- **`spearmanr`** (`float`): Spearman correlation coefficient.
- **`p-value`** (`float`): p-value. **Note**: is only returned
if `return_pvalue=True` is input.
If `return_pvalue=False`, the output is a `dict` with one value, as below:
```python
{'spearmanr': -0.7}
```
Otherwise, if `return_pvalue=True`, the output is a `dict` containing a the `spearmanr` value as well as the corresponding `pvalue`:
```python
{'spearmanr': -0.7, 'spearmanr_pvalue': 0.1881204043741873}
```
Spearman rank-order correlations can take on any value from `-1` to `1`, inclusive.
The p-values can take on any value from `0` to `1`, inclusive.
#### Values from Popular Papers
### Examples
A basic example:
```python
>>> spearmanr_metric = datasets.load_metric("spearmanr")
>>> results = spearmanr_metric.compute(references=[1, 2, 3, 4, 5], predictions=[10, 9, 2.5, 6, 4])
>>> print(results)
{'spearmanr': -0.7}
```
The same example, but that also returns the pvalue:
```python
>>> spearmanr_metric = datasets.load_metric("spearmanr")
>>> results = spearmanr_metric.compute(references=[1, 2, 3, 4, 5], predictions=[10, 9, 2.5, 6, 4], return_pvalue=True)
>>> print(results)
{'spearmanr': -0.7, 'spearmanr_pvalue': 0.1881204043741873
>>> print(results['spearmanr'])
-0.7
>>> print(results['spearmanr_pvalue'])
0.1881204043741873
```
## Limitations and Bias
## Citation
```bibtex
@book{kokoska2000crc,
title={CRC standard probability and statistics tables and formulae},
author={Kokoska, Stephen and Zwillinger, Daniel},
year={2000},
publisher={Crc Press}
}
@article{2020SciPy-NMeth,
author = {Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E. and
Haberland, Matt and Reddy, Tyler and Cournapeau, David and
Burovski, Evgeni and Peterson, Pearu and Weckesser, Warren and
Bright, Jonathan and {van der Walt}, St{\'e}fan J. and
Brett, Matthew and Wilson, Joshua and Millman, K. Jarrod and
Mayorov, Nikolay and Nelson, Andrew R. J. and Jones, Eric and
Kern, Robert and Larson, Eric and Carey, C J and
Polat, {\.I}lhan and Feng, Yu and Moore, Eric W. and
{VanderPlas}, Jake and Laxalde, Denis and Perktold, Josef and
Cimrman, Robert and Henriksen, Ian and Quintero, E. A. and
Harris, Charles R. and Archibald, Anne M. and
Ribeiro, Ant{\^o}nio H. and Pedregosa, Fabian and
{van Mulbregt}, Paul and {SciPy 1.0 Contributors}},
title = {{{SciPy} 1.0: Fundamental Algorithms for Scientific
Computing in Python}},
journal = {Nature Methods},
year = {2020},
volume = {17},
pages = {261--272},
adsurl = {https://rdcu.be/b08Wh},
doi = {10.1038/s41592-019-0686-2},
}
```
## Further References
*Add any useful further references.*
| huggingface/datasets/blob/main/metrics/spearmanr/README.md |
--
title: "Accelerating Hugging Face Transformers with AWS Inferentia2"
thumbnail: /blog/assets/140_accelerate_transformers_with_inferentia2/thumbnail.png
authors:
- user: philschmid
- user: juliensimon
---
# Accelerating Hugging Face Transformers with AWS Inferentia2
<script async defer src="https://unpkg.com/medium-zoom-element@0/dist/medium-zoom-element.min.js"></script>
In the last five years, Transformer models [[1](https://arxiv.org/abs/1706.03762)] have become the _de facto_ standard for many machine learning (ML) tasks, such as natural language processing (NLP), computer vision (CV), speech, and more. Today, many data scientists and ML engineers rely on popular transformer architectures like BERT [[2](https://arxiv.org/abs/1810.04805)], RoBERTa [[3](https://arxiv.org/abs/1907.11692)], the Vision Transformer [[4](https://arxiv.org/abs/2010.11929)], or any of the 130,000+ pre-trained models available on the [Hugging Face](https://huggingface.co) hub to solve complex business problems with state-of-the-art accuracy.
However, for all their greatness, Transformers can be challenging to deploy in production. On top of the infrastructure plumbing typically associated with model deployment, which we largely solved with our [Inference Endpoints](https://huggingface.co/inference-endpoints) service, Transformers are large models which routinely exceed the multi-gigabyte mark. Large language models (LLMs) like [GPT-J-6B](https://huggingface.co/EleutherAI/gpt-j-6B), [Flan-T5](https://huggingface.co/google/flan-t5-xxl), or [Opt-30B](https://huggingface.co/facebook/opt-30b) are in the tens of gigabytes, not to mention behemoths like [BLOOM](https://huggingface.co/bigscience/bloom), our very own LLM, which clocks in at 350 gigabytes.
Fitting these models on a single accelerator can be quite difficult, let alone getting the high throughput and low inference latency that applications require, like conversational applications and search. So far, ML experts have designed complex manual techniques to slice large models, distribute them on a cluster of accelerators, and optimize their latency. Unfortunately, this work is extremely difficult, time-consuming, and completely out of reach for many ML practitioners.
At Hugging Face, we're democratizing ML and always looking to partner with companies who also believe that every developer and organization should benefit from state-of-the-art models. For this purpose, we're excited to partner with Amazon Web Services to optimize Hugging Face Transformers for AWS [Inferentia 2](https://aws.amazon.com/machine-learning/inferentia/)! It’s a new purpose-built inference accelerator that delivers unprecedented levels of throughput, latency, performance per watt, and scalability.
## Introducing AWS Inferentia2
AWS Inferentia2 is the next generation to Inferentia1 launched in 2019. Powered by Inferentia1, Amazon EC2 Inf1 instances delivered 25% higher throughput and 70% lower cost than comparable G5 instances based on NVIDIA A10G GPU, and with Inferentia2, AWS is pushing the envelope again.
The new Inferentia2 chip delivers a 4x throughput increase and a 10x latency reduction compared to Inferentia. Likewise, the new [Amazon EC2 Inf2](https://aws.amazon.com/de/ec2/instance-types/inf2/) instances have up to 2.6x better throughput, 8.1x lower latency, and 50% better performance per watt than comparable G5 instances. Inferentia 2 gives you the best of both worlds: cost-per-inference optimization thanks to high throughput and response time for your application thanks to low inference latency.
Inf2 instances are available in multiple sizes, which are equipped with between 1 to 12 Inferentia 2 chips. When several chips are present, they are interconnected by a blazing-fast direct Inferentia2 to Inferentia2 connectivity for distributed inference on large models. For example, the largest instance size, inf2.48xlarge, has 12 chips and enough memory to load a 175-billion parameter model like GPT-3 or BLOOM.
Thankfully none of this comes at the expense of development complexity. With [optimum neuron](https://github.com/huggingface/optimum-neuron), you don't need to slice or modify your model. Because of the native integration in [AWS Neuron SDK](https://github.com/aws-neuron/aws-neuron-sdk), all it takes is a single line of code to compile your model for Inferentia 2. You can experiment in minutes! Test the performance your model could reach on Inferentia 2 and see for yourself.
Speaking of, let’s show you how several Hugging Face models run on Inferentia 2. Benchmarking time!
## Benchmarking Hugging Face Models on AWS Inferentia 2
We evaluated some of the most popular NLP models from the [Hugging Face Hub](https://huggingface.co/models) including BERT, RoBERTa, DistilBERT, and vision models like Vision Transformers.
The first benchmark compares the performance of Inferentia, Inferentia 2, and GPUs. We ran all experiments on AWS with the following instance types:
* Inferentia1 - [inf1.2xlarge](https://aws.amazon.com/ec2/instance-types/inf1/?nc1=h_ls) powered by a single Inferentia chip.
* Inferentia2 - [inf2.xlarge](https://aws.amazon.com/ec2/instance-types/inf2/?nc1=h_ls) powered by a single Inferentia2 chip.
* GPU - [g5.2xlarge](https://aws.amazon.com/ec2/instance-types/g5/) powered by a single NVIDIA A10G GPU.
_Note: that we did not optimize the model for the GPU environment, the models were evaluated in fp32._
When it comes to benchmarking Transformer models, there are two metrics that are most adopted:
* **Latency**: the time it takes for the model to perform a single prediction (pre-process, prediction, post-process).
* **Throughput**: the number of executions performed in a fixed amount of time for one benchmark configuration
We looked at latency across different setups and models to understand the benefits and tradeoffs of the new Inferentia2 instance. If you want to run the benchmark yourself, we created a [Github repository](https://github.com/philschmid/aws-neuron-samples/tree/main/benchmark) with all the information and scripts to do so.
### Results
The benchmark confirms that the performance improvements claimed by AWS can be reproduced and validated by real use-cases and examples. On average, AWS Inferentia2 delivers 4.5x better latency than NVIDIA A10G GPUs and 4x better latency than Inferentia1 instances.
We ran 144 experiments on 6 different model architectures:
* Accelerators: Inf1, Inf2, NVIDIA A10G
* Models: [BERT-base](https://huggingface.co/bert-base-uncased), [BERT-Large](https://huggingface.co/bert-large-uncased), [RoBERTa-base](https://huggingface.co/roberta-base), [DistilBERT](https://huggingface.co/distilbert-base-uncased), [ALBERT-base](https://huggingface.co/albert-base-v2), [ViT-base](https://huggingface.co/google/vit-base-patch16-224)
* Sequence length: 8, 16, 32, 64, 128, 256, 512
* Batch size: 1
In each experiment, we collected numbers for p95 latency. You can find the full details of the benchmark in this spreadsheet: [HuggingFace: Benchmark Inferentia2](https://docs.google.com/spreadsheets/d/1AULEHBu5Gw6ABN8Ls6aSB2CeZyTIP_y5K7gC7M3MXqs/edit?usp=sharing).
Let’s highlight a few insights of the benchmark.
### BERT-base
Here is the latency comparison for running [BERT-base](https://huggingface.co/bert-base-uncased) on each of the infrastructure setups, with a logarithmic scale for latency. It is remarkable to see how Inferentia2 outperforms all other setups by ~6x for sequence lengths up to 256.
<br>
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="BERT-base p95 latency" src="assets/140_accelerate_transformers_with_inferentia2/bert.png"></medium-zoom>
<figcaption>Figure 1. BERT-base p95 latency</figcaption>
</figure>
<br>
### Vision Transformer
Here is the latency comparison for running [ViT-base](https://huggingface.co/google/vit-base-patch16-224) on the different infrastructure setups. Inferentia2 delivers 2x better latency than the NVIDIA A10G, with the potential to greatly help companies move from traditional architectures, like CNNs, to Transformers for - real-time applications.
<br>
<figure class="image table text-center m-0 w-full">
<medium-zoom background="rgba(0,0,0,.7)" alt="ViT p95 latency" src="assets/140_accelerate_transformers_with_inferentia2/vit.png"></medium-zoom>
<figcaption>Figure 1. ViT p95 latency</figcaption>
</figure>
<br>
## Conclusion
Transformer models have emerged as the go-to solution for many machine learning tasks. However, deploying them in production has been challenging due to their large size and latency requirements. Thanks to AWS Inferentia2 and the collaboration between Hugging Face and AWS, developers and organizations can now leverage the benefits of state-of-the-art models without the prior need for extensive machine learning expertise. You can start testing for as low as 0.76$/h.
The initial benchmarking results are promising, and show that Inferentia2 delivers superior latency performance when compared to both Inferentia and NVIDIA A10G GPUs. This latest breakthrough promises high-quality machine learning models can be made available to a much broader audience delivering AI accessibility to everyone. | huggingface/blog/blob/main/accelerate-transformers-with-inferentia2.md |
Metadata Parsing
Given the simplicity of the format, it's very simple and efficient to fetch and parse metadata about Safetensors weights – i.e. the list of tensors, their types, and their shapes or numbers of parameters – using small [(Range) HTTP requests](https://developer.mozilla.org/en-US/docs/Web/HTTP/Range_requests).
This parsing has been implemented in JS in [`huggingface.js`](https://huggingface.co/docs/huggingface.js/main/en/hub/modules#parsesafetensorsmetadata) (sample code follows below), but it would be similar in any language.
## Example use case
There can be many potential use cases. For instance, we use it on the HuggingFace Hub to display info about models which have safetensors weights:
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/safetensors/model-page-light.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/safetensors/model-page-dark.png"/>
</div>
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/safetensors/view-all-tensors-light.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/safetensors/view-all-tensors-dark.png"/>
</div>
## Usage
### JavaScript/TypeScript[[js]]
Using [`huggingface.js`](https://huggingface.co/docs/huggingface.js)
```ts
import { parseSafetensorsMetadata } from "@huggingface/hub";
const info = await parseSafetensorsMetadata({
repo: { type: "model", name: "bigscience/bloom" },
});
console.log(info)
// {
// sharded: true,
// index: {
// metadata: { total_size: 352494542848 },
// weight_map: {
// 'h.0.input_layernorm.bias': 'model_00002-of-00072.safetensors',
// ...
// }
// },
// headers: {
// __metadata__: {'format': 'pt'},
// 'h.2.attn.c_attn.weight': {'dtype': 'F32', 'shape': [768, 2304], 'data_offsets': [541012992, 548090880]},
// ...
// }
// }
```
Depending on whether the safetensors weights are sharded into multiple files or not, the output of the call above will be:
```ts
export type SafetensorsParseFromRepo =
| {
sharded: false;
header: SafetensorsFileHeader;
}
| {
sharded: true;
index: SafetensorsIndexJson;
headers: SafetensorsShardedHeaders;
};
```
where the underlying `types` are the following:
```ts
type FileName = string;
type TensorName = string;
type Dtype = "F64" | "F32" | "F16" | "BF16" | "I64" | "I32" | "I16" | "I8" | "U8" | "BOOL";
interface TensorInfo {
dtype: Dtype;
shape: number[];
data_offsets: [number, number];
}
type SafetensorsFileHeader = Record<TensorName, TensorInfo> & {
__metadata__: Record<string, string>;
};
interface SafetensorsIndexJson {
weight_map: Record<TensorName, FileName>;
}
export type SafetensorsShardedHeaders = Record<FileName, SafetensorsFileHeader>;
```
### Python
In this example python script, we are parsing metadata of [gpt2](https://huggingface.co/gpt2/blob/main/model.safetensors).
```python
import requests # pip install requests
import struct
def parse_single_file(url):
# Fetch the first 8 bytes of the file
headers = {'Range': 'bytes=0-7'}
response = requests.get(url, headers=headers)
# Interpret the bytes as a little-endian unsigned 64-bit integer
length_of_header = struct.unpack('<Q', response.content)[0]
# Fetch length_of_header bytes starting from the 9th byte
headers = {'Range': f'bytes=8-{7 + length_of_header}'}
response = requests.get(url, headers=headers)
# Interpret the response as a JSON object
header = response.json()
return header
url = "https://huggingface.co/gpt2/resolve/main/model.safetensors"
header = parse_single_file(url)
print(header)
# {
# "__metadata__": { "format": "pt" },
# "h.10.ln_1.weight": {
# "dtype": "F32",
# "shape": [768],
# "data_offsets": [223154176, 223157248]
# },
# ...
# }
```
## Example output
For instance, here are the number of params per dtype for a few models on the HuggingFace Hub. Also see [this issue](https://github.com/huggingface/safetensors/issues/44) for more examples of usage.
model | safetensors | params
--- | --- | ---
[gpt2](https://huggingface.co/gpt2?show_tensors=true) | single-file | { 'F32' => 137022720 }
[roberta-base](https://huggingface.co/roberta-base?show_tensors=true) | single-file | { 'F32' => 124697433, 'I64' => 514 }
[Jean-Baptiste/camembert-ner](https://huggingface.co/Jean-Baptiste/camembert-ner?show_tensors=true) | single-file | { 'F32' => 110035205, 'I64' => 514 }
[roberta-large](https://huggingface.co/roberta-large?show_tensors=true) | single-file | { 'F32' => 355412057, 'I64' => 514 }
[distilbert-base-german-cased](https://huggingface.co/distilbert-base-german-cased?show_tensors=true) | single-file | { 'F32' => 67431550 }
[EleutherAI/gpt-neox-20b](https://huggingface.co/EleutherAI/gpt-neox-20b?show_tensors=true) | sharded | { 'F16' => 20554568208, 'U8' => 184549376 }
[bigscience/bloom-560m](https://huggingface.co/bigscience/bloom-560m?show_tensors=true) | single-file | { 'F16' => 559214592 }
[bigscience/bloom](https://huggingface.co/bigscience/bloom?show_tensors=true) | sharded | { 'BF16' => 176247271424 }
[bigscience/bloom-3b](https://huggingface.co/bigscience/bloom-3b?show_tensors=true) | single-file | { 'F16' => 3002557440 }
| huggingface/safetensors/blob/main/docs/source/metadata_parsing.mdx |
@gradio/textbox
## 0.4.7
### Patch Changes
- Updated dependencies [[`828fb9e`](https://github.com/gradio-app/gradio/commit/828fb9e6ce15b6ea08318675a2361117596a1b5d), [`73268ee`](https://github.com/gradio-app/gradio/commit/73268ee2e39f23ebdd1e927cb49b8d79c4b9a144)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.4.6
### Patch Changes
- Updated dependencies [[`053bec9`](https://github.com/gradio-app/gradio/commit/053bec98be1127e083414024e02cf0bebb0b5142), [`4d1cbbc`](https://github.com/gradio-app/gradio/commit/4d1cbbcf30833ef1de2d2d2710c7492a379a9a00)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 0.4.5
### Fixes
- [#6635](https://github.com/gradio-app/gradio/pull/6635) [`b639e04`](https://github.com/gradio-app/gradio/commit/b639e040741e6c0d9104271c81415d7befbd8cf3) - Quick Image + Text Component Fixes. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
## 0.4.4
### Patch Changes
- Updated dependencies [[`9caddc17b`](https://github.com/gradio-app/gradio/commit/9caddc17b1dea8da1af8ba724c6a5eab04ce0ed8)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 0.4.3
### Patch Changes
- Updated dependencies [[`f816136a0`](https://github.com/gradio-app/gradio/commit/f816136a039fa6011be9c4fb14f573e4050a681a)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 0.4.2
### Fixes
- [#6323](https://github.com/gradio-app/gradio/pull/6323) [`55fda81fa`](https://github.com/gradio-app/gradio/commit/55fda81fa5918b48952729232d6e2fc55af9351d) - Textbox and Code Component Blur/Focus Fixes. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
## 0.4.1
### Patch Changes
- Updated dependencies [[`3cdeabc68`](https://github.com/gradio-app/gradio/commit/3cdeabc6843000310e1a9e1d17190ecbf3bbc780), [`fad92c29d`](https://github.com/gradio-app/gradio/commit/fad92c29dc1f5cd84341aae417c495b33e01245f)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.4.0
### Features
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Publish all components to npm. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Custom components. Thanks [@pngwn](https://github.com/pngwn)!
## 0.4.0-beta.8
### Features
- [#6136](https://github.com/gradio-app/gradio/pull/6136) [`667802a6c`](https://github.com/gradio-app/gradio/commit/667802a6cdbfb2ce454a3be5a78e0990b194548a) - JS Component Documentation. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#6149](https://github.com/gradio-app/gradio/pull/6149) [`90318b1dd`](https://github.com/gradio-app/gradio/commit/90318b1dd118ae08a695a50e7c556226234ab6dc) - swap `mode` on the frontned to `interactive` to match the backend. Thanks [@pngwn](https://github.com/pngwn)!
## 0.4.0-beta.7
### Features
- [#6016](https://github.com/gradio-app/gradio/pull/6016) [`83e947676`](https://github.com/gradio-app/gradio/commit/83e947676d327ca2ab6ae2a2d710c78961c771a0) - Format js in v4 branch. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
### Fixes
- [#6046](https://github.com/gradio-app/gradio/pull/6046) [`dbb7de5e0`](https://github.com/gradio-app/gradio/commit/dbb7de5e02c53fee05889d696d764d212cb96c74) - fix tests. Thanks [@pngwn](https://github.com/pngwn)!
## 0.4.0-beta.6
### Features
- [#5960](https://github.com/gradio-app/gradio/pull/5960) [`319c30f3f`](https://github.com/gradio-app/gradio/commit/319c30f3fccf23bfe1da6c9b132a6a99d59652f7) - rererefactor frontend files. Thanks [@pngwn](https://github.com/pngwn)!
- [#5938](https://github.com/gradio-app/gradio/pull/5938) [`13ed8a485`](https://github.com/gradio-app/gradio/commit/13ed8a485d5e31d7d75af87fe8654b661edcca93) - V4: Use beta release versions for '@gradio' packages. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
## 0.4.3
### Patch Changes
- Updated dependencies [[`e70805d54`](https://github.com/gradio-app/gradio/commit/e70805d54cc792452545f5d8eccc1aa0212a4695)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 0.4.2
### Patch Changes
- Updated dependencies []:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 0.4.1
### Patch Changes
- Updated dependencies [[`8f0fed857`](https://github.com/gradio-app/gradio/commit/8f0fed857d156830626eb48b469d54d211a582d2)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 0.4.0
### Features
- [#5652](https://github.com/gradio-app/gradio/pull/5652) [`2e25d4305`](https://github.com/gradio-app/gradio/commit/2e25d430582264945ae3316acd04c4453a25ce38) - Pause autoscrolling if a user scrolls up in a `gr.Textbox` and resume autoscrolling if they go all the way down. Thanks [@abidlabs](https://github.com/abidlabs)!
- [#5554](https://github.com/gradio-app/gradio/pull/5554) [`75ddeb390`](https://github.com/gradio-app/gradio/commit/75ddeb390d665d4484667390a97442081b49a423) - Accessibility Improvements. Thanks [@hannahblair](https://github.com/hannahblair)!
## 0.3.0
### Features
- [#5488](https://github.com/gradio-app/gradio/pull/5488) [`8909e42a`](https://github.com/gradio-app/gradio/commit/8909e42a7c6272358ad413588d27a5124d151205) - Adds `autoscroll` param to `gr.Textbox()`. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
## 0.2.0
### Features
- [#5417](https://github.com/gradio-app/gradio/pull/5417) [`d14d63e3`](https://github.com/gradio-app/gradio/commit/d14d63e30c4af3f9c2a664fd11b0a01943a8300c) - Auto scroll to bottom of textbox. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
## 0.1.3
### Patch Changes
- Updated dependencies [[`abf1c57d`](https://github.com/gradio-app/gradio/commit/abf1c57d7d85de0df233ee3b38aeb38b638477db)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 0.1.2
### Fixes
- [#5324](https://github.com/gradio-app/gradio/pull/5324) [`31996c99`](https://github.com/gradio-app/gradio/commit/31996c991d6bfca8cef975eb8e3c9f61a7aced19) - ensure login form has correct styles. Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.1
### Highlights
#### Improve startup performance and markdown support ([#5279](https://github.com/gradio-app/gradio/pull/5279) [`fe057300`](https://github.com/gradio-app/gradio/commit/fe057300f0672c62dab9d9b4501054ac5d45a4ec))
##### Improved markdown support
We now have better support for markdown in `gr.Markdown` and `gr.Dataframe`. Including syntax highlighting and Github Flavoured Markdown. We also have more consistent markdown behaviour and styling.
##### Various performance improvements
These improvements will be particularly beneficial to large applications.
- Rather than attaching events manually, they are now delegated, leading to a significant performance improvement and addressing a performance regression introduced in a recent version of Gradio. App startup for large applications is now around twice as fast.
- Optimised the mounting of individual components, leading to a modest performance improvement during startup (~30%).
- Corrected an issue that was causing markdown to re-render infinitely.
- Ensured that the `gr.3DModel` does re-render prematurely.
Thanks [@pngwn](https://github.com/pngwn)!
## 0.1.0
### Features
- [#5005](https://github.com/gradio-app/gradio/pull/5005) [`f5539c76`](https://github.com/gradio-app/gradio/commit/f5539c7618e31451420bd3228754774da14dc65f) - Enhancement: Add focus event to textbox and number component. Thanks [@JodyZ0203](https://github.com/JodyZ0203)!
### Fixes
- [#5114](https://github.com/gradio-app/gradio/pull/5114) [`56d2609d`](https://github.com/gradio-app/gradio/commit/56d2609de93387a75dc82b1c06c1240c5b28c0b8) - Reset textbox value to empty string when value is None. Thanks [@hannahblair](https://github.com/hannahblair)!
| gradio-app/gradio/blob/main/js/textbox/CHANGELOG.md |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# EnCodec
## Overview
The EnCodec neural codec model was proposed in [High Fidelity Neural Audio Compression](https://arxiv.org/abs/2210.13438) by Alexandre Défossez, Jade Copet, Gabriel Synnaeve, Yossi Adi.
The abstract from the paper is the following:
*We introduce a state-of-the-art real-time, high-fidelity, audio codec leveraging neural networks. It consists in a streaming encoder-decoder architecture with quantized latent space trained in an end-to-end fashion. We simplify and speed-up the training by using a single multiscale spectrogram adversary that efficiently reduces artifacts and produce high-quality samples. We introduce a novel loss balancer mechanism to stabilize training: the weight of a loss now defines the fraction of the overall gradient it should represent, thus decoupling the choice of this hyper-parameter from the typical scale of the loss. Finally, we study how lightweight Transformer models can be used to further compress the obtained representation by up to 40%, while staying faster than real time. We provide a detailed description of the key design choices of the proposed model including: training objective, architectural changes and a study of various perceptual loss functions. We present an extensive subjective evaluation (MUSHRA tests) together with an ablation study for a range of bandwidths and audio domains, including speech, noisy-reverberant speech, and music. Our approach is superior to the baselines methods across all evaluated settings, considering both 24 kHz monophonic and 48 kHz stereophonic audio.*
This model was contributed by [Matthijs](https://huggingface.co/Matthijs), [Patrick Von Platen](https://huggingface.co/patrickvonplaten) and [Arthur Zucker](https://huggingface.co/ArthurZ).
The original code can be found [here](https://github.com/facebookresearch/encodec).
## Usage example
Here is a quick example of how to encode and decode an audio using this model:
```python
>>> from datasets import load_dataset, Audio
>>> from transformers import EncodecModel, AutoProcessor
>>> librispeech_dummy = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> model = EncodecModel.from_pretrained("facebook/encodec_24khz")
>>> processor = AutoProcessor.from_pretrained("facebook/encodec_24khz")
>>> librispeech_dummy = librispeech_dummy.cast_column("audio", Audio(sampling_rate=processor.sampling_rate))
>>> audio_sample = librispeech_dummy[-1]["audio"]["array"]
>>> inputs = processor(raw_audio=audio_sample, sampling_rate=processor.sampling_rate, return_tensors="pt")
>>> encoder_outputs = model.encode(inputs["input_values"], inputs["padding_mask"])
>>> audio_values = model.decode(encoder_outputs.audio_codes, encoder_outputs.audio_scales, inputs["padding_mask"])[0]
>>> # or the equivalent with a forward pass
>>> audio_values = model(inputs["input_values"], inputs["padding_mask"]).audio_values
```
## EncodecConfig
[[autodoc]] EncodecConfig
## EncodecFeatureExtractor
[[autodoc]] EncodecFeatureExtractor
- __call__
## EncodecModel
[[autodoc]] EncodecModel
- decode
- encode
- forward
| huggingface/transformers/blob/main/docs/source/en/model_doc/encodec.md |
Torch API
[[autodoc]] safetensors.torch.load_file
[[autodoc]] safetensors.torch.load
[[autodoc]] safetensors.torch.save_file
[[autodoc]] safetensors.torch.save
[[autodoc]] safetensors.torch.load_model
[[autodoc]] safetensors.torch.save_model
| huggingface/safetensors/blob/main/docs/source/api/torch.mdx |
](https://github.com/huggingface/optimum/actions/workflows/test_onnxruntime.yml)
# Hugging Face Optimum
🤗 Optimum is an extension of 🤗 Transformers and Diffusers, providing a set of optimization tools enabling maximum efficiency to train and run models on targeted hardware, while keeping things easy to use.
## Installation
🤗 Optimum can be installed using `pip` as follows:
```bash
python -m pip install optimum
```
If you'd like to use the accelerator-specific features of 🤗 Optimum, you can install the required dependencies according to the table below:
| Accelerator | Installation |
|:-----------------------------------------------------------------------------------------------------------------------|:--------------------------------------------------|
| [ONNX Runtime](https://huggingface.co/docs/optimum/onnxruntime/overview) | `pip install --upgrade-strategy eager optimum[onnxruntime]` |
| [Intel Neural Compressor](https://huggingface.co/docs/optimum/intel/index) | `pip install --upgrade-strategy eager optimum[neural-compressor]`|
| [OpenVINO](https://huggingface.co/docs/optimum/intel/index) | `pip install --upgrade-strategy eager optimum[openvino,nncf]` |
| [AMD Instinct GPUs and Ryzen AI NPU](https://huggingface.co/docs/optimum/amd/index) | `pip install --upgrade-strategy eager optimum[amd]` |
| [Habana Gaudi Processor (HPU)](https://huggingface.co/docs/optimum/habana/index) | `pip install --upgrade-strategy eager optimum[habana]` |
| [FuriosaAI](https://huggingface.co/docs/optimum/furiosa/index) | `pip install --upgrade-strategy eager optimum[furiosa]` |
The `--upgrade-strategy eager` option is needed to ensure the different packages are upgraded to the latest possible version.
To install from source:
```bash
python -m pip install git+https://github.com/huggingface/optimum.git
```
For the accelerator-specific features, append `optimum[accelerator_type]` to the above command:
```bash
python -m pip install optimum[onnxruntime]@git+https://github.com/huggingface/optimum.git
```
## Accelerated Inference
🤗 Optimum provides multiple tools to export and run optimized models on various ecosystems:
- [ONNX](https://huggingface.co/docs/optimum/exporters/onnx/usage_guides/export_a_model) / [ONNX Runtime](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/models)
- TensorFlow Lite
- [OpenVINO](https://huggingface.co/docs/optimum/intel/inference)
- Habana first-gen Gaudi / Gaudi2, more details [here](https://huggingface.co/docs/optimum/main/en/habana/usage_guides/accelerate_inference)
The [export](https://huggingface.co/docs/optimum/exporters/overview) and optimizations can be done both programmatically and with a command line.
### Features summary
| Features | [ONNX Runtime](https://huggingface.co/docs/optimum/main/en/onnxruntime/overview)| [Neural Compressor](https://huggingface.co/docs/optimum/main/en/intel/optimization_inc)| [OpenVINO](https://huggingface.co/docs/optimum/main/en/intel/inference)| [TensorFlow Lite](https://huggingface.co/docs/optimum/main/en/exporters/tflite/overview)|
|:----------------------------------:|:------------------:|:------------------:|:------------------:|:------------------:|
| Graph optimization | :heavy_check_mark: | N/A | :heavy_check_mark: | N/A |
| Post-training dynamic quantization | :heavy_check_mark: | :heavy_check_mark: | N/A | :heavy_check_mark: |
| Post-training static quantization | :heavy_check_mark: | :heavy_check_mark: | :heavy_check_mark: | :heavy_check_mark: |
| Quantization Aware Training (QAT) | N/A | :heavy_check_mark: | :heavy_check_mark: | N/A |
| FP16 (half precision) | :heavy_check_mark: | N/A | :heavy_check_mark: | :heavy_check_mark: |
| Pruning | N/A | :heavy_check_mark: | :heavy_check_mark: | N/A |
| Knowledge Distillation | N/A | :heavy_check_mark: | :heavy_check_mark: | N/A |
### OpenVINO
Before you begin, make sure you have all the necessary libraries installed :
```bash
pip install --upgrade-strategy eager optimum[openvino,nncf]
```
It is possible to export 🤗 Transformers and Diffusers models to the OpenVINO format easily:
```bash
optimum-cli export openvino --model distilbert-base-uncased-finetuned-sst-2-english distilbert_sst2_ov
```
If you add `--int8`, the weights will be quantized to INT8. Static quantization can also be applied on the activations using [NNCF](https://github.com/openvinotoolkit/nncf), more information can be found in the [documentation](https://huggingface.co/docs/optimum/main/en/intel/optimization_ov).
To load a model and run inference with OpenVINO Runtime, you can just replace your `AutoModelForXxx` class with the corresponding `OVModelForXxx` class. To load a PyTorch checkpoint and convert it to the OpenVINO format on-the-fly, you can set `export=True` when loading your model.
```diff
- from transformers import AutoModelForSequenceClassification
+ from optimum.intel import OVModelForSequenceClassification
from transformers import AutoTokenizer, pipeline
model_id = "distilbert-base-uncased-finetuned-sst-2-english"
tokenizer = AutoTokenizer.from_pretrained(model_id)
- model = AutoModelForSequenceClassification.from_pretrained(model_id)
+ model = OVModelForSequenceClassification.from_pretrained("distilbert_sst2_ov")
classifier = pipeline("text-classification", model=model, tokenizer=tokenizer)
results = classifier("He's a dreadful magician.")
```
You can find more examples in the [documentation](https://huggingface.co/docs/optimum/intel/inference) and in the [examples](https://github.com/huggingface/optimum-intel/tree/main/examples/openvino).
### Neural Compressor
Before you begin, make sure you have all the necessary libraries installed :
```bash
pip install --upgrade-strategy eager optimum[neural-compressor]
```
Dynamic quantization can be applied on your model:
```bash
optimum-cli inc quantize --model distilbert-base-cased-distilled-squad --output ./quantized_distilbert
```
To load a model quantized with Intel Neural Compressor, hosted locally or on the 🤗 hub, you can do as follows :
```python
from optimum.intel import INCModelForSequenceClassification
model_id = "Intel/distilbert-base-uncased-finetuned-sst-2-english-int8-dynamic"
model = INCModelForSequenceClassification.from_pretrained(model_id)
```
You can find more examples in the [documentation](https://huggingface.co/docs/optimum/intel/optimization_inc) and in the [examples](https://github.com/huggingface/optimum-intel/tree/main/examples/neural_compressor).
### ONNX + ONNX Runtime
Before you begin, make sure you have all the necessary libraries installed :
```bash
pip install optimum[exporters,onnxruntime]
```
It is possible to export 🤗 Transformers and Diffusers models to the [ONNX](https://onnx.ai/) format and perform graph optimization as well as quantization easily:
```plain
optimum-cli export onnx -m deepset/roberta-base-squad2 --optimize O2 roberta_base_qa_onnx
```
The model can then be quantized using `onnxruntime`:
```bash
optimum-cli onnxruntime quantize \
--avx512 \
--onnx_model roberta_base_qa_onnx \
-o quantized_roberta_base_qa_onnx
```
These commands will export `deepset/roberta-base-squad2` and perform [O2 graph optimization](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/optimization#optimization-configuration) on the exported model, and finally quantize it with the [avx512 configuration](https://huggingface.co/docs/optimum/main/en/onnxruntime/package_reference/configuration#optimum.onnxruntime.AutoQuantizationConfig.avx512).
For more information on the ONNX export, please check the [documentation](https://huggingface.co/docs/optimum/exporters/onnx/usage_guides/export_a_model).
#### Run the exported model using ONNX Runtime
Once the model is exported to the ONNX format, we provide Python classes enabling you to run the exported ONNX model in a seemless manner using [ONNX Runtime](https://onnxruntime.ai/) in the backend:
```diff
- from transformers import AutoModelForQuestionAnswering
+ from optimum.onnxruntime import ORTModelForQuestionAnswering
from transformers import AutoTokenizer, pipeline
model_id = "deepset/roberta-base-squad2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
- model = AutoModelForQuestionAnswering.from_pretrained(model_id)
+ model = ORTModelForQuestionAnswering.from_pretrained("roberta_base_qa_onnx")
qa_pipe = pipeline("question-answering", model=model, tokenizer=tokenizer)
question = "What's Optimum?"
context = "Optimum is an awesome library everyone should use!"
results = qa_pipe(question=question, context=context)
```
More details on how to run ONNX models with `ORTModelForXXX` classes [here](https://huggingface.co/docs/optimum/main/en/onnxruntime/usage_guides/models).
### TensorFlow Lite
Before you begin, make sure you have all the necessary libraries installed :
```bash
pip install optimum[exporters-tf]
```
Just as for ONNX, it is possible to export models to [TensorFlow Lite](https://www.tensorflow.org/lite) and quantize them:
```plain
optimum-cli export tflite \
-m deepset/roberta-base-squad2 \
--sequence_length 384 \
--quantize int8-dynamic roberta_tflite_model
```
## Accelerated training
🤗 Optimum provides wrappers around the original 🤗 Transformers [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) to enable training on powerful hardware easily.
We support many providers:
- Habana's Gaudi processors
- ONNX Runtime (optimized for GPUs)
### Habana
Before you begin, make sure you have all the necessary libraries installed :
```bash
pip install --upgrade-strategy eager optimum[habana]
```
```diff
- from transformers import Trainer, TrainingArguments
+ from optimum.habana import GaudiTrainer, GaudiTrainingArguments
# Download a pretrained model from the Hub
model = AutoModelForXxx.from_pretrained("bert-base-uncased")
# Define the training arguments
- training_args = TrainingArguments(
+ training_args = GaudiTrainingArguments(
output_dir="path/to/save/folder/",
+ use_habana=True,
+ use_lazy_mode=True,
+ gaudi_config_name="Habana/bert-base-uncased",
...
)
# Initialize the trainer
- trainer = Trainer(
+ trainer = GaudiTrainer(
model=model,
args=training_args,
train_dataset=train_dataset,
...
)
# Use Habana Gaudi processor for training!
trainer.train()
```
You can find more examples in the [documentation](https://huggingface.co/docs/optimum/habana/quickstart) and in the [examples](https://github.com/huggingface/optimum-habana/tree/main/examples).
### ONNX Runtime
```diff
- from transformers import Trainer, TrainingArguments
+ from optimum.onnxruntime import ORTTrainer, ORTTrainingArguments
# Download a pretrained model from the Hub
model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased")
# Define the training arguments
- training_args = TrainingArguments(
+ training_args = ORTTrainingArguments(
output_dir="path/to/save/folder/",
optim="adamw_ort_fused",
...
)
# Create a ONNX Runtime Trainer
- trainer = Trainer(
+ trainer = ORTTrainer(
model=model,
args=training_args,
train_dataset=train_dataset,
...
)
# Use ONNX Runtime for training!
trainer.train()
```
You can find more examples in the [documentation](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/trainer) and in the [examples](https://github.com/huggingface/optimum/tree/main/examples/onnxruntime/training).
| huggingface/optimum/blob/main/README.md |
--
title: "Introducing SafeCoder"
thumbnail: /blog/assets/159_safecoder/thumbnail.jpg
authors:
- user: jeffboudier
- user: philschmid
---
# Introducing SafeCoder
Today we are excited to announce SafeCoder - a code assistant solution built for the enterprise.
The goal of SafeCoder is to unlock software development productivity for the enterprise, with a fully compliant and self-hosted pair programmer. In marketing speak: “your own on-prem GitHub copilot”.
Before we dive deeper, here’s what you need to know:
- SafeCoder is not a model, but a complete end-to-end commercial solution
- SafeCoder is built with security and privacy as core principles - code never leaves the VPC during training or inference
- SafeCoder is designed for self-hosting by the customer on their own infrastructure
- SafeCoder is designed for customers to own their own Code Large Language Model
## Why SafeCoder?
Code assistant solutions built upon LLMs, such as GitHub Copilot, are delivering strong [productivity boosts](https://github.blog/2022-09-07-research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness/). For the enterprise, the ability to tune Code LLMs on the company code base to create proprietary Code LLMs improves reliability and relevance of completions to create another level of productivity boost. For instance, Google internal LLM code assistant reports a completion [acceptance rate of 25-34%](https://ai.googleblog.com/2022/07/ml-enhanced-code-completion-improves.html) by being trained on an internal code base.
However, relying on closed-source Code LLMs to create internal code assistants exposes companies to compliance and security issues. First during training, as fine-tuning a closed-source Code LLM on an internal codebase requires exposing this codebase to a third party. And then during inference, as fine-tuned Code LLMs are likely to “leak” code from their training dataset during inference. To meet compliance requirements, enterprises need to deploy fine-tuned Code LLMs within their own infrastructure - which is not possible with closed source LLMs.
With SafeCoder, Hugging Face will help customers build their own Code LLMs, fine-tuned on their proprietary codebase, using state of the art open models and libraries, without sharing their code with Hugging Face or any other third party. With SafeCoder, Hugging Face delivers a containerized, hardware-accelerated Code LLM inference solution, to be deployed by the customer directly within the Customer secure infrastructure, without code inputs and completions leaving their secure IT environment.
## From StarCoder to SafeCoder
At the core of the SafeCoder solution is the [StarCoder](https://huggingface.co/bigcode/starcoder) family of Code LLMs, created by the [BigCode](https://huggingface.co/bigcode) project, a collaboration between Hugging Face, ServiceNow and the open source community.
The StarCoder models offer unique characteristics ideally suited to enterprise self-hosted solution:
- State of the art code completion results - see benchmarks in the [paper](https://huggingface.co/papers/2305.06161) and [multilingual code evaluation leaderboard](https://huggingface.co/spaces/bigcode/multilingual-code-evals)
- Designed for inference performance: a 15B parameters model with code optimizations, Multi-Query Attention for reduced memory footprint, and Flash Attention to scale to 8,192 tokens context.
- Trained on [the Stack](https://huggingface.co/datasets/bigcode/the-stack), an ethically sourced, open source code dataset containing only commercially permissible licensed code, with a developer opt-out mechanism from the get-go, refined through intensive PII removal and deduplication efforts.
Note: While StarCoder is the inspiration and model powering the initial version of SafeCoder, an important benefit of building a LLM solution upon open source models is that it can adapt to the latest and greatest open source models available. In the future, SafeCoder may offer other similarly commercially permissible open source models built upon ethically sourced and transparent datasets as the base LLM available for fine-tuning.
## Privacy and Security as a Core Principle
For any company, the internal codebase is some of its most important and valuable intellectual property. A core principle of SafeCoder is that the customer internal codebase will never be accessible to any third party (including Hugging Face) during training or inference.
In the initial set up phase of SafeCoder, the Hugging Face team provides containers, scripts and examples to work hand in hand with the customer to select, extract, prepare, duplicate, deidentify internal codebase data into a training dataset to be used in a Hugging Face provided training container configured to the hardware infrastructure available to the customer.
In the deployment phase of SafeCoder, the customer deploys containers provided by Hugging Face on their own infrastructure to expose internal private endpoints within their VPC. These containers are configured to the exact hardware configuration available to the customer, including NVIDIA GPUs, AMD Instinct GPUs, Intel Xeon CPUs, AWS Inferentia2 or Habana Gaudi accelerators.
## Compliance as a Core Principle
As the regulation framework around machine learning models and datasets is still being written across the world, global companies need to make sure the solutions they use minimize legal risks.
Data sources, data governance, management of copyrighted data are just a few of the most important compliance areas to consider. BigScience, the older cousin and inspiration for BigCode, addressed these areas in working groups before they were broadly recognized by the draft AI EU Act, and as a result was [graded as most compliant among Foundational Model Providers in a Stanford CRFM study](https://crfm.stanford.edu/2023/06/15/eu-ai-act.html).
BigCode expanded upon this work by implementing novel techniques for the code domain and building The Stack with compliance as a core principle, such as commercially permissible license filtering, consent mechanisms (developers can [easily find out if their code is present and request to be opted out](https://huggingface.co/spaces/bigcode/in-the-stack) of the dataset), and extensive documentation and tools to inspect the [source data](https://huggingface.co/datasets/bigcode/the-stack-metadata), and dataset improvements (such as [deduplication](https://huggingface.co/blog/dedup) and [PII removal](https://huggingface.co/bigcode/starpii)).
All these efforts translate into legal risk minimization for users of the StarCoder models, and customers of SafeCoder. And for SafeCoder users, these efforts translate into compliance features: when software developers get code completions these suggestions are checked against The Stack, so users know if the suggested code matches existing code in the source dataset, and what the license is. Customers can specify which licenses are preferred and surface those preferences to their users.
## How does it work?
SafeCoder is a complete commercial solution, including service, software and support.
### Training your own SafeCoder model
StarCoder was trained in more than 80 programming languages and offers state of the art performance on [multiple benchmarks](https://huggingface.co/spaces/bigcode/multilingual-code-evals). To offer better code suggestions specifically for a SafeCoder customer, we start the engagement with an optional training phase, where the Hugging Face team works directly with the customer team to guide them through the steps to prepare and build a training code dataset, and to create their own code generation model through fine-tuning, without ever exposing their codebase to third parties or the internet.
The end result is a model that is adapted to the code languages, standards and practices of the customer. Through this process, SafeCoder customers learn the process and build a pipeline for creating and updating their own models, ensuring no vendor lock-in, and keeping control of their AI capabilities.
### Deploying SafeCoder
During the setup phase, SafeCoder customers and Hugging Face design and provision the optimal infrastructure to support the required concurrency to offer a great developer experience. Hugging Face then builds SafeCoder inference containers that are hardware-accelerated and optimized for throughput, to be deployed by the customer on their own infrastructure.
SafeCoder inference supports various hardware to give customers a wide range of options: NVIDIA Ampere GPUs, AMD Instinct GPUs, Habana Gaudi2, AWS Inferentia 2, Intel Xeon Sapphire Rapids CPUs and more.
### Using SafeCoder
Once SafeCoder is deployed and its endpoints are live within the customer VPC, developers can install compatible SafeCoder IDE plugins to get code suggestions as they work. Today, SafeCoder supports popular IDEs, including [VSCode](https://marketplace.visualstudio.com/items?itemName=HuggingFace.huggingface-vscode), IntelliJ and with more plugins coming from our partners.
## How can I get SafeCoder?
Today, we are announcing SafeCoder in collaboration with VMware at the VMware Explore conference and making SafeCoder available to VMware enterprise customers. Working with VMware helps ensure the deployment of SafeCoder on customers’ VMware Cloud infrastructure is successful – whichever cloud, on-premises or hybrid infrastructure scenario is preferred by the customer. In addition to utilizing SafeCoder, VMware has published a [reference architecture](https://www.vmware.com/content/dam/digitalmarketing/vmware/en/pdf/docs/vmware-baseline-reference-architecture-for-generative-ai.pdf) with code samples to enable the fastest possible time-to-value when deploying and operating SafeCoder on VMware infrastructure. VMware’s Private AI Reference Architecture makes it easy for organizations to quickly leverage popular open source projects such as ray and kubeflow to deploy AI services adjacent to their private datasets, while working with Hugging Face to ensure that organizations maintain the flexibility to take advantage of the latest and greatest in open-source models. This is all without tradeoffs in total cost of ownership or performance.
“Our collaboration with Hugging Face around SafeCoder fully aligns to VMware’s goal of enabling customer choice of solutions while maintaining privacy and control of their business data. In fact, we have been running SafeCoder internally for months and have seen excellent results. Best of all, our collaboration with Hugging Face is just getting started, and I’m excited to take our solution to our hundreds of thousands of customers worldwide,” says Chris Wolf, Vice President of VMware AI Labs. Learn more about private AI and VMware’s differentiation in this emerging space [here](https://octo.vmware.com/vmware-private-ai-foundation/).
---
If you’re interested in SafeCoder for your company, please contact us [here](mailto:[email protected]?subject=SafeCoder) - our team will contact you to discuss your requirements!
| huggingface/blog/blob/main/safecoder.md |
Gradio Demo: dataset_component
```
!pip install -q gradio
```
```
import gradio as gr
with gr.Blocks() as demo:
gr.Dataset(components=[gr.Textbox(visible=False)],
label="Text Dataset",
samples=[
["The quick brown fox jumps over the lazy dog"],
["Build & share delightful machine learning apps"],
["She sells seashells by the seashore"],
["Supercalifragilisticexpialidocious"],
["Lorem ipsum"],
["That's all folks!"]
],
)
demo.launch()
```
| gradio-app/gradio/blob/main/demo/dataset_component/run.ipynb |
!---
Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
-->
# Model parallel language model training example
The following example showcases how to train/fine-tune GPTNeo model with model parallelism using
the JAX/Flax backend and the [`pjit`](https://jax.readthedocs.io/en/latest/jax.experimental.pjit.html) transformation.
> Note: The example is experimental and might have bugs. Also currently it only supports single V3-8.
The `partition.py` file defines the `PyTree` of `ParitionSpec` for the GPTNeo model which describes how the model will be sharded.
The actual sharding is auto-matically handled by `pjit`. The weights are sharded across all local devices.
To adapt the script for other models, we need to also change the `ParitionSpec` accordingly.
TODO: Add more explantion.
Before training, let's prepare our model first. To be able to shard the model, the sharded dimention needs to be a multiple of devices it'll be sharded on. But GPTNeo's vocab size is 50257, so we need to resize the embeddings accordingly.
```python
from transformers import FlaxGPTNeoForCausalLM, GPTNeoConfig
model = FlaxGPTNeoForCausalLM.from_pretrained("EleutherAI/gpt-neo-1.3B")
emb = jnp.zeros((50264, model.config.hidden_size))
# update the first 50257 weights using pre-trained weights
emb = emb.at[:50257, :].set(model.params["transformer"]["wte"]["embedding"])
params = model.params
params["transformer"]["wte"]["embedding"] = emb
# initialize a random model with the right vocab_size
config = GPTNeoConfig.from_pretrained("EleutherAI/gpt-neo-1.3B", vocab_size=50264)
model = FlaxGPTNeoForCausalLM(config)
# assign the pre-trained weights and save the model.
model.params = params
model.save_pretrained("gpt-neo-1.3B")
```
### Train Model
```bash
python run_clm_mp.py \
--model_name_or_path gpt-neo-1.3B \
--tokenizer_name gpt2 \
--dataset_name wikitext --dataset_config_name wikitext-2-raw-v1 \
--do_train --do_eval \
--block_size 1024 \
--num_train_epochs 5 \
--learning_rate 4e-6 \
--per_device_train_batch_size 3 --per_device_eval_batch_size 3 \
--overwrite_output_dir --output_dir ~/tmp/flax-clm \
--cache_dir ~/datasets_cache/wikitext --dtype bfloat16 \
--logging_steps 96 --eval_steps 96
``` | huggingface/transformers/blob/main/examples/research_projects/jax-projects/model_parallel/README.md |
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [yyyy] [name of copyright owner]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
| huggingface/deep-rl-class/blob/main/LICENSE.md |
Libraries
The Hub has support for dozens of libraries in the Open Source ecosystem. Thanks to the `huggingface_hub` Python library, it's easy to enable sharing your models on the Hub. The Hub supports many libraries, and we're working on expanding this support. We're happy to welcome to the Hub a set of Open Source libraries that are pushing Machine Learning forward.
The table below summarizes the supported libraries and their level of integration. Find all our supported libraries in [the model-libraries.ts file](https://github.com/huggingface/huggingface.js/blob/main/packages/tasks/src/model-libraries.ts).
| Library | Description | Inference API | Widgets | Download from Hub | Push to Hub |
|-----------------------------------------------------------------------------|--------------------------------------------------------------------------------------|---|---:|---|---|
| [Adapter Transformers](https://github.com/Adapter-Hub/adapter-transformers) | Extends 🤗Transformers with Adapters. | ❌ | ❌ | ✅ | ✅ |
| [AllenNLP](https://github.com/allenai/allennlp) | An open-source NLP research library, built on PyTorch. | ✅ | ✅ | ✅ | ❌ |
| [Asteroid](https://github.com/asteroid-team/asteroid) | Pytorch-based audio source separation toolkit | ✅ | ✅ | ✅ | ❌ |
| [BERTopic](https://github.com/MaartenGr/BERTopic) | BERTopic is a topic modeling library for text and images | ✅ | ✅ | ✅ | ✅ |
| [Diffusers](https://github.com/huggingface/diffusers) | A modular toolbox for inference and training of diffusion models | ✅ | ✅ | ✅ | ✅ |
| [docTR](https://github.com/mindee/doctr) | Models and datasets for OCR-related tasks in PyTorch & TensorFlow | ✅ | ✅ | ✅ | ❌ |
| [ESPnet](https://github.com/espnet/espnet) | End-to-end speech processing toolkit (e.g. TTS) | ✅ | ✅ | ✅ | ❌ |
| [fastai](https://github.com/fastai/fastai) | Library to train fast and accurate models with state-of-the-art outputs. | ✅ | ✅ | ✅ | ✅ |
| [Keras](https://huggingface.co/docs/hub/keras) | Library that uses a consistent and simple API to build models leveraging TensorFlow and its ecosystem. | ❌ | ❌ | ✅ | ✅ |
| [Flair](https://github.com/flairNLP/flair) | Very simple framework for state-of-the-art NLP. | ✅ | ✅ | ✅ | ✅ |
| [MBRL-Lib](https://github.com/facebookresearch/mbrl-lib) | PyTorch implementations of MBRL Algorithms. | ❌ | ❌ | ✅ | ✅ |
| [MidiTok](https://github.com/Natooz/MidiTok) | Tokenizers for symbolic music / MIDI files. | ❌ | ❌ | ✅ | ✅ |
| [ML-Agents](https://github.com/huggingface/ml-agents) | Enables games and simulations made with Unity to serve as environments for training intelligent agents. | ❌ | ❌ | ✅ | ✅ |
| [NeMo](https://github.com/NVIDIA/NeMo) | Conversational AI toolkit built for researchers | ✅ | ✅ | ✅ | ❌ |
| [OpenCLIP](https://github.com/mlfoundations/open_clip) | Library for open-source implementation of OpenAI's CLIP | ❌ | ❌ | ✅ | ✅ |
| [PaddleNLP](https://github.com/PaddlePaddle/PaddleNLP) | Easy-to-use and powerful NLP library built on PaddlePaddle | ✅ | ✅ | ✅ | ✅ |
| [PEFT](https://github.com/huggingface/peft) | Cutting-edge Parameter Efficient Fine-tuning Library | ✅ | ✅ | ✅ | ✅ |
| [Pyannote](https://github.com/pyannote/pyannote-audio) | Neural building blocks for speaker diarization. | ❌ | ❌ | ✅ | ❌ |
| [PyCTCDecode](https://github.com/kensho-technologies/pyctcdecode) | Language model supported CTC decoding for speech recognition | ❌ | ❌ | ✅ | ❌ |
| [Pythae](https://github.com/clementchadebec/benchmark_VAE) | Unified framework for Generative Autoencoders in Python | ❌ | ❌ | ✅ | ✅ |
| [RL-Baselines3-Zoo](https://github.com/DLR-RM/rl-baselines3-zoo) | Training framework for Reinforcement Learning, using [Stable Baselines3](https://github.com/DLR-RM/stable-baselines3).| ❌ | ✅ | ✅ | ✅ |
| [Sample Factory](https://github.com/alex-petrenko/sample-factory) | Codebase for high throughput asynchronous reinforcement learning. | ❌ | ✅ | ✅ | ✅ |
| [Sentence Transformers](https://github.com/UKPLab/sentence-transformers) | Compute dense vector representations for sentences, paragraphs, and images. | ✅ | ✅ | ✅ | ✅ |
| [SetFit](https://github.com/huggingface/setfit) | Efficient few-shot text classification with Sentence Transformers | ✅ | ✅ | ✅ | ✅ |
| [spaCy](https://github.com/explosion/spaCy) | Advanced Natural Language Processing in Python and Cython. | ✅ | ✅ | ✅ | ✅ |
| [SpanMarker](https://github.com/tomaarsen/SpanMarkerNER) | Familiar, simple and state-of-the-art Named Entity Recognition. | ✅ | ✅ | ✅ | ✅ |
| [Scikit Learn (using skops)](https://skops.readthedocs.io/en/stable/) | Machine Learning in Python. | ✅ | ✅ | ✅ | ✅ |
| [Speechbrain](https://speechbrain.github.io/) | A PyTorch Powered Speech Toolkit. | ✅ | ✅ | ✅ | ❌ |
| [Stable-Baselines3](https://github.com/DLR-RM/stable-baselines3) | Set of reliable implementations of deep reinforcement learning algorithms in PyTorch | ❌ | ✅ | ✅ | ✅ |
| [TensorFlowTTS](https://github.com/TensorSpeech/TensorFlowTTS) | Real-time state-of-the-art speech synthesis architectures. | ❌ | ❌ | ✅ | ❌ |
| [Timm](https://github.com/rwightman/pytorch-image-models) | Collection of image models, scripts, pretrained weights, etc. | ✅ | ✅ | ✅ | ✅ |
| [Transformers](https://github.com/huggingface/transformers) | State-of-the-art Natural Language Processing for Pytorch, TensorFlow, and JAX | ✅ | ✅ | ✅ | ✅ |
| [Transformers.js](https://github.com/xenova/transformers.js) | State-of-the-art Machine Learning for the web. Run 🤗 Transformers directly in your browser, with no need for a server! | ❌ | ❌ | ✅ | ❌ |
### How can I add a new library to the Inference API?
If you're interested in adding your library, please reach out to us! Read about it in [Adding a Library Guide](./models-adding-libraries).
| huggingface/hub-docs/blob/main/docs/hub/models-libraries.md |
Datasets server API - rows endpoint
> /rows endpoint
## Configuration
The service can be configured using environment variables. They are grouped by scope.
### API service
See [../../libs/libapi/README.md](../../libs/libapi/README.md) for more information about the API configuration.
### Common
See [../../libs/libcommon/README.md](../../libs/libcommon/README.md) for more information about the common configuration.
## Endpoints
See https://huggingface.co/docs/datasets-server
- /healthcheck: ensure the app is running
- /metrics: return a list of metrics in the Prometheus format
- /rows: get a slice of rows of a dataset split
| huggingface/datasets-server/blob/main/services/rows/README.md |
Optuna Tutorial [[optuna]]
The content below comes from [Antonin's Raffin ICRA 2022 presentations](https://araffin.github.io/tools-for-robotic-rl-icra2022/), he's one of the founders of Stable-Baselines and RL-Baselines3-Zoo.
## The theory behind Hyperparameter tuning
<Youtube id="AidFTOdGNFQ" />
## Optuna Tutorial
<Youtube id="ihP7E76KGOI" />
The notebook 👉 [here](https://colab.research.google.com/github/araffin/tools-for-robotic-rl-icra2022/blob/main/notebooks/optuna_lab.ipynb)
| huggingface/deep-rl-class/blob/main/units/en/unitbonus2/optuna.mdx |
Dual Path Network (DPN)
A **Dual Path Network (DPN)** is a convolutional neural network which presents a new topology of connection paths internally. The intuition is that [ResNets](https://paperswithcode.com/method/resnet) enables feature re-usage while DenseNet enables new feature exploration, and both are important for learning good representations. To enjoy the benefits from both path topologies, Dual Path Networks share common features while maintaining the flexibility to explore new features through dual path architectures.
The principal building block is an [DPN Block](https://paperswithcode.com/method/dpn-block).
## How do I use this model on an image?
To load a pretrained model:
```py
>>> import timm
>>> model = timm.create_model('dpn107', pretrained=True)
>>> model.eval()
```
To load and preprocess the image:
```py
>>> import urllib
>>> from PIL import Image
>>> from timm.data import resolve_data_config
>>> from timm.data.transforms_factory import create_transform
>>> config = resolve_data_config({}, model=model)
>>> transform = create_transform(**config)
>>> url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
>>> urllib.request.urlretrieve(url, filename)
>>> img = Image.open(filename).convert('RGB')
>>> tensor = transform(img).unsqueeze(0) # transform and add batch dimension
```
To get the model predictions:
```py
>>> import torch
>>> with torch.no_grad():
... out = model(tensor)
>>> probabilities = torch.nn.functional.softmax(out[0], dim=0)
>>> print(probabilities.shape)
>>> # prints: torch.Size([1000])
```
To get the top-5 predictions class names:
```py
>>> # Get imagenet class mappings
>>> url, filename = ("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt", "imagenet_classes.txt")
>>> urllib.request.urlretrieve(url, filename)
>>> with open("imagenet_classes.txt", "r") as f:
... categories = [s.strip() for s in f.readlines()]
>>> # Print top categories per image
>>> top5_prob, top5_catid = torch.topk(probabilities, 5)
>>> for i in range(top5_prob.size(0)):
... print(categories[top5_catid[i]], top5_prob[i].item())
>>> # prints class names and probabilities like:
>>> # [('Samoyed', 0.6425196528434753), ('Pomeranian', 0.04062102362513542), ('keeshond', 0.03186424449086189), ('white wolf', 0.01739676296710968), ('Eskimo dog', 0.011717947199940681)]
```
Replace the model name with the variant you want to use, e.g. `dpn107`. You can find the IDs in the model summaries at the top of this page.
To extract image features with this model, follow the [timm feature extraction examples](../feature_extraction), just change the name of the model you want to use.
## How do I finetune this model?
You can finetune any of the pre-trained models just by changing the classifier (the last layer).
```py
>>> model = timm.create_model('dpn107', pretrained=True, num_classes=NUM_FINETUNE_CLASSES)
```
To finetune on your own dataset, you have to write a training loop or adapt [timm's training
script](https://github.com/rwightman/pytorch-image-models/blob/master/train.py) to use your dataset.
## How do I train this model?
You can follow the [timm recipe scripts](../scripts) for training a new model afresh.
## Citation
```BibTeX
@misc{chen2017dual,
title={Dual Path Networks},
author={Yunpeng Chen and Jianan Li and Huaxin Xiao and Xiaojie Jin and Shuicheng Yan and Jiashi Feng},
year={2017},
eprint={1707.01629},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
<!--
Type: model-index
Collections:
- Name: DPN
Paper:
Title: Dual Path Networks
URL: https://paperswithcode.com/paper/dual-path-networks
Models:
- Name: dpn107
In Collection: DPN
Metadata:
FLOPs: 23524280296
Parameters: 86920000
File Size: 348612331
Architecture:
- Batch Normalization
- Convolution
- DPN Block
- Dense Connections
- Global Average Pooling
- Max Pooling
- Softmax
Tasks:
- Image Classification
Training Techniques:
- SGD with Momentum
- Weight Decay
Training Data:
- ImageNet
Training Resources: 40x K80 GPUs
ID: dpn107
LR: 0.316
Layers: 107
Crop Pct: '0.875'
Batch Size: 1280
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/dpn.py#L310
Weights: https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn107_extra-1ac7121e2.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 80.16%
Top 5 Accuracy: 94.91%
- Name: dpn131
In Collection: DPN
Metadata:
FLOPs: 20586274792
Parameters: 79250000
File Size: 318016207
Architecture:
- Batch Normalization
- Convolution
- DPN Block
- Dense Connections
- Global Average Pooling
- Max Pooling
- Softmax
Tasks:
- Image Classification
Training Techniques:
- SGD with Momentum
- Weight Decay
Training Data:
- ImageNet
Training Resources: 40x K80 GPUs
ID: dpn131
LR: 0.316
Layers: 131
Crop Pct: '0.875'
Batch Size: 960
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/dpn.py#L302
Weights: https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn131-71dfe43e0.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.83%
Top 5 Accuracy: 94.71%
- Name: dpn68
In Collection: DPN
Metadata:
FLOPs: 2990567880
Parameters: 12610000
File Size: 50761994
Architecture:
- Batch Normalization
- Convolution
- DPN Block
- Dense Connections
- Global Average Pooling
- Max Pooling
- Softmax
Tasks:
- Image Classification
Training Techniques:
- SGD with Momentum
- Weight Decay
Training Data:
- ImageNet
Training Resources: 40x K80 GPUs
ID: dpn68
LR: 0.316
Layers: 68
Crop Pct: '0.875'
Batch Size: 1280
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/dpn.py#L270
Weights: https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn68-66bebafa7.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 76.31%
Top 5 Accuracy: 92.97%
- Name: dpn68b
In Collection: DPN
Metadata:
FLOPs: 2990567880
Parameters: 12610000
File Size: 50781025
Architecture:
- Batch Normalization
- Convolution
- DPN Block
- Dense Connections
- Global Average Pooling
- Max Pooling
- Softmax
Tasks:
- Image Classification
Training Techniques:
- SGD with Momentum
- Weight Decay
Training Data:
- ImageNet
Training Resources: 40x K80 GPUs
ID: dpn68b
LR: 0.316
Layers: 68
Crop Pct: '0.875'
Batch Size: 1280
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/dpn.py#L278
Weights: https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights/dpn68b_ra-a31ca160.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.21%
Top 5 Accuracy: 94.42%
- Name: dpn92
In Collection: DPN
Metadata:
FLOPs: 8357659624
Parameters: 37670000
File Size: 151248422
Architecture:
- Batch Normalization
- Convolution
- DPN Block
- Dense Connections
- Global Average Pooling
- Max Pooling
- Softmax
Tasks:
- Image Classification
Training Techniques:
- SGD with Momentum
- Weight Decay
Training Data:
- ImageNet
Training Resources: 40x K80 GPUs
ID: dpn92
LR: 0.316
Layers: 92
Crop Pct: '0.875'
Batch Size: 1280
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/dpn.py#L286
Weights: https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn92_extra-b040e4a9b.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.99%
Top 5 Accuracy: 94.84%
- Name: dpn98
In Collection: DPN
Metadata:
FLOPs: 15003675112
Parameters: 61570000
File Size: 247021307
Architecture:
- Batch Normalization
- Convolution
- DPN Block
- Dense Connections
- Global Average Pooling
- Max Pooling
- Softmax
Tasks:
- Image Classification
Training Techniques:
- SGD with Momentum
- Weight Decay
Training Data:
- ImageNet
Training Resources: 40x K80 GPUs
ID: dpn98
LR: 0.4
Layers: 98
Crop Pct: '0.875'
Batch Size: 1280
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/dpn.py#L294
Weights: https://github.com/rwightman/pytorch-dpn-pretrained/releases/download/v0.1/dpn98-5b90dec4d.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.65%
Top 5 Accuracy: 94.61%
--> | huggingface/pytorch-image-models/blob/main/hfdocs/source/models/dpn.mdx |
--
title: "The Reformer - Pushing the limits of language modeling"
thumbnail: /blog/assets/03_reformer/thumbnail.png
authors:
- user: patrickvonplaten
---
# The Reformer - Pushing the limits of language modeling
<a href="https://colab.research.google.com/github/patrickvonplaten/blog/blob/main/notebooks/03_reformer.ipynb" target="_parent"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/></a>
## How the Reformer uses less than 8GB of RAM to train on sequences of half a million tokens
The Reformer model as introduced by [Kitaev, Kaiser et al. (2020)](https://arxiv.org/pdf/2001.04451.pdf) is one of the most memory-efficient transformer models for long sequence modeling as of today.
Recently, long sequence modeling has experienced a surge of interest as can be seen by the many submissions from this year alone - [Beltagy et al. (2020)](https://arxiv.org/abs/2004.05150), [Roy et al. (2020)](https://arxiv.org/abs/2003.05997), [Tay et al.](https://arxiv.org/abs/2002.11296), [Wang et al.](https://arxiv.org/abs/2006.04768) to name a few.
The motivation behind long sequence modeling is that many tasks in NLP, *e.g.* summarization, question answering, require the model to process longer input sequences than models, such as BERT, are able to handle. In tasks that require the model to process a large input sequence, long sequence models do not have to cut the input sequence to avoid memory overflow and thus have been shown to outperform standard "BERT"-like models *cf.* [Beltagy et al. (2020)](https://arxiv.org/abs/2004.05150).
The Reformer pushes the limit of longe sequence modeling by its ability to process up to half a million tokens at once as shown in this [demo](https://github.com/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb). As a comparison, a conventional `bert-base-uncased` model limits the input length to only 512 tokens. In Reformer, each part of the standard transformer architecture is re-engineered to optimize for minimal memory requirement without a significant drop in performance.
The memory improvements can be attributed to **4** features which the Reformer authors introduced to the transformer world:
1. **Reformer Self-Attention Layer** - *How to efficiently implement self-attention without being restricted to a local context?*
2. **Chunked Feed Forward Layers** - *How to get a better time-memory trade-off for large feed forward layers?*
3. **Reversible Residual Layers** - *How to drastically reduce memory consumption in training by a smart residual architecture?*
4. **Axial Positional Encodings** - *How to make positional encodings usable for extremely large input sequences?*
The goal of this blog post is to give the reader an **in-depth** understanding of each of the four Reformer features mentioned above. While the explanations are focussed on the Reformer, the reader should get a better intuition under which circumstances each of the four features can be effective for other transformer models as well.
The four sections are only loosely connected, so they can very well be read individually.
Reformer is part of the 🤗Transformers library. For all users of the Reformer, it is advised to go through this very detailed blog post to better understand how the model works and how to correctly set its configuration. All equations are accompanied by their equivalent name for the Reformer config, *e.g.* `config.<param_name>`, so that the reader can quickly relate to the official docs and configuration file.
**Note**: *Axial Positional Encodings* are not explained in the official Reformer paper, but are extensively used in the official codebase. This blog post gives the first in-depth explanation of Axial Positional Encodings.
## 1. Reformer Self-Attention Layer
Reformer uses two kinds of special self-attention layers: *local* self-attention layers and Locality Sensitive Hashing (*LSH*) self-attention layers.
To better introduce these new self-attention layers, we will briefly recap
conventional self-attention as introduced in [Vaswani et al. 2017](https://arxiv.org/abs/1706.03762).
This blog post uses the same notation and coloring as the popular blog post [The illustrated transformer](http://jalammar.github.io/illustrated-transformer/), so the reader is strongly advised to read this blog first.
**Important**: While Reformer was originally introduced for causal self-attention, it can very well be used for bi-directional self-attention as well. In this post, Reformer's self-attention is presented for *bidirectional* self-attention.
### Recap Global Self-Attention
The core of every Transformer model is the **self-attention** layer. To recap the conventional self-attention layer, which we refer to here as the **global self-attention** layer, let us assume we apply a transformer layer on the embedding vector sequence \\(\mathbf{X} = \mathbf{x}_1, \ldots, \mathbf{x}_n\\) where each vector \\(\mathbf{x}_{i}\\) is of size `config.hidden_size`, *i.e.* \\(d_h\\).
In short, a global self-attention layer projects \\(\mathbf{X}\\) to the query, key and value matrices \\(\mathbf{Q}, \mathbf{K}, \mathbf{V}\\) and computes the output \\(\mathbf{Z}\\) using the *softmax* operation as follows:
\\(\mathbf{Z} = \text{SelfAttn}(\mathbf{X}) = \text{softmax}(\mathbf{Q}\mathbf{K}^T) \mathbf{V}\\) with \\(\mathbf{Z}\\) being of dimension \\(d_h \times n\\) (leaving out the key normalization factor and self-attention weights \\(\mathbf{W}^{O}\\) for simplicity). For more detail on the complete transformer operation, see [the illustrated transformer](http://jalammar.github.io/illustrated-transformer/).
Visually, we can illustrate this operation as follows for \\(n=16, d_h=3\\):
Note that for all visualizations `batch_size` and `config.num_attention_heads` is assumed to be 1. Some vectors, *e.g.* \\(\mathbf{x_3}\\) and its corresponding output vector \\(\mathbf{z_3}\\) are marked so that *LSH self-attention* can later be better explained. The presented logic can effortlessly be extended for multi-head self-attention (`config.num_attention_{h}eads` > 1). The reader is advised to read [the illustrated transformer](http://jalammar.github.io/illustrated-transformer/) as a reference for multi-head self-attention.
Important to remember is that for each output vector \\(\mathbf{z}_{i}\\), the whole input sequence \\(\mathbf{X}\\) is processed. The tensor of the inner dot-product \\(\mathbf{Q}\mathbf{K}^T\\) has an asymptotic memory complexity of \\(\mathcal{O}(n^2)\\) which usually represents the memory bottleneck in a transformer model.
This is also the reason why `bert-base-cased` has a `config.max_position_embedding_size` of only 512.
### Local Self-Attention
**Local self-attention** is the obvious solution to reducing the \\(\mathcal{O}(n^2)\\) memory bottleneck, allowing us to model longer sequences with a reduced computational cost.
In local self-attention the input \\( \mathbf{X} = \mathbf{X}_{1:n} = \mathbf{x}_{1}, \ldots, \mathbf{x}_{n} \\)
is cut into \\(n_{c}\\) chunks: \\( \mathbf{X} = \left[\mathbf{X}_{1:l_{c}}, \ldots, \mathbf{X}_{(n_{c} - 1) * l_{c} : n_{c} * l_{c}}\right] \\) each
of length `config.local_chunk_length`, *i.e.* \\(l_{c}\\), and subsequently global self-attention is applied on each chunk separately.
Let's take our input sequence for \\(n=16, d_h=3\\) again for visualization:
Assuming \\(l_{c} = 4, n_{c} = 4\\), chunked attention can be illustrated as follows:
As can be seen, the attention operation is applied for each chunk \\(\mathbf{X}_{1:4}, \mathbf{X}_{5:8}, \mathbf{X}_{9:12}, \mathbf{X}_{13:16}\\) individually.
The first drawback of this architecture becomes obvious: Some input vectors have no access to their immediate context, *e.g.* \\(\mathbf{x}_9\\) has no access to \\(\mathbf{x}_{8}\\) and vice-versa in our example. This is problematic because these tokens are not able to learn word representations that take their immediate context into account.
A simple remedy is to augment each chunk with `config.local_num_chunks_before`, *i.e.* \\(n_{p}\\), chunks and `config.local_num_chunks_after`, *i.e.* \\(n_{a}\\), so that every input vector has at least access to \\(n_{p}\\) previous input vectors and \\(n_{a}\\) following input vectors. This can also be understood as chunking with overlap whereas \\(n_{p}\\) and \\(n_{a}\\) define the amount of overlap each chunk has with all previous chunks and following chunks. We denote this extended local self-attention as follows:
$$\mathbf{Z}^{\text{loc}} = \left[\mathbf{Z}_{1:l_{c}}^{\text{loc}}, \ldots, \mathbf{Z}_{(n_{c} - 1) * l_{c} : n_{c} * l_{c}}^{\text{loc}}\right], $$
with
$$\mathbf{Z}_{l_{c} * (i - 1) + 1 : l_{c} * i}^{\text{loc}} = \text{SelfAttn}(\mathbf{X}_{l_{c} * (i - 1 - n_{p}) + 1: l_{c} * (i + n_{a})})\left[n_{p} * l_{c}: -n_{a} * l_{c}\right], \forall i \in \{1, \ldots, n_{c} \}$$
Okay, this formula looks quite complicated. Let's make it easier.
In Reformer's self-attention layers \\(n_{a}\\) is usually set to 0 and \\(n_{p}\\) is set to 1, so let's write down the formula again for \\(i = 1\\):
$$\mathbf{Z}_{1:l_{c}}^{\text{loc}} = \text{SelfAttn}(\mathbf{X}_{-l_{c} + 1: l_{c}})\left[l_{c}:\right]$$
We notice that we have a circular relationship so that the first segment can attend the last segment as well. Let's illustrate this slightly enhanced local attention again. First, we apply self-attention within each windowed segment and keep only the central output segment.
Finally, the relevant output is concatenated to \\(\mathbf{Z}^{\text{loc}}\\) and looks as follows.
Note that local self-attention is implemented efficiently way so that no output is computed and subsequently "thrown-out" as shown here for illustration purposes by the red cross.
It's important to note here that extending the input vectors for each chunked self-attention function allows *each* single output vector \\( \mathbf{z}_{i} \\) of this self-attention function to learn better vector representations. E.g. each of the output vectors \\( \mathbf{z}_{5}^{\text{loc}}, \mathbf{z}_{6}^{\text{loc}}, \mathbf{z}_{7}^{\text{loc}}, \mathbf{z}_{8}^{\text{loc}} \\) can take into account all of the input vectors \\( \mathbf{X}_{1:8} \\) to learn better representations.
The gain in memory consumption is quite obvious: The \\( \mathcal{O}(n^2) \\) memory complexity is broken down for each segment individually so that the total asymptotic memory consumption is reduced to \\( \mathcal{O}(n_{c} * l_{c}^2) = \mathcal{O}(n * l_{c}) \\).
This enhanced local self-attention is better than the vanilla local self-attention architecture but still has a major drawback in that every input vector can only attend to a local context of predefined size. For NLP tasks that do not require the transformer model to learn long-range dependencies between the input vectors, which include arguably *e.g.* speech recognition, named entity recognition and causal language modeling of short sentences, this might not be a big issue. Many NLP tasks do require the model to learn long-range dependencies, so that local self-attention could lead to significant performance degradation, *e.g.*
* *Question-answering*: the model has to learn the relationship between the question tokens and relevant answer tokens which will most likely not be in the same local range
* *Multiple-Choice*: the model has to compare multiple answer token segments to each other which are usually separated by a significant length
* *Summarization*: the model has to learn the relationship between a long sequence of context tokens and a shorter sequence of summary tokens, whereas the relevant relationships between context and summary can most likely not be captured by local self-attention
* etc...
Local self-attention on its own is most likely not sufficient for the transformer model to learn the relevant relationships of input vectors (tokens) to each other.
Therefore, Reformer additionally employs an efficient self-attention layer that approximates global self-attention, called *LSH self-attention*.
### LSH Self-Attention
Alright, now that we have understood how local self-attention works, we can take a stab at the probably most innovative piece of Reformer: **Locality sensitive hashing (LSH) Self-Attention**.
The premise of LSH self-attention is to be more or less as efficient as local self-attention while approximating global self-attention.
LSH self-attention relies on the LSH algorithm as presented in [Andoni et al (2015)](https://arxiv.org/abs/1509.02897), hence its name.
The idea behind LSH self-attention is based on the insight that if \\(n\\) is large, the softmax applied on the \\(\mathbf{Q}\mathbf{K}^T\\) attention dot-product weights only very few value vectors with values significantly larger than 0 for each query vector.
Let's explain this in more detail.
Let \\(\mathbf{k}_{i} \in \mathbf{K} = \left[\mathbf{k}_1, \ldots, \mathbf{k}_n \right]^T\\) and \\(\mathbf{q}_{i} \in \mathbf{Q} = \left[\mathbf{q}_1, \ldots, \mathbf{q}_n\right]^T\\) be the key and query vectors. For each \\(\mathbf{q}_{i}\\), the computation \\(\text{softmax}(\mathbf{q}_{i}^T \mathbf{K}^T)\\) can be approximated by using only those key vectors of \\(\mathbf{k}_{j}\\) that have a high cosine similarity with \\(\mathbf{q}_{i}\\). This owes to the fact that the softmax function puts exponentially more weight on larger input values.
So far so good, the next problem is to efficiently find the vectors that have a
high cosine similarity with \\(\mathbf{q}_{i}\\) for all \\(i\\).
First, the authors of Reformer notice that sharing the query and key projections: \\(\mathbf{Q} = \mathbf{K}\\) does not impact the performance of a transformer model \\({}^1\\). Now, instead of having to find the key vectors of high cosine similarity for each query vector \\(q_i\\), only the cosine similarity of query vectors to each other has to be found.
This is important because there is a transitive property to the query-query vector dot product approximation: If \\(\mathbf{q}_{i}\\) has a high cosine similarity to the query vectors \\(\mathbf{q}_{j}\\) and \\(\mathbf{q}_{k}\\), then \\(\mathbf{q}_{j}\\) also has a high cosine similarity to \\(\mathbf{q}_{k}\\). Therefore, the query vectors can be clustered into buckets, such that all query vectors that belong to the same bucket have a high cosine similarity to each other. Let's define \\(C_{m}\\) as the *mth* set of position indices, such that their corresponding query vectors are in the same bucket: \\(C_{m} = \{ i | \text{ s.t. } \mathbf{q}_{i} \in \text{mth cluster}\}\\) and `config.num_buckets`, *i.e.* \\(n_{b}\\), as the number of buckets.
For each set of indices \\(C_{m}\\), the softmax function on the corresponding bucket of query vectors \\(\text{softmax}(\mathbf{Q}_{i \in C_{m}} \mathbf{Q}^T_{i \in C_{m}})\\) approximates the softmax function of global self-attention with shared query and key projections \\(\text{softmax}(\mathbf{q}_{i}^T \mathbf{Q}^T)\\) for all position indices \\(i\\) in \\(C_{m}\\).
Second, the authors make use of the **LSH** algorithm to cluster the query vectors into a predefined number of buckets \\(n_{b}\\). The LSH algorithm is an ideal choice here because it is very efficient and is an approximation of the nearest neighbor algorithm for cosine similarity. Explaining the LSH scheme is out-of-scope for this notebook, so let's just keep in mind that for each vector \\(\mathbf{q}_{i}\\) the LSH algorithm attributes its position index \\(i\\) to one of \\(n_{b}\\) predefined buckets, *i.e.* \\(\text{LSH}(\mathbf{q}_{i}) = m\\) with \\(i \in \{1, \ldots, n\}\\) and \\(m \in \{1, \ldots, n_{b}\}\\).
Visually, we can illustrate this as follows for our original example:
Third, it can be noted that having clustered all query vectors in \\(n_{b}\\) buckets, the corresponding set of indices \\(C_{m}\\) can be used to permute the input vectors \\(\mathbf{x}_1, \ldots, \mathbf{x}_n\\) accordingly \\({}^2\\) so that shared query-key self-attention can be applied piecewise similar to local attention.
Let's clarify with our example input vectors \\(\mathbf{X} = \mathbf{x}_1, ..., \mathbf{x}_{16}\\) and assume `config.num_buckets=4` and `config.lsh_chunk_length = 4`. Looking at the graphic above we can see that we have assigned each query vector \\( \mathbf{q}_1, \ldots, \mathbf{q}_{16} \\) to one of the clusters \\( \mathcal{C}_{1}, \mathcal{C}_{2}, \mathcal{C}_{3}, \mathcal{C}_{4} \\) .
If we now sort the corresponding input vectors \\( \mathbf{x}_1, \ldots, \mathbf{x}_{16} \\) accordingly, we get the following permuted input \\( \mathbf{X'} \\):
The self-attention mechanism should be applied for each cluster individually so that for each cluster \\( \mathcal{C}_m \\) the corresponding output is calculated as follows: \\( \mathbf{Z}^{\text{LSH}}_{i \in \mathcal{C}_m} = \text{SelfAttn}_{\mathbf{Q}=\mathbf{K}}(\mathbf{X}_{i \in \mathcal{C}_m}) \\).
Let's illustrate this again for our example.
As can be seen, the self-attention function operates on different sizes of matrices, which is suboptimal for efficient batching in GPU and TPU.
To overcome this problem, the permuted input can be chunked the same way it is done for local attention so that each chunk is of size `config.lsh_chunk_length`. By chunking the permuted input, a bucket might be split into two different chunks. To remedy this problem, in LSH self-attention each chunk attends to its previous chunk `config.lsh_num_chunks_before=1` in addition to itself, the same way local self-attention does (`config.lsh_num_chunks_after` is usually set to 0). This way, we can be assured that all vectors in a bucket attend to each other with a high probability \\({}^3\\).
All in all for all chunks \\( k \in \{1, \ldots, n_{c}\} \\), LSH self-attention can be noted down as follows:
$$ \mathbf{Z'}_{l_{c} * k + 1:l_{c} * (k + 1)}^{\text{LSH}} = \text{SelfAttn}_{\mathbf{Q} = \mathbf{K}}(\mathbf{X'}_{l_{c} * k + 1): l_{c} * (k + 1)})\left[l_{c}:\right] $$
with \\(\mathbf{X'}\\) and \\( \mathbf{Z'} \\) being the input and output vectors permuted according to the LSH algorithm.
Enough complicated formulas, let's illustrate LSH self-attention.
The permuted vectors \\(\mathbf{X'}\\) as shown above are chunked and shared query key self-attention is applied to each chunk.
Finally, the output \\(\mathbf{Z'}^{\text{LSH}}\\) is reordered to its original permutation.
One important feature to mention here as well is that the accuracy of LSH self-attention can be improved by running LSH self-attention `config.num_hashes`, e.g. \\(n_{h} \\) times in parallel, each with a different random LSH hash.
By setting `config.num_hashes > 1`, for each output position \\( i \\), multiple output vectors \\( \mathbf{z}^{\text{LSH}, 1}_{i}, \ldots, \mathbf{z}^{\text{LSH}, n_{h}}_{i} \\) are computed
and subsequently merged: \\( \mathbf{z}^{\text{LSH}}_{i} = \sum_k^{n_{h}} \mathbf{Z}^{\text{LSH}, k}_{i} * \text{weight}^k_i \\). The \\( \text{weight}^k_i \\) represents the importance of the output vectors \\( \mathbf{z}^{\text{LSH}, k}_{i} \\) of hashing round \\( k \\) in comparison to the other hashing rounds, and is exponentially proportional to the normalization term of their softmax computation. The intuition behind this is that if the corresponding query vector \\( \mathbf{q}_{i}^{k} \\) have a high cosine similarity with all other query vectors in its respective chunk, then the softmax normalization term of this chunk tends to be high, so that the corresponding output vectors \\( \mathbf{q}_{i}^{k} \\) should be a better approximation to global attention and thus receive more weight than output vectors of hashing rounds with a lower softmax normalization term. For more detail see Appendix A of the [paper](https://arxiv.org/pdf/2001.04451.pdf). For our example, multi-round LSH self-attention can be illustrated as follows.
Great. That's it. Now we know how LSH self-attention works in Reformer.
Regarding the memory complexity, we now have two terms that compete which each other to be the memory bottleneck: the dot-product: \\( \mathcal{O}(n_{h} * n_{c} * l_{c}^2) = \mathcal{O}(n * n_{h} * l_{c}) \\) and the required memory for LSH bucketing: \\( \mathcal{O}(n * n_{h} * \frac{n_{b}}{2}) \\) with \\( l_{c} \\) being the chunk length. Because for large \\( n \\), the number of buckets \\( \frac{n_{b}}{2} \\) grows much faster than the chunk length \\( l_{c} \\), the user can again factorize the number of buckets `config.num_buckets` as explained [here](https://huggingface.co/transformers/model_doc/reformer.html#lsh-self-attention).
Let's recap quickly what we have gone through above:
1. We want to approximate global attention using the knowledge that the softmax operation only puts significant weights on very few key vectors.
2. If key vectors are equal to query vectors this means that *for each* query vector \\( \mathbf{q}_{i} \\), the softmax only puts significant weight on other query vectors that are similar in terms of cosine similarity.
3. This relationship works in both ways, meaning if \\( \mathbf{q}_{j} \\) is similar to \\( \mathbf{q}_{i} \\) than \\(\mathbf{q}_{j} \\) is also similar to \\( \mathbf{q}_{i} \\), so that we can do a global clustering before applying self-attention on a permuted input.
4. We apply local self-attention on the permuted input and re-order the output to its original permutation.
---
\\( {}^{1} \\) The authors run some preliminary experiments confirming that shared query key self-attention performs more or less as well as standard self-attention.
\\( {}^{2} \\) To be more exact the query vectors within a bucket are sorted according to their original order. This means if, *e.g.* the vectors \\( \mathbf{q}_1, \mathbf{q}_3, \mathbf{q}_7 \\) are all hashed to bucket 2, the order of the vectors in bucket 2 would still be \\( \mathbf{q}_1 \\), followed by \\( \mathbf{q}_3 \\) and \\( \mathbf{q}_7 \\).
\\( {}^3 \\) On a side note, it is to mention the authors put a mask on the query vector \\( \mathbf{q}_{i} \\) to prevent the vector from attending to itself. Because the cosine similarity of a vector to itself will always be as high or higher than the cosine similarity to other vectors, the query vectors in shared query key self-attention are strongly discouraged to attend to themselves.
### Benchmark
Benchmark tools were recently added to Transformers - see [here](https://github.com/huggingface/transformers/blob/master/notebooks/05-benchmark.ipynb) for a more detailed explanation.
To show how much memory can be saved using "local" + "LSH" self-attention, the Reformer model `google/reformer-enwik8` is benchmarked for different `local_attn_chunk_length` and `lsh_attn_chunk_length`. The default configuration and usage of the `google/reformer-enwik8` model can be checked in more detail [here](https://huggingface.co/google/reformer-enwik8).
Let's first do some necessary imports and installs.
```
#@title Installs and Imports
# pip installs
!pip -qq install git+https://github.com/huggingface/transformers.git
!pip install -qq py3nvml
from transformers import ReformerConfig, PyTorchBenchmark, PyTorchBenchmarkArguments
```
First, let's benchmark the memory usage of the Reformer model using *global* self-attention. This can be achieved by setting `lsh_attn_chunk_length` = `local_attn_chunk_length` = 8192 so that for all input sequences smaller or equal to 8192, the model automatically switches to global self-attention.
```
config = ReformerConfig.from_pretrained("google/reformer-enwik8", lsh_attn_chunk_length=16386, local_attn_chunk_length=16386, lsh_num_chunks_before=0, local_num_chunks_before=0)
benchmark_args = PyTorchBenchmarkArguments(sequence_lengths=[2048, 4096, 8192, 16386], batch_sizes=[1], models=["Reformer"], no_speed=True, no_env_print=True)
benchmark = PyTorchBenchmark(configs=[config], args=benchmark_args)
result = benchmark.run()
```
HBox(children=(FloatProgress(value=0.0, description='Downloading', max=1279.0, style=ProgressStyle(description…
1 / 1
Doesn't fit on GPU. CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 11.17 GiB total capacity; 8.87 GiB already allocated; 1.92 GiB free; 8.88 GiB reserved in total by PyTorch)
==================== INFERENCE - MEMORY - RESULT ====================
--------------------------------------------------------------------------------
Model Name Batch Size Seq Length Memory in MB
--------------------------------------------------------------------------------
Reformer 1 2048 1465
Reformer 1 4096 2757
Reformer 1 8192 7893
Reformer 1 16386 N/A
--------------------------------------------------------------------------------
The longer the input sequence, the more visible is the quadratic relationship \\( \mathcal{O}(n^2) \\) between input sequence and peak memory usage. As can be seen, in practice it would require a much longer input sequence to clearly observe that doubling the input sequence quadruples the peak memory usage.
For this a `google/reformer-enwik8` model using global attention, a sequence length of over 16K results in a memory overflow.
Now, let's activate *local* and *LSH* self-attention by using the model's default parameters.
```
config = ReformerConfig.from_pretrained("google/reformer-enwik8")
benchmark_args = PyTorchBenchmarkArguments(sequence_lengths=[2048, 4096, 8192, 16384, 32768, 65436], batch_sizes=[1], models=["Reformer"], no_speed=True, no_env_print=True)
benchmark = PyTorchBenchmark(configs=[config], args=benchmark_args)
result = benchmark.run()
```
1 / 1
Doesn't fit on GPU. CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 11.17 GiB total capacity; 7.85 GiB already allocated; 1.74 GiB free; 9.06 GiB reserved in total by PyTorch)
Doesn't fit on GPU. CUDA out of memory. Tried to allocate 4.00 GiB (GPU 0; 11.17 GiB total capacity; 6.56 GiB already allocated; 3.99 GiB free; 6.81 GiB reserved in total by PyTorch)
==================== INFERENCE - MEMORY - RESULT ====================
--------------------------------------------------------------------------------
Model Name Batch Size Seq Length Memory in MB
--------------------------------------------------------------------------------
Reformer 1 2048 1785
Reformer 1 4096 2621
Reformer 1 8192 4281
Reformer 1 16384 7607
Reformer 1 32768 N/A
Reformer 1 65436 N/A
--------------------------------------------------------------------------------
As expected using local and LSH self-attention is much more memory efficient for longer input sequences, so that the model runs out of memory only at 16K tokens for a 11GB RAM GPU in this notebook.
## 2. Chunked Feed Forward Layers
Transformer-based models often employ very large feed forward layers after the self-attention layer in parallel. Thereby, this layer can take up a significant amount of the overall memory and sometimes even represent the memory bottleneck of a model.
First introduced in the Reformer paper, feed forward chunking is a technique that allows to effectively trade better memory consumption for increased time consumption.
### Chunked Feed Forward Layer in Reformer
In Reformer, the _LSH_- or _local_ self-attention layer is usually followed by a residual connection, which then defines the first part in a *transformer block*. For more detail on this please refer to this [blog](http://jalammar.github.io/illustrated-transformer/).
The output of the first part of the *transformer block*, called *normed self-attention* output can be written as \\( \mathbf{\overline{Z}} = \mathbf{Z} + \mathbf{X} \\), with \\( \mathbf{Z} \\) being either \\( \mathbf{Z}^{\text{LSH}} \\) or \\( \mathbf{Z}^\text{loc} \\) in Reformer.
For our example input \\( \mathbf{x}_1, \ldots, \mathbf{x}_{16} \\), we illustrate the normed self-attention output as follows.
Now, the second part of a *transformer block* usually consists of two feed forward layers \\( ^{1} \\), defined as \\( \text{Linear}_{\text{int}}(\ldots) \\) that processes \\( \mathbf{\overline{Z}} \\), to an intermediate output \\( \mathbf{Y}_{\text{int}} \\) and \\( \text{Linear}_{\text{out}}(\ldots) \\) that processes the intermediate output to the output \\( \mathbf{Y}_{\text{out}} \\). The two feed forward layers can be defined by
$$\mathbf{Y}_{\text{out}} = \text{Linear}_{\text{out}}(\mathbf{Y}_\text{int}) =
\text{Linear}_{\text{out}}(\text{Linear}_{\text{int}}(\mathbf{\overline{Z}})).$$
It is important to remember at this point that mathematically the output of a feed forward layer at position \\( \mathbf{y}_{\text{out}, i} \\) only depends on the input at this position \\( \mathbf{\overline{y}}_{i} \\). In contrast to the self-attention layer, every output \\( \mathbf{y}_{\text{out}, i} \\) is therefore completely independent of all inputs \\( \mathbf{\overline{y}}_{j \ne i} \\) of different positions.
Let's illustrate the feed forward layers for \\( \mathbf{\overline{z}}_1, \ldots, \mathbf{\overline{z}}_{16} \\).
As can be depicted from the illustration, all input vectors \\( \mathbf{\overline{z}}_{i} \\) are processed by the same feed forward layer in parallel.
It becomes interesting when one takes a look at the output dimensions of the feed forward layers. In Reformer, the output dimension of \\( \text{Linear}_{\text{int}} \\) is defined as `config.feed_forward_size`, *e.g.* \\( d_{f} \\), and the output dimension of \\( \text{Linear}_{\text{out}} \\) is defined as `config.hidden_size`, *i.e.* \\( d_{h} \\).
The Reformer authors observed that in a transformer model the intermediate dimension \\( d_{f} \\) usually tends to be much larger than the output dimension \\(^{2}\\) \\( d_{h} \\). This means that the tensor \\( \mathbf{\mathbf{Y}}_\text{int} \\) of dimension \\( d_{f} \times n \\) allocates a significant amount of the total memory and can even become the memory bottleneck.
To get a better feeling for the differences in dimensions let's picture the matrices \\( \mathbf{Y}_\text{int} \\) and \\( \mathbf{Y}_\text{out} \\) for our example.
It is becoming quite obvious that the tensor \\( \mathbf{Y}_\text{int} \\) holds much more memory ( \\( \frac{d_{f}}{d_{h}} \times n \\) as much to be exact) than \\( \mathbf{Y}_{\text{out}} \\). But, is it even necessary to compute the full intermediate matrix \\( \mathbf{Y}_\text{int} \\) ? Not really, because relevant is only the output matrix \\( \mathbf{Y}_\text{out} \\).
To trade memory for speed, one can thus chunk the linear layers computation to only process one chunk at the time. Defining `config.chunk_size_feed_forward` as \\( c_{f} \\), chunked linear layers are defined as \\( \mathbf{Y}_{\text{out}} = \left[\mathbf{Y}_{\text{out}, 1: c_{f}}, \ldots, \mathbf{Y}_{\text{out}, (n - c_{f}): n}\right] \\) with \\( \mathbf{Y}_{\text{out}, (c_{f} * i): (i * c_{f} + i)} = \text{Linear}_{\text{out}}(\text{Linear}_{\text{int}}(\mathbf{\overline{Z}}_{(c_{f} * i): (i * c_{f} + i)})) \\).
In practice, it just means that the output is incrementally computed and concatenated to avoid having to store the whole intermediate tensor \\( \mathbf{Y}_{\text{int}} \\) in memory.
Assuming \\( c_{f}=1 \\) for our example we can illustrate the incremental computation of the output for position \\( i=9 \\) as follows.
By processing the inputs in chunks of size 1, the only tensors that have to be stored in memory at the same time are \\( \mathbf{Y}_\text{out} \\) of a maximum size of \\( 16 \times d_{h} \\), \\( \mathbf{y}_{\text{int}, i} \\) of size \\( d_{f} \\) and the input \\( \mathbf{\overline{Z}} \\) of size \\( 16 \times d_{h} \\), with \\( d_{h} \\) being `config.hidden_size` \\(^{3}\\).
Finally, it is important to remember that *chunked linear layers* yield a mathematically equivalent output to conventional linear layers and can therefore be applied to all transformer linear layers. Making use of `config.chunk_size_feed_forward` therefore allows a better trade-off between memory and speed in certain use cases.
---
\\( {}^1 \\) For a simpler explanation, the layer norm layer which is normally applied to \\( \mathbf{\overline{Z}} \\) before being processed by the feed forward layers is omitted for now.
\\( {}^2 \\) In `bert-base-uncased`, *e.g.* the intermediate dimension \\( d_{f} \\) is with 3072 four times larger than the output dimension \\( d_{h} \\).
\\( {}^3 \\) As a reminder, the output `config.num_attention_heads` is assumed to be 1 for the sake of clarity and illustration in this notebook, so that the output of the self-attention layers can be assumed to be of size `config.hidden_size`.
More information on chunked linear / feed forward layers can also be found [here](https://huggingface.co/transformers/glossary.html#feed-forward-chunking) on the 🤗Transformers docs.
### Benchmark
Let's test how much memory can be saved by using chunked feed forward layers.
```
#@title Installs and Imports
# pip installs
!pip -qq install git+https://github.com/huggingface/transformers.git
!pip install -qq py3nvml
from transformers import ReformerConfig, PyTorchBenchmark, PyTorchBenchmarkArguments
```
Building wheel for transformers (setup.py) ... [?25l[?25hdone
First, let's compare the default `google/reformer-enwik8` model without chunked feed forward layers to the one with chunked feed forward layers.
```
config_no_chunk = ReformerConfig.from_pretrained("google/reformer-enwik8") # no chunk
config_chunk = ReformerConfig.from_pretrained("google/reformer-enwik8", chunk_size_feed_forward=1) # feed forward chunk
benchmark_args = PyTorchBenchmarkArguments(sequence_lengths=[1024, 2048, 4096], batch_sizes=[8], models=["Reformer-No-Chunk", "Reformer-Chunk"], no_speed=True, no_env_print=True)
benchmark = PyTorchBenchmark(configs=[config_no_chunk, config_chunk], args=benchmark_args)
result = benchmark.run()
```
1 / 2
Doesn't fit on GPU. CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 11.17 GiB total capacity; 7.85 GiB already allocated; 1.74 GiB free; 9.06 GiB reserved in total by PyTorch)
2 / 2
Doesn't fit on GPU. CUDA out of memory. Tried to allocate 2.00 GiB (GPU 0; 11.17 GiB total capacity; 7.85 GiB already allocated; 1.24 GiB free; 9.56 GiB reserved in total by PyTorch)
==================== INFERENCE - MEMORY - RESULT ====================
--------------------------------------------------------------------------------
Model Name Batch Size Seq Length Memory in MB
--------------------------------------------------------------------------------
Reformer-No-Chunk 8 1024 4281
Reformer-No-Chunk 8 2048 7607
Reformer-No-Chunk 8 4096 N/A
Reformer-Chunk 8 1024 4309
Reformer-Chunk 8 2048 7669
Reformer-Chunk 8 4096 N/A
--------------------------------------------------------------------------------
Interesting, chunked feed forward layers do not seem to help here at all. The reason is that `config.feed_forward_size` is not sufficiently large to make a real difference. Only at longer sequence lengths of 4096, a slight decrease in memory usage can be seen.
Let's see what happens to the memory peak usage if we increase the size of the feed forward layer by a factor of 4 and reduce the number of attention heads also by a factor of 4 so that the feed forward layer becomes the memory bottleneck.
```
config_no_chunk = ReformerConfig.from_pretrained("google/reformer-enwik8", chunk_size_feed_forward=0, num_attention_{h}eads=2, feed_forward_size=16384) # no chuck
config_chunk = ReformerConfig.from_pretrained("google/reformer-enwik8", chunk_size_feed_forward=1, num_attention_{h}eads=2, feed_forward_size=16384) # feed forward chunk
benchmark_args = PyTorchBenchmarkArguments(sequence_lengths=[1024, 2048, 4096], batch_sizes=[8], models=["Reformer-No-Chunk", "Reformer-Chunk"], no_speed=True, no_env_print=True)
benchmark = PyTorchBenchmark(configs=[config_no_chunk, config_chunk], args=benchmark_args)
result = benchmark.run()
```
1 / 2
2 / 2
==================== INFERENCE - MEMORY - RESULT ====================
--------------------------------------------------------------------------------
Model Name Batch Size Seq Length Memory in MB
--------------------------------------------------------------------------------
Reformer-No-Chunk 8 1024 3743
Reformer-No-Chunk 8 2048 5539
Reformer-No-Chunk 8 4096 9087
Reformer-Chunk 8 1024 2973
Reformer-Chunk 8 2048 3999
Reformer-Chunk 8 4096 6011
--------------------------------------------------------------------------------
Now a clear decrease in peak memory usage can be seen for longer input sequences.
As a conclusion, it should be noted chunked feed forward layers only makes sense for models having few attention heads and large feed forward layers.
## 3. Reversible Residual Layers
Reversible residual layers were first introduced in [N. Gomez et al](https://arxiv.org/abs/1707.04585) and used to reduce memory consumption when training the popular *ResNet* model. Mathematically, reversible residual layers are slightly different
to "real" residual layers but do not require the activations to be saved during the forward pass, which can drastically reduce memory consumption for training.
### Reversible Residual Layers in Reformer
Let's start by investigating why training a model requires
much more memory than the inference of the model.
When running a model in inference, the required memory equals more or less the memory it takes to compute the **single** largest tensor in the model.
On the other hand, when training a model, the required memory equals more or less the **sum** of all differentiable tensors.
This is not surprising when considering how auto differentiation works in deep learning frameworks. These lecture [slides](https://www.cs.toronto.edu/~rgrosse/courses/csc321_2018/slides/lec10.pdf) by Roger Grosse of the University of Toronto are great to better understand auto differentiation.
In a nutshell, in order to calculate the gradient of a differentiable function (*e.g.* a layer), auto differentiation requires the gradient of the function's output and the function's input and output tensor. While the gradients are dynamically computed and subsequently discarded, the input and output tensors (*a.k.a* activations) of a function are stored during the forward pass.
Alright, let's apply this to a transformer model. A transformer model includes a stack of multiple so-called transformer layers. Each additional transformer layer forces the model to store more activations during the forward pass and thus increases the required memory for training.
Let's take a more detailed look. A transformer layer essentially consists of two residual layers. The first residual layer represents the *self-attention* mechanism as explained in section 1) and the second residual layer represents the *linear* or feed-forward layers as explained in section 2).
Using the same notation as before, the input of a transformer layer *i.e.* \\( \mathbf{X} \\) is first normalized \\( ^{1} \\) and subsequently processed by the self-attention layer to get the output \\( \mathbf{Z} = \text{SelfAttn}(\text{LayerNorm}(\mathbf{X})) \\). We will abbreviate these two layers with \\( G \\) so that \\( \mathbf{Z} = G(\mathbf{X}) \\).
Next, the residual \\( \mathbf{Z} \\) is added to the input \\( \mathbf{\overline{Z}} = \mathbf{Z} + \mathbf{X} \\) and the sum is fed into the second residual layer - the two linear layers. \\( \mathbf{\overline{Z}} \\) is processed by a second normalization layer, followed by the two linear layers to get \\( \mathbf{Y} = \text{Linear}(\text{LayerNorm}(\mathbf{Z} + \mathbf{X})) \\). We will abbreviate the second normalization layer and the two linear layers with \\( F \\) yielding \\( \mathbf{Y} = F(\mathbf{\overline{Z}}) \\).
Finally, the residual \\( \mathbf{Y} \\) is added to \\( \mathbf{\overline{Z}} \\) to give the output of the transformer layer \\( \mathbf{\overline{Y}} = \mathbf{Y} + \mathbf{\overline{Z}} \\).
Let's illustrate a complete transformer layer using the example of \\( \mathbf{x}_1, \ldots, \mathbf{x}_{16} \\).
To calculate the gradient of *e.g.* the self-attention block \\( G \\), three tensors have to be known beforehand: the gradient \\( \partial \mathbf{Z} \\), the output \\( \mathbf{Z} \\), and the input \\( \mathbf{X} \\). While \\( \partial \mathbf{Z} \\) can be calculated on-the-fly and discarded afterward, the values for \\( \mathbf{Z} \\) and \\( \mathbf{X} \\) have to be calculated and stored during the forward pass since it is not possible to recalculate them easily on-the-fly during backpropagation. Therefore, during the forward pass, large tensor outputs, such as the query-key dot product matrix \\( \mathbf{Q}\mathbf{K}^T \\) or the intermediate output of the linear layers \\( \mathbf{Y}^{\text{int}} \\), have to be stored in memory \\( ^{2} \\).
Here, reversible residual layers come to our help. The idea is relatively straight-forward. The residual block is designed in a way so that instead of having to store the input and output tensor of a function, both can easily be recalculated during the backward pass so that no tensor has to be stored in memory during the forward pass.
This is achieved by using two input streams \\( \mathbf{X}^{(1)}, \mathbf{X}^{(2)} \\), and two output streams \\( \mathbf{\overline{Y}}^{(1)}, \mathbf{\overline{Y}}^{(2)} \\). The first residual \\( \mathbf{Z} \\) is computed by the first output stream \\( \mathbf{Z} = G(\mathbf{X}^{(1)}) \\) and subsequently added to the input of the second input stream, so that \\( \mathbf{\overline{Z}} = \mathbf{Z} + \mathbf{X}^{(2)} \\).
Similarly, the residual \\( \mathbf{Y} = F(\mathbf{\overline{Z}}) \\) is added to the first input stream again, so that the two output streams are defined by \\( \mathbf{Y}^{(1)} = \mathbf{Y} + \mathbf{X}^{(1)} \\) and \\( \mathbf{Y}^{(2)} = \mathbf{X}^{(2)} + \mathbf{Z} = \mathbf{\overline{Z}} \\).
The reversible transformer layer can be visualized for \\( \mathbf{x}_1, \ldots, \mathbf{x}_{16} \\) as follows.
As can be seen, the outputs \\( \mathbf{\overline{Y}}^{(1)}, \mathbf{\overline{Y}}^{(2)} \\) are calculated in a very similar way than \\( \mathbf{\overline{Y}} \\) of the non-reversible layer, but they are mathematically different. The authors of Reformer observe in some initial experiments that the performance of a reversible transformer model matches the performance of a standard transformer model.
The first visible difference to the standard transformer layer is that there are two input streams and output streams \\( ^{3} \\), which at first slightly increases the required memory for both the forward pass.
The two-stream architecture is crucial though for not having to save any activations during the forward pass. Let's explain. For backpropagation, the reversible transformer layer has to calculate the gradients \\( \partial G \\) and \\( \partial F \\). In addition to the gradients \\( \partial \mathbf{Y} \\) and \\( \partial \mathbf{Z} \\) which can be calculated on-the-fly, the tensor values \\( \mathbf{Y} \\), \\( \mathbf{\overline{Z}} \\) have to be known for \\( \partial F \\) and the tensor values \\( \mathbf{Z} \\) and \\( \mathbf{X}^{(1)} \\) for \\( \partial G \\) to make auto-differentiation work.
If we assume to know \\( \mathbf{\overline{Y}}^{(1)}, \mathbf{\overline{Y}}^{(2)} \\), it can easily be depicted from the graph that one can calculate \\( \mathbf{X}^{(1)}, \mathbf{X}^{(2)} \\) as follows. \\( \mathbf{X}^{(1)} = F(\mathbf{\overline{Y}}^{(1)}) - \mathbf{\overline{Y}}^{(1)} \\). Great, now that \\( \mathbf{X}^{(1)} \\) is known, \\( \mathbf{X}^{(2)} \\) can be computed by \\( \mathbf{X}^{(2)} = \mathbf{\overline{Y}}^{(1)} - G(\mathbf{X}^{(1)}) \\). Alright now, \\( \mathbf{Z} \\) and \\( \mathbf{Y} \\) are trivial to compute via \\( \mathbf{Y} = \mathbf{\overline{Y}}^{(1)} - \mathbf{X}^{(1)} \\) and \\( \mathbf{Z} = \mathbf{\overline{Y}}^{(2)} - \mathbf{X}^{(2)} \\). So as a conclusion, if only the outputs \\( \mathbf{\overline{Y}}^{(1)}, \mathbf{\overline{Y}}^{(2)} \\) of the **last** reversible transformer layer are stored during the forward pass, all other relevant activations can be derived by making use of \\( G \\) and \\( F \\) during the backward pass and passing \\( \mathbf{X}^{(1)} \\) and \\( \mathbf{X}^{(2)} \\). The overhead of two forward passes of \\( G \\) and \\( F \\) per reversible transformer layer during the backpropagation is traded against not having to store any activations during the forward pass. Not a bad deal!
**Note**: Since recently, major deep learning frameworks have released code that allows to store only certain activations and recompute larger ones during the backward propagation (Tensoflow [here](https://www.tensorflow.org/api_docs/python/tf/recompute_grad) and PyTorch [here](https://pytorch.org/docs/stable/checkpoint.html)). For standard reversible layers, this still means that at least one activation has to be stored for each transformer layer, but by defining which activations can dynamically be recomputed a lot of memory can be saved.
---
\\( ^{1} \\) In the previous two sections, we have omitted the layer norm layers preceding both the self-attention layer and the linear layers. The reader should know that both \\( \mathbf{X} \\) and \\( \mathbf{\overline{Z}} \\) are both processed by layer normalization before being fed into self-attention and the linear layers respectively.
\\( ^{2} \\) While in the design the dimension of \\( \mathbf{Q}\mathbf{K} \\) is written as \\( n \times n \\), in a *LSH self-attention* or *local self-attention* layer the dimension would only be \\( n \times l_{c} \times n_{h} \\) or \\( n \times l_{c} \\) respectively with \\( l_{c} \\) being the chunk length and \\( n_{h} \\) the number of hashes
\\( ^{3} \\) In the first reversible transformer layer \\( \mathbf{X}^{(2)} \\) is set to be equal to \\( \mathbf{X}^{(1)} \\).
### Benchmark
In order to measure the effect of reversible residual layers, we will compare the memory consumption of BERT with Reformer in training for an increasing number of layers.
```
#@title Installs and Imports
# pip installs
!pip -qq install git+https://github.com/huggingface/transformers.git
!pip install -qq py3nvml
from transformers import ReformerConfig, BertConfig, PyTorchBenchmark, PyTorchBenchmarkArguments
```
Let's measure the required memory for the standard `bert-base-uncased` BERT model by increasing the number of layers from 4 to 12.
```
config_4_layers_bert = BertConfig.from_pretrained("bert-base-uncased", num_hidden_layers=4)
config_8_layers_bert = BertConfig.from_pretrained("bert-base-uncased", num_hidden_layers=8)
config_12_layers_bert = BertConfig.from_pretrained("bert-base-uncased", num_hidden_layers=12)
benchmark_args = PyTorchBenchmarkArguments(sequence_lengths=[512], batch_sizes=[8], models=["Bert-4-Layers", "Bert-8-Layers", "Bert-12-Layers"], training=True, no_inference=True, no_speed=True, no_env_print=True)
benchmark = PyTorchBenchmark(configs=[config_4_layers_bert, config_8_layers_bert, config_12_layers_bert], args=benchmark_args)
result = benchmark.run()
```
HBox(children=(FloatProgress(value=0.0, description='Downloading', max=433.0, style=ProgressStyle(description_…
1 / 3
2 / 3
3 / 3
==================== TRAIN - MEMORY - RESULTS ====================
--------------------------------------------------------------------------------
Model Name Batch Size Seq Length Memory in MB
--------------------------------------------------------------------------------
Bert-4-Layers 8 512 4103
Bert-8-Layers 8 512 5759
Bert-12-Layers 8 512 7415
--------------------------------------------------------------------------------
It can be seen that adding a single layer of BERT linearly increases the required memory by more than 400MB.
```
config_4_layers_reformer = ReformerConfig.from_pretrained("google/reformer-enwik8", num_hidden_layers=4, num_hashes=1)
config_8_layers_reformer = ReformerConfig.from_pretrained("google/reformer-enwik8", num_hidden_layers=8, num_hashes=1)
config_12_layers_reformer = ReformerConfig.from_pretrained("google/reformer-enwik8", num_hidden_layers=12, num_hashes=1)
benchmark_args = PyTorchBenchmarkArguments(sequence_lengths=[512], batch_sizes=[8], models=["Reformer-4-Layers", "Reformer-8-Layers", "Reformer-12-Layers"], training=True, no_inference=True, no_speed=True, no_env_print=True)
benchmark = PyTorchBenchmark(configs=[config_4_layers_reformer, config_8_layers_reformer, config_12_layers_reformer], args=benchmark_args)
result = benchmark.run()
```
1 / 3
2 / 3
3 / 3
==================== TRAIN - MEMORY - RESULTS ====================
--------------------------------------------------------------------------------
Model Name Batch Size Seq Length Memory in MB
--------------------------------------------------------------------------------
Reformer-4-Layers 8 512 4607
Reformer-8-Layers 8 512 4987
Reformer-12-Layers 8 512 5367
--------------------------------------------------------------------------------
For Reformer, on the other hand, adding a layer adds significantly less memory in practice. Adding a single layer increases the required memory on average by less than 100MB so that a much larger 12-Layer `reformer-enwik8` model requires less memory than a 12-Layer `bert-base-uncased` model.
## 4. Axial Positional Encodings
Reformer makes it possible to process huge input sequences. However, for such long input sequences standard positional encoding weight matrices alone would use more than 1GB to store its weights.
To prevent such large positional encoding matrices, the official Reformer code introduced *Axial Position Encodings*.
**Important:** *Axial Position Encodings were not explained in the official paper, but can be well understood from looking into the code and talking to the authors*
### Axial Positional Encodings in Reformer
Transformers need positional encodings to account for the order of words in the input because self-attention layers have *no notion of order*.
Positional encodings are usually defined by a simple look-up matrix \\( \mathbf{E} = \left[\mathbf{e}_1, \ldots, \mathbf{e}_{n_\text{max}}\right] \\) The positional encoding vector \\( \mathbf{e}_{i} \\) is then simply added to the *ith* input vector \\( \mathbf{x}_{i} + \mathbf{e}_{i} \\) so that the model can distinguish if an input vector (*a.k.a* token) is at position \\( i \\) or \\( j \\).
For every input position, the model needs to be able to look up the corresponding positional encoding vector so that the dimension of \\( \mathbf{E} \\) is defined by the maximum length of input vectors the model can process `config.max_position_embeddings`, *i.e.* \\( n_\text{max} \\), and the `config.hidden_size`, *i.e.* \\( d_{h} \\) of the input vectors.
Assuming \\( d_{h}=4 \\) and \\( n_\text{max}=49 \\), such a positional encoding matrix can be visualized as follows:
Here, we showcase only the positional encodings \\( \mathbf{e}_{1} \\), \\( \mathbf{e}_{2} \\), and \\( \mathbf{e}_{49} \\) each of dimension, *a.k.a* height 4.
Let's imagine, we want to train a Reformer model on sequences of a length of up to 0.5M tokens and an input vector `config.hidden_size` of 1024 (see notebook [here](https://github.com/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb)). The corresponding positional embeddings have a size of \\( 0.5M \times 1024 \sim 512M \\) parameters, which corresponds to a size of 2GB.
Such positional encodings would use an unnecessarily large amount of memory both when loading the model in memory and when saving the model on a hard drive.
The Reformer authors managed to drastically shrink the positional encodings in size by cutting the `config.hidden_size` dimension in two and smartly factorizing
the \\( n_\text{max} \\) dimension.
In Transformer, the user can decide into which shape \\( n_\text{max} \\) can be factorized into by setting `config.axial_pos_shape` to an appropriate
list of two values \\( n_\text{max}^1 \\) and \\( n_\text{max}^2 \\) so that \\( n_\text{max}^1 \times n_\text{max}^2 = n_\text{max} \\). By setting `config.axial_pos_embds_dim` to an
appropriate list of two values \\( d_{h}^{1} \\) and \\( d_{h}^2 \\) so that \\( d_{h}^1 + d_{h}^2 = d_{h} \\), the user can decide how the hidden size dimension should be cut.
Now, let's visualize and explain more intuitively.
One can think of factorizing \\( n_{\text{max}} \\) as folding the dimension into a third axis, which is shown in the following for the factorization `config.axial_pos_shape = [7, 7]`:
Each of the three standing rectangular prisms corresponds to one of the encoding vectors \\( \mathbf{e}_{1}, \mathbf{e}_{2}, \mathbf{e}_{49} \\), but we can see that the 49 encoding vectors are divided into 7 rows of 7 vectors each.
Now the idea is to use only one row of 7 encoding vectors and expand those vectors to the other 6 rows, essentially reusing their values.
Because it is discouraged to have the same values for different encoding vectors, each vector of dimension (*a.k.a* height) `config.hidden_size=4` is cut into the lower encoding vector \\( \mathbf{e}_\text{down} \\) of size \\( 1 \\) and \\( \mathbf{e}_\text{up} \\) of size \\( 3 \\), so that the lower part can be expanded along the row dimension and the upper part can be expanded along the column dimension.
Let's visualize for more clarity.
We can see that we have cut the embedding vectors into \\( \mathbf{e}_\text{down} \\) (*in blue*) and \\( \mathbf{e}_\text{up} \\) (*in yellow*).
Now for the "sub"-vectors \\( \mathbf{E}_\text{down} = \left[\mathbf{e}_{\text{down},1}, \ldots, \mathbf{e}_{\text{down},49}\right] \\) only the first row, *a.k.a.* the width in the graphic, of \\( 7 \\) is kept and expanded along the column dimension, *a.k.a.* the depth of the graphic. Inversely, for the "sub"-vectors \\( \mathbf{E}_\text{up} = \left[\mathbf{e}_{\text{up},1}, \ldots, \mathbf{e}_{\text{up},49}\right] \\) only the first column of \\( 7 \\) is kept and expanded along the row dimension.
The resulting embedding vectors \\( \mathbf{e'}_{i} \\) then correspond to
$$\mathbf{e'}_{i} = \left[ \left[\mathbf{e}_{\text{down, } i \% n_\text{max}^1}\right]^T, \left[\mathbf{e}_{\text{up, } \left \lfloor{\frac{i}{{n}^2_{\text{max}}}}\right \rfloor} \right]^T \right]^T $$
whereas \\( n_\text{max}^1 = 7 \\) and \\( n_\text{max}^2 = 7 \\) in our example.
These new encodings \\( \mathbf{E'} = \left[\mathbf{e'}_{1}, \ldots, \mathbf{e'}_{n_\text{max}}\right] \\) are called **Axial Position Encodings**.
In the following, these axial position encodings are illustrated in more detail for our example.
Now it should be more understandable how the final positional encoding vectors \\( \mathbf{E'} \\) are calculated only from \\( \mathbf{E}_{\text{down}} \\) of dimension \\( d_{h}^1 \times n_{\text{max}^1} \\) and \\( \mathbf{E}_{\text{up}} \\) of dimension \\( d_{h}^2 \times n_{\text{max}}^2 \\).
The crucial aspect to see here is that Axial Positional Encodings make sure that none of the vectors \\( \left[\mathbf{e'}_1, \ldots, \mathbf{e'}_{n_{\text{max}}}\right] \\) are equal to each other by design and that the overall size of the encoding matrix is reduced from \\( n_{\text{max}} \times d_{h} \\) to \\( n_{\text{max}}^1 \times d_{h}^1 + n_\text{max}^2 \times d_{h}^2 \\).
By allowing each axial positional encoding vector to be different by design the model is given much more flexibility to learn efficient positional representations if axial positional encodings are learned by the model.
To demonstrate the drastic reduction in size,
let's assume we would have set `config.axial_pos_shape = [1024, 512]` and `config.axial_pos_embds_dim = [512, 512]` for a Reformer model that can process inputs up to a length of 0.5M tokens. The resulting axial positional encoding matrix would have had a size of only \\( 1024 \times 512 + 512 \times 512 \sim 800K \\) parameters which corresponds to roughly 3MB. This is a drastic reduction from the 2GB a standard positional encoding matrix would require in this case.
For a more condensed and math-heavy explanation please refer to the 🤗Transformers docs [here](https://huggingface.co/transformers/model_doc/reformer.html#axial-positional-encodings).
### Benchmark
Lastly, let's also compare the peak memory consumption of conventional positional embeddings to *axial positional embeddings*.
```
#@title Installs and Imports
# pip installs
!pip -qq install git+https://github.com/huggingface/transformers.git
!pip install -qq py3nvml
from transformers import ReformerConfig, PyTorchBenchmark, PyTorchBenchmarkArguments, ReformerModel
```
Positional embeddings depend only on two configuration parameters: The maximum allowed length of input sequences `config.max_position_embeddings` and `config.hidden_size`. Let's use a model that pushes the maximum allowed length of input sequences to half a million tokens, called `google/reformer-crime-and-punishment`, to see the effect of using axial positional embeddings.
To begin with, we will compare the shape of axial position encodings with standard positional encodings and the number of parameters in the model.
```
config_no_pos_axial_embeds = ReformerConfig.from_pretrained("google/reformer-crime-and-punishment", axial_pos_embds=False) # disable axial positional embeddings
config_pos_axial_embeds = ReformerConfig.from_pretrained("google/reformer-crime-and-punishment", axial_pos_embds=True, axial_pos_embds_dim=(64, 192), axial_pos_shape=(512, 1024)) # enable axial positional embeddings
print("Default Positional Encodings")
print(20 * '-')
model = ReformerModel(config_no_pos_axial_embeds)
print(f"Positional embeddings shape: {model.embeddings.position_embeddings}")
print(f"Num parameters of model: {model.num_parameters()}")
print(20 * '-' + '\n\n')
print("Axial Positional Encodings")
print(20 * '-')
model = ReformerModel(config_pos_axial_embeds)
print(f"Positional embeddings shape: {model.embeddings.position_embeddings}")
print(f"Num parameters of model: {model.num_parameters()}")
print(20 * '-' + '\n\n')
```
HBox(children=(FloatProgress(value=0.0, description='Downloading', max=1151.0, style=ProgressStyle(description…
Default Positional Encodings
--------------------
Positional embeddings shape: PositionEmbeddings(
(embedding): Embedding(524288, 256)
)
Num parameters of model: 136572416
--------------------
Axial Positional Encodings
--------------------
Positional embeddings shape: AxialPositionEmbeddings(
(weights): ParameterList(
(0): Parameter containing: [torch.FloatTensor of size 512x1x64]
(1): Parameter containing: [torch.FloatTensor of size 1x1024x192]
)
)
Num parameters of model: 2584064
--------------------
Having read the theory, the shape of the axial positional encoding weights should not be a surprise to the reader.
Regarding the results, it can be seen that for models being capable of processing such long input sequences, it is not practical to use default positional encodings.
In the case of `google/reformer-crime-and-punishment`, standard positional encodings alone contain more than 100M parameters.
Axial positional encodings reduce this number to just over 200K.
Lastly, let's also compare the required memory at inference time.
```
benchmark_args = PyTorchBenchmarkArguments(sequence_lengths=[512], batch_sizes=[8], models=["Reformer-No-Axial-Pos-Embeddings", "Reformer-Axial-Pos-Embeddings"], no_speed=True, no_env_print=True)
benchmark = PyTorchBenchmark(configs=[config_no_pos_axial_embeds, config_pos_axial_embeds], args=benchmark_args)
result = benchmark.run()
```
1 / 2
2 / 2
==================== INFERENCE - MEMORY - RESULT ====================
--------------------------------------------------------------------------------
Model Name Batch Size Seq Length Memory in MB
--------------------------------------------------------------------------------
Reformer-No-Axial-Pos-Embeddin 8 512 959
Reformer-Axial-Pos-Embeddings 8 512 447
--------------------------------------------------------------------------------
It can be seen that using axial positional embeddings reduces the memory requirement to approximately half in the case of `google/reformer-crime-and-punishment`.
| huggingface/blog/blob/main/reformer.md |
!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Using 🤗 Simulate to learn Agent behaviors with Stable-Baselines3
We provide several example RL integrations with the Stable-Baselines3 (LINK) library. To install this dependancy use `pip install simulate[sb3]`.
Including:
* Learning to navigate in a simple T-Maze
* Collecting objects
* Navigating in procedurally generated mazes
* Physical interaction with movable objects
* Reward functions based on line of sight observation of objects.
## Learning to navigate in a simple T-Maze
<img class="!m-0 !border-0 !dark:border-0 !shadow-none !max-w-lg w-[600px]" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/simulate/simulate_sb3_basic_maze.png"/>
Example: [`sb3_basic_maze.py`](https://github.com/huggingface/simulate/examples/rl/sb3_basic_maze.py)
Objective: Navigate to a spherical object in a simple T-Maze. Upon object collection, the environment resets.
Actors: An EgoCentric Camera Actor (LINK) equipped with a monocular camera.
Observation space:
- An RGB camera of shape (3, 40, 40) (C, H, W) in uint8 format.
Action space:
- A discrete action space with 3 possible actions
- Turn left 10 degrees
- Turn right 10 degrees
- Move forward
Reward function:
- A dense reward based on improvement in best euclidean distance to the object
- A sparse reward of +1 when the object is collected
- A timeout penaly of -1 if the agent does not reach the object in 200 time-steps
Parallel: 4 independent instances of the same environment configuration.
## Collecting objects
<img class="!m-0 !border-0 !dark:border-0 !shadow-none !max-w-lg w-[600px]" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/simulate/simulate_sb3_collectables.png"/>
Example: [`sb3_collectables.py`](https://github.com/huggingface/simulate/examples/rl/sb3_collectables.py)
Objective: Collect all 20 objects in a large square room.
Actors: An EgoCentric Camera Actor (LINK) equipped with a monocular camera.
Observation space:
- An RGB camera of shape (3, 40, 40) (C, H, W) in uint8 format.
Action space:
- A discrete action space with 3 possible actions
- Turn left 10 degrees
- Turn right 10 degrees
- Move forward
Reward function:
- A sparse reward of +1 when an object is collected
- A timeout penaly of -1 if the agent does not reach the object in 500 time-steps
Parallel: 4 independent instances of the same environment configuration.
## Navigating in procedurally generated mazes
<img class="!m-0 !border-0 !dark:border-0 !shadow-none !max-w-lg w-[600px]" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/simulate/simulate_sb3_procgen.png"/>
Example: [`sb3_procgen.py`](https://github.com/huggingface/simulate/examples/rl/sb3_procgen.py)
Objective: Navigate to an object in a 3D maze, when the object is collected the environment resets.
Actors: An EgoCentric Camera Actor (LINK) equipped with a monocular camera
Observation space:
- An RGB camera of shape (3, 40, 40) (C, H, W) in uint8 format.
Action space:
- A discrete action space with 3 possible actions
- Turn left 10 degrees
- Turn right 10 degrees
- Move forward
Reward function:
- A sparse reward of +1 when the object is reached
- A timeout penaly of -1 if the agent does not reach the object in 500 time-steps
Parallel: 4 independent instances of randomly generated environment configurations.
## Physical interaction with movable objects
<img class="!m-0 !border-0 !dark:border-0 !shadow-none !max-w-lg w-[600px]" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/simulate/simulate_sb3_move_boxes.png"/>
Example: [`sb3_move_boxes.py`](https://github.com/huggingface/simulate/examples/rl/sb3_move_boxes.py)
Objective: Push boxes in a room near to each other.
Actors: An EgoCentric Camera Actor (LINK) equipped with a monocular camera
Observation space:
- An RGB camera of shape (3, 40, 40) (C, H, W) in uint8 format.
Action space:
- A discrete action space with 3 possible actions
- Turn left 10 degrees
- Turn right 10 degrees
- Move forward
Reward function:
- A reward for moving the red and yellow boxes close to eachother
- A reward for moving the green and white boxes close to eachother
- A timeout penaly of -1 if the agent does not reach the object in 100 time-steps
Parallel: 16 independent instances of the same environment configuration.
## Reward functions based on line of sight observation of objects.
<img class="!m-0 !border-0 !dark:border-0 !shadow-none !max-w-lg w-[600px]" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/simulate/simulate_sb3_see_reward.png"/>
Example: [`sb3_visual_reward.py`](https://github.com/huggingface/simulate/examples/rl/sb3_visual_reward.py)
Objective: Move the agent so the box is within the agents its field of view
Actors: An EgoCentric Camera Actor (LINK) equipped with a monocular camera
Observation space:
- An RGB camera of shape (3, 40, 40) (C, H, W) in uint8 format.
Action space:
- A discrete action space with 3 possible actions
- Turn left 10 degrees
- Turn right 10 degrees
- Move forward
Reward function:
- A sparse reward for moving the box within a 60 degree fov cone in front of the agent.
- A timeout penaly of -1 if the agent does not reach the object in 100 time-steps
Parallel: 4 independent instances of the same environment configuration. | huggingface/simulate/blob/main/docs/source/howto/rl.mdx |
Gated models
To give more control over how models are used, the Hub allows model authors to enable **access requests** for their models. Users must agree to share their contact information (username and email address) with the model authors to access the model files when enabled. Model authors can configure this request with additional fields. A model with access requests enabled is called a **gated model**. Access requests are always granted to individual users rather than to entire organizations. A common use case of gated models is to provide access to early research models before the wider release.
## Manage gated models as a model author
<a id="manual-approval"></a> <!-- backward compatible anchor -->
<a id="notifications-settings"></a> <!-- backward compatible anchor -->
To enable access requests, go to the model settings page. By default, the model is not gated. Click on **Enable Access request** in the top-right corner.
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-disabled.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-disabled-dark.png"/>
</div>
By default, access to the model is automatically granted to the user when requesting it. This is referred to as **automatic approval**. In this mode, any user can access your model once they've shared their personal information with you.
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-enabled.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-enabled-dark.png"/>
</div>
If you want to manually approve which users can access your model, you must set it to **manual approval**. When this is the case, you will notice more options:
- **Add access** allows you to search for a user and grant them access even if they did not request it.
- **Notification frequency** lets you configure when to get notified if new users request access. It can be set to once a day or real-time. By default, an email is sent to your primary email address. You can set a different email address in the **Notifications email** field. For models hosted under an organization, emails are sent to the first 5 admins of the organization.
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-manual-approval.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-manual-approval-dark.png"/>
</div>
### Review access requests
Once access requests are enabled, you have full control of who can access your model or not, whether the approval mode is manual or automatic. You can review and manage requests either from the UI or via the API.
#### From the UI
You can review who has access to your gated model from its settings page by clicking on the **Review access requests** button. This will open a modal with 3 lists of users:
- **pending**: the list of users waiting for approval to access your model. This list is empty unless you've selected **manual approval**. You can either **Accept** or **Reject** the demand. If the demand is rejected, the user cannot access your model and cannot request access again.
- **accepted**: the complete list of users with access to your model. You can choose to **Reject** access at any time for any user, whether the approval mode is manual or automatic. You can also **Cancel** the approval, which will move the user to the *pending* list.
- **rejected**: the list of users you've manually rejected. Those users cannot access your models. If they go to your model repository, they will see a message *Your request to access this repo has been rejected by the repo's authors*.
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-enabled-pending-users.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-enabled-pending-users-dark.png"/>
</div>
#### Via the API
You can automate the approval of access requests by using the API. You must pass a `token` with `write` access to the gated repository. To generate a token, go to [your user settings](https://huggingface.co/settings/tokens).
| Method | URI | Description | Headers | Payload
| ------ | --- | ----------- | ------- | ------- |
| `GET` | `/api/models/{repo_id}/user-access-request/pending` | Retrieve the list of pending requests. | `{"authorization": "Bearer $token"}` | |
| `GET` | `/api/models/{repo_id}/user-access-request/accepted` | Retrieve the list of accepted requests. | `{"authorization": "Bearer $token"}` | |
| `GET` | `/api/models/{repo_id}/user-access-request/rejected` | Retrieve the list of rejected requests. | `{"authorization": "Bearer $token"}` | |
| `POST` | `/api/models/{repo_id}/user-access-request/handle` | Change the status of a given access request to `status`. | `{"authorization": "Bearer $token"}` | `{"status": "accepted"/"rejected"/"pending", "user": "username"}` |
| `POST` | `/api/models/{repo_id}/user-access-request/grant` | Allow a specific user to access your repo. | `{"authorization": "Bearer $token"}` | `{"user": "username"} ` |
The base URL for the HTTP endpoints above is `https://huggingface.co`.
Those endpoints are not officially supported in `huggingface_hub` or `huggingface.js` yet but [this code snippet](https://github.com/huggingface/huggingface_hub/issues/1535#issuecomment-1614693412) (in Python) might help you getting started.
NEW! There's an [open PR](https://github.com/huggingface/huggingface_hub/pull/1905) in `huggingface_hub` to add official support from our Python library.
### Download access report
You can download a report of all access requests for a gated model with the **download user access report** button. Click on it to download a json file with a list of users. For each entry, you have:
- **user**: the user id. Example: *julien-c*.
- **fullname**: name of the user on the Hub. Example: *Julien Chaumond*.
- **status**: status of the request. Either `"pending"`, `"accepted"` or `"rejected"`.
- **email**: email of the user.
- **time**: datetime when the user initially made the request.
<a id="modifying-the-prompt"></a> <!-- backward compatible anchor -->
### Customize requested information
By default, users landing on your gated model will be asked to share their contact information (email and username) by clicking the **Agree and send request to access repo** button.
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-user-side.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-user-side-dark.png"/>
</div>
If you want to collect more user information, you can configure additional fields. This information will be accessible from the **Settings** tab. To do so, add an `extra_gated_fields` property to your [model card metadata](./model-cards#model-card-metadata) containing a list of key/value pairs. The *key* is the name of the field and *value* its type. A field can be either `text` (free text area) or `checkbox`. Finally, you can also personalize the message displayed to the user with the `extra_gated_prompt` extra field.
Here is an example of customized request form where the user is asked to provide their company name and country and acknowledge that the model is for non-commercial use only.
```yaml
---
extra_gated_prompt: "You agree to not use the model to conduct experiments that cause harm to human subjects."
extra_gated_fields:
Company: text
Country: text
I agree to use this model for non-commercial use ONLY: checkbox
---
```
In some cases, you might also want to modify the text in the gate heading and the text in the button. For those use cases, you can modify `extra_gated_heading` and `extra_gated_button_content` like this:
```yaml
---
extra_gated_heading: "Acknowledge license to accept the repository"
extra_gated_button_content: "Acknowledge license"
---
```
### Example use cases of programmatically managing access requests
Here are a few interesting use cases of programmatically managing access requests for gated repos we've seen organically emerge in the community.
As a reminder, the model repo needs to be set to manual approval, otherwise users get access to it automatically.
Possible use cases of programmatic management include:
- If you have advanced user request screening requirements (for advanced compliance requirements, etc) or you wish to handle the user requests outside the Hub.
- An example for this was Meta's [Llama 2](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) initial release where users had to request access on a Meta website.
- You can ask users for their HF username in your access flow, and then use a script to programmatically accept user requests on the Hub based on your set of conditions.
- If you want to condition access to a model based on completing a payment flow (note that the actual payment flow happens outside of the Hub).
- Here's an [example repo](https://huggingface.co/Trelis/openchat_3.5-function-calling-v3) from TrelisResearch that uses this use case.
- [@RonanMcGovern](https://huggingface.co/RonanMcGovern) has posted a [video about the flow](https://www.youtube.com/watch?v=2OT2SI5auQU) and tips on how to implement it.
## Access gated models as a user
As a user, if you want to use a gated model, you will need to request access to it. This means that you must be logged in to a Hugging Face user account.
Requesting access can only be done from your browser. Go to the model on the Hub and you will be prompted to share your information:
<div class="flex justify-center">
<img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-user-side.png"/>
<img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/hub/models-gated-user-side-dark.png"/>
</div>
By clicking on **Agree**, you agree to share your username and email address with the model authors. In some cases, additional fields might be requested. To help the model authors decide whether to grant you access, try to fill out the form as completely as possible.
Once the access request is sent, there are two possibilities. If the approval mechanism is automatic, you immediately get access to the model files. Otherwise, the requests have to be approved manually by the authors, which can take more time.
<Tip warning={true}>
The model authors have complete control over model access. In particular, they can decide at any time to block your access to the model without prior notice, regardless of approval mechanism or if your request has already been approved.
</Tip>
### Download files
To download files from a gated model you'll need to be authenticated. In the browser, this is automatic as long as you are logged in with your account. If you are using a script, you will need to provide a [user token](./security-tokens). In the Hugging Face Python ecosystem (`transformers`, `diffusers`, `datasets`, etc.), you can login your machine using the [`huggingface_hub`](https://huggingface.co/docs/huggingface_hub/index) library and running in your terminal:
```bash
huggingface-cli login
```
Alternatively, you can programmatically login using `login()` in a notebook or a script:
```python
>>> from huggingface_hub import login
>>> login()
```
You can also provide the `token` parameter to most loading methods in the libraries (`from_pretrained`, `hf_hub_download`, `load_dataset`, etc.), directly from your scripts.
For more details about how to login, check out the [login guide](https://huggingface.co/docs/huggingface_hub/quick-start#login). | huggingface/hub-docs/blob/main/docs/hub/models-gated.md |
Gradio Demo: blocks_layout
```
!pip install -q gradio
```
```
import gradio as gr
demo = gr.Blocks()
with demo:
with gr.Row():
gr.Image(interactive=True, scale=2)
gr.Image()
with gr.Row():
gr.Textbox(label="Text")
gr.Number(label="Count", scale=2)
gr.Radio(choices=["One", "Two"])
with gr.Row():
gr.Button("500", scale=0, min_width=500)
gr.Button("A", scale=0)
gr.Button("grow")
with gr.Row():
gr.Textbox()
gr.Textbox()
gr.Button()
with gr.Row():
with gr.Row():
with gr.Column():
gr.Textbox(label="Text")
gr.Number(label="Count")
gr.Radio(choices=["One", "Two"])
gr.Image()
with gr.Column():
gr.Image(interactive=True)
gr.Image()
gr.Image()
gr.Textbox(label="Text")
gr.Number(label="Count")
gr.Radio(choices=["One", "Two"])
if __name__ == "__main__":
demo.launch()
```
| gradio-app/gradio/blob/main/demo/blocks_layout/run.ipynb |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# LoRA
Low-Rank Adaptation ([LoRA](https://huggingface.co/papers/2309.15223)) is a PEFT method that decomposes a large matrix into two smaller low-rank matrices in the attention layers. This drastically reduces the number of parameters that need to be fine-tuned.
The abstract from the paper is:
*We propose a neural language modeling system based on low-rank adaptation (LoRA) for speech recognition output rescoring. Although pretrained language models (LMs) like BERT have shown superior performance in second-pass rescoring, the high computational cost of scaling up the pretraining stage and adapting the pretrained models to specific domains limit their practical use in rescoring. Here we present a method based on low-rank decomposition to train a rescoring BERT model and adapt it to new domains using only a fraction (0.08%) of the pretrained parameters. These inserted matrices are optimized through a discriminative training objective along with a correlation-based regularization loss. The proposed low-rank adaptation Rescore-BERT (LoRB) architecture is evaluated on LibriSpeech and internal datasets with decreased training times by factors between 5.4 and 3.6.*.
## LoraConfig
[[autodoc]] tuners.lora.config.LoraConfig
## LoraModel
[[autodoc]] tuners.lora.model.LoraModel | huggingface/peft/blob/main/docs/source/package_reference/lora.md |
(Gluon) ResNeXt
A **ResNeXt** repeats a [building block](https://paperswithcode.com/method/resnext-block) that aggregates a set of transformations with the same topology. Compared to a [ResNet](https://paperswithcode.com/method/resnet), it exposes a new dimension, *cardinality* (the size of the set of transformations) $C$, as an essential factor in addition to the dimensions of depth and width.
The weights from this model were ported from [Gluon](https://cv.gluon.ai/model_zoo/classification.html).
## How do I use this model on an image?
To load a pretrained model:
```python
import timm
model = timm.create_model('gluon_resnext101_32x4d', pretrained=True)
model.eval()
```
To load and preprocess the image:
```python
import urllib
from PIL import Image
from timm.data import resolve_data_config
from timm.data.transforms_factory import create_transform
config = resolve_data_config({}, model=model)
transform = create_transform(**config)
url, filename = ("https://github.com/pytorch/hub/raw/master/images/dog.jpg", "dog.jpg")
urllib.request.urlretrieve(url, filename)
img = Image.open(filename).convert('RGB')
tensor = transform(img).unsqueeze(0) # transform and add batch dimension
```
To get the model predictions:
```python
import torch
with torch.no_grad():
out = model(tensor)
probabilities = torch.nn.functional.softmax(out[0], dim=0)
print(probabilities.shape)
# prints: torch.Size([1000])
```
To get the top-5 predictions class names:
```python
# Get imagenet class mappings
url, filename = ("https://raw.githubusercontent.com/pytorch/hub/master/imagenet_classes.txt", "imagenet_classes.txt")
urllib.request.urlretrieve(url, filename)
with open("imagenet_classes.txt", "r") as f:
categories = [s.strip() for s in f.readlines()]
# Print top categories per image
top5_prob, top5_catid = torch.topk(probabilities, 5)
for i in range(top5_prob.size(0)):
print(categories[top5_catid[i]], top5_prob[i].item())
# prints class names and probabilities like:
# [('Samoyed', 0.6425196528434753), ('Pomeranian', 0.04062102362513542), ('keeshond', 0.03186424449086189), ('white wolf', 0.01739676296710968), ('Eskimo dog', 0.011717947199940681)]
```
Replace the model name with the variant you want to use, e.g. `gluon_resnext101_32x4d`. You can find the IDs in the model summaries at the top of this page.
To extract image features with this model, follow the [timm feature extraction examples](https://rwightman.github.io/pytorch-image-models/feature_extraction/), just change the name of the model you want to use.
## How do I finetune this model?
You can finetune any of the pre-trained models just by changing the classifier (the last layer).
```python
model = timm.create_model('gluon_resnext101_32x4d', pretrained=True, num_classes=NUM_FINETUNE_CLASSES)
```
To finetune on your own dataset, you have to write a training loop or adapt [timm's training
script](https://github.com/rwightman/pytorch-image-models/blob/master/train.py) to use your dataset.
## How do I train this model?
You can follow the [timm recipe scripts](https://rwightman.github.io/pytorch-image-models/scripts/) for training a new model afresh.
## Citation
```BibTeX
@article{DBLP:journals/corr/XieGDTH16,
author = {Saining Xie and
Ross B. Girshick and
Piotr Doll{\'{a}}r and
Zhuowen Tu and
Kaiming He},
title = {Aggregated Residual Transformations for Deep Neural Networks},
journal = {CoRR},
volume = {abs/1611.05431},
year = {2016},
url = {http://arxiv.org/abs/1611.05431},
archivePrefix = {arXiv},
eprint = {1611.05431},
timestamp = {Mon, 13 Aug 2018 16:45:58 +0200},
biburl = {https://dblp.org/rec/journals/corr/XieGDTH16.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
```
<!--
Type: model-index
Collections:
- Name: Gloun ResNeXt
Paper:
Title: Aggregated Residual Transformations for Deep Neural Networks
URL: https://paperswithcode.com/paper/aggregated-residual-transformations-for-deep
Models:
- Name: gluon_resnext101_32x4d
In Collection: Gloun ResNeXt
Metadata:
FLOPs: 10298145792
Parameters: 44180000
File Size: 177367414
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: gluon_resnext101_32x4d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/gluon_resnet.py#L193
Weights: https://github.com/rwightman/pytorch-pretrained-gluonresnet/releases/download/v0.1/gluon_resnext101_32x4d-b253c8c4.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 80.33%
Top 5 Accuracy: 94.91%
- Name: gluon_resnext101_64x4d
In Collection: Gloun ResNeXt
Metadata:
FLOPs: 19954172928
Parameters: 83460000
File Size: 334737852
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: gluon_resnext101_64x4d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/gluon_resnet.py#L201
Weights: https://github.com/rwightman/pytorch-pretrained-gluonresnet/releases/download/v0.1/gluon_resnext101_64x4d-f9a8e184.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 80.63%
Top 5 Accuracy: 95.0%
- Name: gluon_resnext50_32x4d
In Collection: Gloun ResNeXt
Metadata:
FLOPs: 5472648192
Parameters: 25030000
File Size: 100441719
Architecture:
- 1x1 Convolution
- Batch Normalization
- Convolution
- Global Average Pooling
- Grouped Convolution
- Max Pooling
- ReLU
- ResNeXt Block
- Residual Connection
- Softmax
Tasks:
- Image Classification
Training Data:
- ImageNet
ID: gluon_resnext50_32x4d
Crop Pct: '0.875'
Image Size: '224'
Interpolation: bicubic
Code: https://github.com/rwightman/pytorch-image-models/blob/d8e69206be253892b2956341fea09fdebfaae4e3/timm/models/gluon_resnet.py#L185
Weights: https://github.com/rwightman/pytorch-pretrained-gluonresnet/releases/download/v0.1/gluon_resnext50_32x4d-e6a097c1.pth
Results:
- Task: Image Classification
Dataset: ImageNet
Metrics:
Top 1 Accuracy: 79.35%
Top 5 Accuracy: 94.42%
--> | huggingface/pytorch-image-models/blob/main/docs/models/gloun-resnext.md |
@gradio/app
## 1.17.0
### Features
- [#6831](https://github.com/gradio-app/gradio/pull/6831) [`f3abde8`](https://github.com/gradio-app/gradio/commit/f3abde80884d96ad69b825020c46486d9dd5cac5) - Add an option to enable header links for markdown. Thanks [@pngwn](https://github.com/pngwn)!
### Fixes
- [#6766](https://github.com/gradio-app/gradio/pull/6766) [`73268ee`](https://github.com/gradio-app/gradio/commit/73268ee2e39f23ebdd1e927cb49b8d79c4b9a144) - Improve source selection UX. Thanks [@hannahblair](https://github.com/hannahblair)!
## 1.16.2
### Patch Changes
- Updated dependencies [[`245d58e`](https://github.com/gradio-app/gradio/commit/245d58eff788e8d44a59d37a2d9b26d0f08a62b4), [`c352811`](https://github.com/gradio-app/gradio/commit/c352811f76d4126613ece0a584f8c552fdd8d1f6)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.16.1
### Patch Changes
- Updated dependencies [[`5d51fbc`](https://github.com/gradio-app/gradio/commit/5d51fbce7826da840a2fd4940feb5d9ad6f1bc5a), [`34f9431`](https://github.com/gradio-app/gradio/commit/34f943101bf7dd6b8a8974a6131c1ed7c4a0dac0)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.16.0
### Features
- [#6398](https://github.com/gradio-app/gradio/pull/6398) [`67ddd40`](https://github.com/gradio-app/gradio/commit/67ddd40b4b70d3a37cb1637c33620f8d197dbee0) - Lite v4. Thanks [@whitphx](https://github.com/whitphx)!
- [#6738](https://github.com/gradio-app/gradio/pull/6738) [`f3c4d78`](https://github.com/gradio-app/gradio/commit/f3c4d78b710854b94d9a15db78178e504a02c680) - reload on css changes + fix css specificity. Thanks [@pngwn](https://github.com/pngwn)!
### Fixes
- [#6639](https://github.com/gradio-app/gradio/pull/6639) [`9a6ff70`](https://github.com/gradio-app/gradio/commit/9a6ff704cd8429289c5376d3af5e4b8492df4773) - Fix issue with `head` param when adding more than one script tag. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
## 1.15.0
### Features
- [#6512](https://github.com/gradio-app/gradio/pull/6512) [`4f040c7`](https://github.com/gradio-app/gradio/commit/4f040c752bb3b0586a4e16eca25a1e5f596eee48) - Update zh-CN.json. Thanks [@cibimo](https://github.com/cibimo)!
## 1.14.0
### Features
- [#6537](https://github.com/gradio-app/gradio/pull/6537) [`6d3fecfa4`](https://github.com/gradio-app/gradio/commit/6d3fecfa42dde1c70a60c397434c88db77289be6) - chore(deps): update all non-major dependencies. Thanks [@renovate](https://github.com/apps/renovate)!
### Fixes
- [#6530](https://github.com/gradio-app/gradio/pull/6530) [`13ef0f0ca`](https://github.com/gradio-app/gradio/commit/13ef0f0caa13e5a1cea70d572684122419419599) - Quick fix: Make component interactive when it is in focus. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
## 1.13.1
### Fixes
- [#6536](https://github.com/gradio-app/gradio/pull/6536) [`1bbd6cab3`](https://github.com/gradio-app/gradio/commit/1bbd6cab3f0abe183b514b82061f0937c8480966) - Fix undefined `data` TypeError in Blocks. Thanks [@hannahblair](https://github.com/hannahblair)!
## 1.13.0
### Highlights
#### New `ImageEditor` component ([#6169](https://github.com/gradio-app/gradio/pull/6169) [`9caddc17b`](https://github.com/gradio-app/gradio/commit/9caddc17b1dea8da1af8ba724c6a5eab04ce0ed8))
A brand new component, completely separate from `Image` that provides simple editing capabilities.
- Set background images from file uploads, webcam, or just paste!
- Crop images with an improved cropping UI. App authors can event set specific crop size, or crop ratios (`1:1`, etc)
- Paint on top of any image (or no image) and erase any mistakes!
- The ImageEditor supports layers, confining draw and erase actions to that layer.
- More flexible access to data. The image component returns a composite image representing the final state of the canvas as well as providing the background and all layers as individual images.
- Fully customisable. All features can be enabled and disabled. Even the brush color swatches can be customised.
<video src="https://user-images.githubusercontent.com/12937446/284027169-31188926-fd16-4a1c-8718-998e7aae4695.mp4" autoplay muted></video>
```py
def fn(im):
im["composite"] # the full canvas
im["background"] # the background image
im["layers"] # a list of individual layers
im = gr.ImageEditor(
# decide which sources you'd like to accept
sources=["upload", "webcam", "clipboard"],
# set a cropsize constraint, can either be a ratio or a concrete [width, height]
crop_size="1:1",
# enable crop (or disable it)
transforms=["crop"],
# customise the brush
brush=Brush(
default_size="25", # or leave it as 'auto'
color_mode="fixed", # 'fixed' hides the user swatches and colorpicker, 'defaults' shows it
default_color="hotpink", # html names are supported
colors=[
"rgba(0, 150, 150, 1)", # rgb(a)
"#fff", # hex rgb
"hsl(360, 120, 120)" # in fact any valid colorstring
]
),
brush=Eraser(default_size="25")
)
```
Thanks [@pngwn](https://github.com/pngwn)!
## 1.12.0
### Features
- [#6427](https://github.com/gradio-app/gradio/pull/6427) [`e0fc14659`](https://github.com/gradio-app/gradio/commit/e0fc146598ba9b081bc5fa9616d0a41c2aba2427) - Allow google analytics to work on Spaces (and other iframe situations). Thanks [@abidlabs](https://github.com/abidlabs)!
### Fixes
- [#6254](https://github.com/gradio-app/gradio/pull/6254) [`f816136a0`](https://github.com/gradio-app/gradio/commit/f816136a039fa6011be9c4fb14f573e4050a681a) - Add volume control to Audio. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#6457](https://github.com/gradio-app/gradio/pull/6457) [`d00fcf89d`](https://github.com/gradio-app/gradio/commit/d00fcf89d1c3ecbc910e81bb1311479ec2b73e4e) - Gradio custom component dev mode now detects changes to Example.svelte file. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
## 1.11.0
### Features
- [#6099](https://github.com/gradio-app/gradio/pull/6099) [`d84209703`](https://github.com/gradio-app/gradio/commit/d84209703b7a0728cdb49221e543500ddb6a8d33) - Lite: SharedWorker mode. Thanks [@whitphx](https://github.com/whitphx)!
### Fixes
- [#6383](https://github.com/gradio-app/gradio/pull/6383) [`324867f63`](https://github.com/gradio-app/gradio/commit/324867f63c920113d89a565892aa596cf8b1e486) - Fix event target. Thanks [@aliabid94](https://github.com/aliabid94)!
## 1.10.3
### Patch Changes
- Updated dependencies [[`6204ccac5`](https://github.com/gradio-app/gradio/commit/6204ccac5967763e0ebde550d04d12584243a120), [`4d3aad33a`](https://github.com/gradio-app/gradio/commit/4d3aad33a0b66639dbbb2928f305a79fb7789b2d), [`854b482f5`](https://github.com/gradio-app/gradio/commit/854b482f598e0dc47673846631643c079576da9c), [`55fda81fa`](https://github.com/gradio-app/gradio/commit/55fda81fa5918b48952729232d6e2fc55af9351d), [`37dd335e5`](https://github.com/gradio-app/gradio/commit/37dd335e5f04a8e689dd7f23ae24ad1934ea08d8), [`f1409f95e`](https://github.com/gradio-app/gradio/commit/f1409f95ed39c5565bed6a601e41f94e30196a57)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.10.2
### Patch Changes
- Updated dependencies [[`4b1011bab`](https://github.com/gradio-app/gradio/commit/4b1011bab03c0b6a09329e0beb9c1b17b2189878), [`bca6c2c80`](https://github.com/gradio-app/gradio/commit/bca6c2c80f7e5062427019de45c282238388af95), [`19af2806a`](https://github.com/gradio-app/gradio/commit/19af2806a58419cc551d2d1d6d8987df0db91ccb), [`d3b53a457`](https://github.com/gradio-app/gradio/commit/d3b53a4577ea05cd27e37ce7fec952028c18ed45), [`3cdeabc68`](https://github.com/gradio-app/gradio/commit/3cdeabc6843000310e1a9e1d17190ecbf3bbc780), [`fad92c29d`](https://github.com/gradio-app/gradio/commit/fad92c29dc1f5cd84341aae417c495b33e01245f)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.10.1
### Patch Changes
- Updated dependencies [[`92278729e`](https://github.com/gradio-app/gradio/commit/92278729ee008126af15ffe6be399236211b2f34), [`e8216be94`](https://github.com/gradio-app/gradio/commit/e8216be948f76ce064595183d11e9148badf9421)]:
- @gradio/[email protected]
- @gradio/[email protected]
## 1.10.0
### Features
- [#6261](https://github.com/gradio-app/gradio/pull/6261) [`8bbeca0e7`](https://github.com/gradio-app/gradio/commit/8bbeca0e772a5a2853d02a058b35abb2c15ffaf1) - Improve Embed and CDN handling and fix a couple of related bugs. Thanks [@pngwn](https://github.com/pngwn)!
### Fixes
- [#6266](https://github.com/gradio-app/gradio/pull/6266) [`e32bac894`](https://github.com/gradio-app/gradio/commit/e32bac8944c85e0ec4831963299889d6bbfa0351) - Fix updating interactive prop. Thanks [@abidlabs](https://github.com/abidlabs)!
- [#6213](https://github.com/gradio-app/gradio/pull/6213) [`27194a987`](https://github.com/gradio-app/gradio/commit/27194a987fa7ba1234b5fc0ce8bf7fabef7033a9) - Ensure the statustracker for `gr.Image` displays in static mode. Thanks [@pngwn](https://github.com/pngwn)!
- [#6234](https://github.com/gradio-app/gradio/pull/6234) [`aaa55ce85`](https://github.com/gradio-app/gradio/commit/aaa55ce85e12f95aba9299445e9c5e59824da18e) - Video/Audio fixes. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#6236](https://github.com/gradio-app/gradio/pull/6236) [`6bce259c5`](https://github.com/gradio-app/gradio/commit/6bce259c5db7b21b327c2067e74ea20417bc89ec) - Ensure `gr.CheckboxGroup` updates as expected. Thanks [@pngwn](https://github.com/pngwn)!
## 1.9.2
### Fixes
- [#6191](https://github.com/gradio-app/gradio/pull/6191) [`b555bc09f`](https://github.com/gradio-app/gradio/commit/b555bc09ffe8e58b10da6227e2f11a0c084aa71d) - fix cdn build. Thanks [@pngwn](https://github.com/pngwn)!
## 1.9.1
### Features
- [#6137](https://github.com/gradio-app/gradio/pull/6137) [`2ba14b284`](https://github.com/gradio-app/gradio/commit/2ba14b284f908aa13859f4337167a157075a68eb) - JS Param. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
## 1.9.0
### Features
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Improve Audio Component. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Adds the ability to build the frontend and backend of custom components in preparation for publishing to pypi using `gradio_component build`. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Improve Video Component. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Custom components. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Swap websockets for SSE. Thanks [@pngwn](https://github.com/pngwn)!
### Fixes
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Pending events behavior. Thanks [@pngwn](https://github.com/pngwn)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`287fe6782`](https://github.com/gradio-app/gradio/commit/287fe6782825479513e79a5cf0ba0fbfe51443d7) - Reinstate types that were removed in error in #5832. Thanks [@pngwn](https://github.com/pngwn)!
## 1.9.0-beta.3
### Features
- [#6124](https://github.com/gradio-app/gradio/pull/6124) [`a7435ba9e`](https://github.com/gradio-app/gradio/commit/a7435ba9e6f8b88a838e80893eb8fedf60ccda67) - Fix static issues with Lite on v4. Thanks [@aliabd](https://github.com/aliabd)!
- [#6136](https://github.com/gradio-app/gradio/pull/6136) [`667802a6c`](https://github.com/gradio-app/gradio/commit/667802a6cdbfb2ce454a3be5a78e0990b194548a) - JS Component Documentation. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#6149](https://github.com/gradio-app/gradio/pull/6149) [`90318b1dd`](https://github.com/gradio-app/gradio/commit/90318b1dd118ae08a695a50e7c556226234ab6dc) - swap `mode` on the frontned to `interactive` to match the backend. Thanks [@pngwn](https://github.com/pngwn)!
- [#6118](https://github.com/gradio-app/gradio/pull/6118) [`88bccfdba`](https://github.com/gradio-app/gradio/commit/88bccfdba3df2df4b2747ea5d649ed528047cf50) - Improve Video Component. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#6126](https://github.com/gradio-app/gradio/pull/6126) [`865a22d5c`](https://github.com/gradio-app/gradio/commit/865a22d5c60fd97aeca968e55580b403743a23ec) - Refactor `Blocks.load()` so that it is in the same style as the other listeners. Thanks [@abidlabs](https://github.com/abidlabs)!
- [#6157](https://github.com/gradio-app/gradio/pull/6157) [`db143bdd1`](https://github.com/gradio-app/gradio/commit/db143bdd13b830f3bfd513bbfbc0cd1403522b84) - Make output components not editable if they are being updated. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
- [#6069](https://github.com/gradio-app/gradio/pull/6069) [`bf127e124`](https://github.com/gradio-app/gradio/commit/bf127e1241a41401e144874ea468dff8474eb505) - Swap websockets for SSE. Thanks [@aliabid94](https://github.com/aliabid94)!
## 1.9.0-beta.2
### Features
- [#6016](https://github.com/gradio-app/gradio/pull/6016) [`83e947676`](https://github.com/gradio-app/gradio/commit/83e947676d327ca2ab6ae2a2d710c78961c771a0) - Format js in v4 branch. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#5966](https://github.com/gradio-app/gradio/pull/5966) [`9cad2127b`](https://github.com/gradio-app/gradio/commit/9cad2127b965023687470b3abfe620e188a9da6e) - Improve Audio Component. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#5955](https://github.com/gradio-app/gradio/pull/5955) [`825c9cddc`](https://github.com/gradio-app/gradio/commit/825c9cddc83a09457d8c85ebeecb4bc705572d82) - Fix dev mode model3D. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#6107](https://github.com/gradio-app/gradio/pull/6107) [`9a40de7bf`](https://github.com/gradio-app/gradio/commit/9a40de7bff5844c8a135e73c7d175eb02b63a966) - Fix: Move to cache in init postprocess + Fallback Fixes. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#6089](https://github.com/gradio-app/gradio/pull/6089) [`cd8146ba0`](https://github.com/gradio-app/gradio/commit/cd8146ba053fbcb56cf5052e658e4570d457fb8a) - Update logos for v4. Thanks [@abidlabs](https://github.com/abidlabs)!
- [#5996](https://github.com/gradio-app/gradio/pull/5996) [`9cf40f76f`](https://github.com/gradio-app/gradio/commit/9cf40f76fed1c0f84b5a5336a9b0100f8a9b4ee3) - V4: Simple dropdown. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#5990](https://github.com/gradio-app/gradio/pull/5990) [`85056de5c`](https://github.com/gradio-app/gradio/commit/85056de5cd4e90a10cbfcefab74037dbc622b26b) - V4: Simple textbox. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
### Fixes
- [#6065](https://github.com/gradio-app/gradio/pull/6065) [`7d07001e8`](https://github.com/gradio-app/gradio/commit/7d07001e8e7ca9cbd2251632667b3a043de49f49) - fix storybook. Thanks [@pngwn](https://github.com/pngwn)!
- [#5826](https://github.com/gradio-app/gradio/pull/5826) [`ce036c5d4`](https://github.com/gradio-app/gradio/commit/ce036c5d47e741e29812654bcc641ea6be876504) - Pending events behavior. Thanks [@dawoodkhan82](https://github.com/dawoodkhan82)!
- [#6046](https://github.com/gradio-app/gradio/pull/6046) [`dbb7de5e0`](https://github.com/gradio-app/gradio/commit/dbb7de5e02c53fee05889d696d764d212cb96c74) - fix tests. Thanks [@pngwn](https://github.com/pngwn)!
- [#6076](https://github.com/gradio-app/gradio/pull/6076) [`f3f98f923`](https://github.com/gradio-app/gradio/commit/f3f98f923c9db506284b8440e18a3ac7ddd8398b) - Lite error handler. Thanks [@whitphx](https://github.com/whitphx)!
## 1.9.0-beta.1
### Patch Changes
- Updated dependencies [[`174b73619`](https://github.com/gradio-app/gradio/commit/174b736194756e23f51bbaf6f850bac5f1ca95b5), [`5fbda0bd2`](https://github.com/gradio-app/gradio/commit/5fbda0bd2b2bbb2282249b8875d54acf87cd7e84)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.9.0-beta.0
### Features
- [#5960](https://github.com/gradio-app/gradio/pull/5960) [`319c30f3f`](https://github.com/gradio-app/gradio/commit/319c30f3fccf23bfe1da6c9b132a6a99d59652f7) - rererefactor frontend files. Thanks [@pngwn](https://github.com/pngwn)!
- [#5956](https://github.com/gradio-app/gradio/pull/5956) [`f769876e0`](https://github.com/gradio-app/gradio/commit/f769876e0fa62336425c4e8ada5e09f38353ff01) - Apply formatter (and small refactoring) to the Lite-related frontend code. Thanks [@whitphx](https://github.com/whitphx)!
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`85ba6de13`](https://github.com/gradio-app/gradio/commit/85ba6de136a45b3e92c74e410bb27e3cbe7138d7) - Adds the ability to build the frontend and backend of custom components in preparation for publishing to pypi using `gradio_component build`. Thanks [@pngwn](https://github.com/pngwn)!
### Fixes
- [#5498](https://github.com/gradio-app/gradio/pull/5498) [`85ba6de13`](https://github.com/gradio-app/gradio/commit/85ba6de136a45b3e92c74e410bb27e3cbe7138d7) - Reinstate types that were removed in error in #5832. Thanks [@pngwn](https://github.com/pngwn)!
## 1.8.0
### Features
- [#5627](https://github.com/gradio-app/gradio/pull/5627) [`b67115e8e`](https://github.com/gradio-app/gradio/commit/b67115e8e6e489fffd5271ea830211863241ddc5) - Lite: Make the Examples component display media files using pseudo HTTP requests to the Wasm server. Thanks [@whitphx](https://github.com/whitphx)!
- [#5886](https://github.com/gradio-app/gradio/pull/5886) [`121f25b2d`](https://github.com/gradio-app/gradio/commit/121f25b2d50a33e1e06721b79e20b4f5651987ba) - Lite: Fix is_self_host() to detect `127.0.0.1` as localhost as well. Thanks [@whitphx](https://github.com/whitphx)!
## 1.7.1
### Patch Changes
- Updated dependencies [[`796145e2c`](https://github.com/gradio-app/gradio/commit/796145e2c48c4087bec17f8ec0be4ceee47170cb)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.7.0
### Highlights
#### new `FileExplorer` component ([#5672](https://github.com/gradio-app/gradio/pull/5672) [`e4a307ed6`](https://github.com/gradio-app/gradio/commit/e4a307ed6cde3bbdf4ff2f17655739addeec941e))
Thanks to a new capability that allows components to communicate directly with the server _without_ passing data via the value, we have created a new `FileExplorer` component.
This component allows you to populate the explorer by passing a glob, but only provides the selected file(s) in your prediction function.
Users can then navigate the virtual filesystem and select files which will be accessible in your predict function. This component will allow developers to build more complex spaces, with more flexible input options.
For more information check the [`FileExplorer` documentation](https://gradio.app/docs/fileexplorer).
Thanks [@aliabid94](https://github.com/aliabid94)!
### Fixes
- [#5794](https://github.com/gradio-app/gradio/pull/5794) [`f096c3ae1`](https://github.com/gradio-app/gradio/commit/f096c3ae168c0df00f90fe131c1e48c572e0574b) - Throw helpful error when media devices are not found. Thanks [@hannahblair](https://github.com/hannahblair)!
## 1.6.4
### Features
- [#5124](https://github.com/gradio-app/gradio/pull/5124) [`6e56a0d9b`](https://github.com/gradio-app/gradio/commit/6e56a0d9b0c863e76c69e1183d9d40196922b4cd) - Lite: Websocket queueing. Thanks [@whitphx](https://github.com/whitphx)!
## 1.6.3
### Patch Changes
- Updated dependencies [[`abb5e9df4`](https://github.com/gradio-app/gradio/commit/abb5e9df47989b2c56c2c312d74944678f9f2d4e), [`e842a561a`](https://github.com/gradio-app/gradio/commit/e842a561af4394f8109291ee5725bcf74743e816), [`8f0fed857`](https://github.com/gradio-app/gradio/commit/8f0fed857d156830626eb48b469d54d211a582d2), [`502054848`](https://github.com/gradio-app/gradio/commit/502054848fdbe39fc03ec42445242b4e49b7affc), [`2a5b9e03b`](https://github.com/gradio-app/gradio/commit/2a5b9e03b15ea324d641fe6982f26d81b1ca7210)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.6.2
### Features
- [#5721](https://github.com/gradio-app/gradio/pull/5721) [`84e03fe50`](https://github.com/gradio-app/gradio/commit/84e03fe506e08f1f81bac6d504c9fba7924f2d93) - Adds copy buttons to website, and better descriptions to API Docs. Thanks [@aliabd](https://github.com/aliabd)!
### Fixes
- [#5705](https://github.com/gradio-app/gradio/pull/5705) [`78e7cf516`](https://github.com/gradio-app/gradio/commit/78e7cf5163e8d205e8999428fce4c02dbdece25f) - ensure internal data has updated before dispatching `success` or `then` events. Thanks [@pngwn](https://github.com/pngwn)!
- [#5726](https://github.com/gradio-app/gradio/pull/5726) [`96c4b97c7`](https://github.com/gradio-app/gradio/commit/96c4b97c742311e90a87d8e8ee562c6ad765e9f0) - Adjust translation. Thanks [@ylhsieh](https://github.com/ylhsieh)!
## 1.6.1
### Patch Changes
- Updated dependencies [[`ee8eec1e5`](https://github.com/gradio-app/gradio/commit/ee8eec1e5e544a0127e0aa68c2522a7085b8ada5)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.6.0
### Features
- [#5639](https://github.com/gradio-app/gradio/pull/5639) [`e1874aff8`](https://github.com/gradio-app/gradio/commit/e1874aff814d13b23f3e59ef239cc13e18ad3fa7) - Add `gr.on` listener method. Thanks [@aliabid94](https://github.com/aliabid94)!
- [#5554](https://github.com/gradio-app/gradio/pull/5554) [`75ddeb390`](https://github.com/gradio-app/gradio/commit/75ddeb390d665d4484667390a97442081b49a423) - Accessibility Improvements. Thanks [@hannahblair](https://github.com/hannahblair)!
## 1.5.4
### Features
- [#5514](https://github.com/gradio-app/gradio/pull/5514) [`52f783175`](https://github.com/gradio-app/gradio/commit/52f7831751b432411e109bd41add4ab286023a8e) - refactor: Use package.json for version management. Thanks [@DarhkVoyd](https://github.com/DarhkVoyd)!
## 1.5.3
### Fixes
- [#5562](https://github.com/gradio-app/gradio/pull/5562) [`50d9747d0`](https://github.com/gradio-app/gradio/commit/50d9747d061962cff7f60a8da648bb3781794102) - chore(deps): update dependency iframe-resizer to v4.3.7. Thanks [@renovate](https://github.com/apps/renovate)!
- [#5550](https://github.com/gradio-app/gradio/pull/5550) [`4ed5902e7`](https://github.com/gradio-app/gradio/commit/4ed5902e7dda2d95cd43e4ccaaef520ddd8eba57) - Adding basque language. Thanks [@EkhiAzur](https://github.com/EkhiAzur)!
## 1.5.2
### Patch Changes
- Updated dependencies [[`a0cc9ac9`](https://github.com/gradio-app/gradio/commit/a0cc9ac931554e06dcb091158c9b9ac0cc580b6c)]:
- @gradio/[email protected]
## 1.5.1
### Patch Changes
- Updated dependencies [[`dc86e4a7`](https://github.com/gradio-app/gradio/commit/dc86e4a7e1c40b910c74558e6f88fddf9b3292bc), [`21f1db40`](https://github.com/gradio-app/gradio/commit/21f1db40de6d1717eba97a550e11422a457ba7e9)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.5.0
### Features
- [#5505](https://github.com/gradio-app/gradio/pull/5505) [`9ee20f49`](https://github.com/gradio-app/gradio/commit/9ee20f499f62c1fe5af6b8f84918b3a334eb1c8d) - Validate i18n file names with ISO-639x. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#5475](https://github.com/gradio-app/gradio/pull/5475) [`c60b89b0`](https://github.com/gradio-app/gradio/commit/c60b89b0a54758a27277f0a6aa20d0653647c7c8) - Adding Central Kurdish. Thanks [@Hrazhan](https://github.com/Hrazhan)!
- [#5400](https://github.com/gradio-app/gradio/pull/5400) [`d112e261`](https://github.com/gradio-app/gradio/commit/d112e2611b0fc79ecedfaed367571f3157211387) - Allow interactive input in `gr.HighlightedText`. Thanks [@hannahblair](https://github.com/hannahblair)!
## 1.4.3
### Patch Changes
- Updated dependencies [[`6e381c4f`](https://github.com/gradio-app/gradio/commit/6e381c4f146cc8177a4e2b8e39f914f09cd7ff0c)]:
- @gradio/[email protected]
## 1.4.2
### Fixes
- [#5447](https://github.com/gradio-app/gradio/pull/5447) [`7a4a89e5`](https://github.com/gradio-app/gradio/commit/7a4a89e5ca1dedb39e5366867501584b0c636bbb) - ensure iframe is correct size on spaces. Thanks [@pngwn](https://github.com/pngwn)!
## 1.4.1
### Patch Changes
- Updated dependencies [[`afac0006`](https://github.com/gradio-app/gradio/commit/afac0006337ce2840cf497cd65691f2f60ee5912), [`d14d63e3`](https://github.com/gradio-app/gradio/commit/d14d63e30c4af3f9c2a664fd11b0a01943a8300c), [`26fef8c7`](https://github.com/gradio-app/gradio/commit/26fef8c7f85a006c7e25cdbed1792df19c512d02)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.4.0
### Features
- [#5267](https://github.com/gradio-app/gradio/pull/5267) [`119c8343`](https://github.com/gradio-app/gradio/commit/119c834331bfae60d4742c8f20e9cdecdd67e8c2) - Faster reload mode. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#5373](https://github.com/gradio-app/gradio/pull/5373) [`79d8f9d8`](https://github.com/gradio-app/gradio/commit/79d8f9d891901683c5a1b7486efb44eab2478c96) - Adds `height` and `zoom_speed` parameters to `Model3D` component, as well as a button to reset the camera position. Thanks [@abidlabs](https://github.com/abidlabs)!
## 1.3.2
### Patch Changes
- Updated dependencies [[`5f25eb68`](https://github.com/gradio-app/gradio/commit/5f25eb6836f6a78ce6208b53495a01e1fc1a1d2f), [`3341148c`](https://github.com/gradio-app/gradio/commit/3341148c109b5458cc88435d27eb154210efc472), [`df090e89`](https://github.com/gradio-app/gradio/commit/df090e89f74a16e4cb2b700a1e3263cabd2bdd91)]:
- @gradio/[email protected]
- @gradio/[email protected]
- @gradio/[email protected]
## 1.3.1
### Fixes
- [#5324](https://github.com/gradio-app/gradio/pull/5324) [`31996c99`](https://github.com/gradio-app/gradio/commit/31996c991d6bfca8cef975eb8e3c9f61a7aced19) - ensure login form has correct styles. Thanks [@pngwn](https://github.com/pngwn)!
## 1.3.0
### Highlights
#### Improve startup performance and markdown support ([#5279](https://github.com/gradio-app/gradio/pull/5279) [`fe057300`](https://github.com/gradio-app/gradio/commit/fe057300f0672c62dab9d9b4501054ac5d45a4ec))
##### Improved markdown support
We now have better support for markdown in `gr.Markdown` and `gr.Dataframe`. Including syntax highlighting and Github Flavoured Markdown. We also have more consistent markdown behaviour and styling.
##### Various performance improvements
These improvements will be particularly beneficial to large applications.
- Rather than attaching events manually, they are now delegated, leading to a significant performance improvement and addressing a performance regression introduced in a recent version of Gradio. App startup for large applications is now around twice as fast.
- Optimised the mounting of individual components, leading to a modest performance improvement during startup (~30%).
- Corrected an issue that was causing markdown to re-render infinitely.
- Ensured that the `gr.3DModel` does re-render prematurely.
Thanks [@pngwn](https://github.com/pngwn)!
#### Add `render` function to `<gradio-app>` ([#5158](https://github.com/gradio-app/gradio/pull/5158) [`804fcc05`](https://github.com/gradio-app/gradio/commit/804fcc058e147f283ece67f1f353874e26235535))
We now have an event `render` on the <gradio-app> web component, which is triggered once the embedded space has finished rendering.
```html
<script>
function handleLoadComplete() {
console.log("Embedded space has finished rendering");
}
const gradioApp = document.querySelector("gradio-app");
gradioApp.addEventListener("render", handleLoadComplete);
</script>
```
Thanks [@hannahblair](https://github.com/hannahblair)!
### Features
- [#5215](https://github.com/gradio-app/gradio/pull/5215) [`fbdad78a`](https://github.com/gradio-app/gradio/commit/fbdad78af4c47454cbb570f88cc14bf4479bbceb) - Lazy load interactive or static variants of a component individually, rather than loading both variants regardless. This change will improve performance for many applications. Thanks [@pngwn](https://github.com/pngwn)!
- [#5216](https://github.com/gradio-app/gradio/pull/5216) [`4b58ea6d`](https://github.com/gradio-app/gradio/commit/4b58ea6d98e7a43b3f30d8a4cb6f379bc2eca6a8) - Update i18n tokens and locale files. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#5219](https://github.com/gradio-app/gradio/pull/5219) [`e8fd4e4e`](https://github.com/gradio-app/gradio/commit/e8fd4e4ec68a6c974bc8c84b61f4a0ec50a85bc6) - Add `api_name` parameter to `gr.Interface`. Additionally, completely hide api page if show_api=False. Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
- [#5264](https://github.com/gradio-app/gradio/pull/5264) [`46a2b600`](https://github.com/gradio-app/gradio/commit/46a2b600a7ff030a9ea1560b882b3bf3ad266bbc) - ensure translations for audio work correctly. Thanks [@hannahblair](https://github.com/hannahblair)!
### Fixes
- [#5285](https://github.com/gradio-app/gradio/pull/5285) [`cdfd4217`](https://github.com/gradio-app/gradio/commit/cdfd42174a9c777eaee9c1209bf8e90d8c7791f2) - Tweaks to `icon` parameter in `gr.Button()`. Thanks [@abidlabs](https://github.com/abidlabs)!
- [#5312](https://github.com/gradio-app/gradio/pull/5312) [`f769cb67`](https://github.com/gradio-app/gradio/commit/f769cb67149d8e209091508f06d87014acaed965) - only start listening for events after the components are mounted. Thanks [@pngwn](https://github.com/pngwn)!
- [#5276](https://github.com/gradio-app/gradio/pull/5276) [`502f1015`](https://github.com/gradio-app/gradio/commit/502f1015bf23b365bc32446dd2e549b0c5d0dc72) - Ensure `Blocks` translation copy renders correctly. Thanks [@hannahblair](https://github.com/hannahblair)!
## 1.2.0
### Highlights
#### Client.predict will now return the final output for streaming endpoints ([#5057](https://github.com/gradio-app/gradio/pull/5057) [`35856f8b`](https://github.com/gradio-app/gradio/commit/35856f8b54548cae7bd3b8d6a4de69e1748283b2))
### This is a breaking change (for gradio_client only)!
Previously, `Client.predict` would only return the first output of an endpoint that streamed results. This was causing confusion for developers that wanted to call these streaming demos via the client.
We realize that developers using the client don't know the internals of whether a demo streams or not, so we're changing the behavior of predict to match developer expectations.
Using `Client.predict` will now return the final output of a streaming endpoint. This will make it even easier to use gradio apps via the client.
Thanks [@freddyaboulton](https://github.com/freddyaboulton)!
### Features
- [#5025](https://github.com/gradio-app/gradio/pull/5025) [`6693660a`](https://github.com/gradio-app/gradio/commit/6693660a790996f8f481feaf22a8c49130d52d89) - Add download button to selected images in `Gallery`. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#5046](https://github.com/gradio-app/gradio/pull/5046) [`5244c587`](https://github.com/gradio-app/gradio/commit/5244c5873c355cf3e2f0acb7d67fda3177ef8b0b) - Allow new lines in `HighlightedText` with `/n` and preserve whitespace. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#5047](https://github.com/gradio-app/gradio/pull/5047) [`883ac364`](https://github.com/gradio-app/gradio/commit/883ac364f69d92128774ac446ce49bdf8415fd7b) - Add `step` param to `Number`. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#5005](https://github.com/gradio-app/gradio/pull/5005) [`f5539c76`](https://github.com/gradio-app/gradio/commit/f5539c7618e31451420bd3228754774da14dc65f) - Enhancement: Add focus event to textbox and number component. Thanks [@JodyZ0203](https://github.com/JodyZ0203)!
- [#5136](https://github.com/gradio-app/gradio/pull/5136) [`eaa1ce14`](https://github.com/gradio-app/gradio/commit/eaa1ce14ac41de1c23321e93f11f1b03a2f3c7f4) - Enhancing Tamil Translation: Language Refinement 🌟. Thanks [@sanjaiyan-dev](https://github.com/sanjaiyan-dev)!
## 1.1.0
### Features
- [#4995](https://github.com/gradio-app/gradio/pull/4995) [`3f8c210b`](https://github.com/gradio-app/gradio/commit/3f8c210b01ef1ceaaf8ee73be4bf246b5b745bbf) - Implement left and right click in `Gallery` component and show implicit images in `Gallery` grid. Thanks [@hannahblair](https://github.com/hannahblair)!
- [#4993](https://github.com/gradio-app/gradio/pull/4993) [`dc07a9f9`](https://github.com/gradio-app/gradio/commit/dc07a9f947de44b419d8384987a02dcf94977851) - Bringing back the "Add download button for audio" PR by [@leuryr](https://github.com/leuryr). Thanks [@abidlabs](https://github.com/abidlabs)!
- [#4979](https://github.com/gradio-app/gradio/pull/4979) [`44ac8ad0`](https://github.com/gradio-app/gradio/commit/44ac8ad08d82ea12c503dde5c78f999eb0452de2) - Allow setting sketch color default. Thanks [@aliabid94](https://github.com/aliabid94)! | gradio-app/gradio/blob/main/js/app/CHANGELOG.md |
VisualBERT Demo
This demo shows usage of VisualBERT VQA model and is adapted from LXMERT demo present [here](https://github.com/huggingface/transformers/blob/main/examples/research_projects/lxmert/demo.ipynb).
1. make a virtualenv: ``virtualenv venv`` and activate ``source venv/bin/activate``
2. install reqs: ``pip install -r ./requirements.txt``
3. usage is as shown in demo.ipynb
| huggingface/transformers/blob/main/examples/research_projects/visual_bert/README.md |
Quantization
Quantization is a technique to reduce the computational and memory costs of running inference by representing the
weights and activations with low-precision data types like 8-bit integer (`int8`) instead of the usual 32-bit floating
point (`float32`).
Reducing the number of bits means the resulting model requires less memory storage, consumes less energy (in theory), and
operations like matrix multiplication can be performed much faster with integer arithmetic. It also allows to run models
on embedded devices, which sometimes only support integer data types.
## Theory
The basic idea behind quantization is quite easy: going from high-precision representation (usually the regular 32-bit
floating-point) for weights and activations to a lower precision data type. The most common lower precision data types
are:
- `float16`, accumulation data type `float16`
- `bfloat16`, accumulation data type `float32`
- `int16`, accumulation data type `int32`
- `int8`, accumulation data type `int32`
The accumulation data type specifies the type of the result of accumulating (adding, multiplying, etc) values of the
data type in question. For example, let's consider two `int8` values `A = 127`, `B = 127`, and let's define `C` as the
sum of `A` and `B`:
```
C = A + B
```
Here the result is much bigger than the biggest representable value in `int8`, which is `127`. Hence the need for a larger
precision data type to avoid a huge precision loss that would make the whole quantization process useless.
## Quantization
The two most common quantization cases are `float32 -> float16` and `float32 -> int8`.
### Quantization to `float16`
Performing quantization to go from `float32` to `float16` is quite straightforward since both data types follow the same
representation scheme. The questions to ask yourself when quantizing an operation to `float16` are:
- Does my operation have a `float16` implementation?
- Does my hardware suport `float16`? For instance, Intel CPUs [have been supporting `float16` as a storage type, but
computation is done after converting to `float32`](https://scicomp.stackexchange.com/a/35193). Full support will come
in Cooper Lake and Sapphire Rapids.
- Is my operation sensitive to lower precision?
For instance the value of epsilon in `LayerNorm` is usually very small (~ `1e-12`), but the smallest representable value in
`float16` is ~ `6e-5`, this can cause `NaN` issues. The same applies for big values.
### Quantization to `int8`
Performing quantization to go from `float32` to `int8` is more tricky. Only 256 values can be represented in `int8`,
while `float32` can represent a very wide range of values. The idea is to find the best way to project our range `[a, b]`
of `float32` values to the `int8` space.
Let's consider a float `x` in `[a, b]`, then we can write the following quantization scheme, also called the *affine
quantization scheme*:
```
x = S * (x_q - Z)
```
where:
- `x_q` is the quantized `int8` value associated to `x`
- `S` and `Z` are the quantization parameters
- `S` is the scale, and is a positive `float32`
- `Z` is called the zero-point, it is the `int8` value corresponding to the value `0` in the `float32` realm. This is
important to be able to represent exactly the value `0` because it is used everywhere throughout machine learning
models.
The quantized value `x_q` of `x` in `[a, b]` can be computed as follows:
```
x_q = round(x/S + Z)
```
And `float32` values outside of the `[a, b]` range are clipped to the closest representable value, so for any
floating-point number `x`:
```
x_q = clip(round(x/S + Z), round(a/S + Z), round(b/S + Z))
```
<Tip>
Usually `round(a/S + Z)` corresponds to the smallest representable value in the considered data type, and `round(b/S + Z)`
to the biggest one. But this can vary, for instance when using a *symmetric quantization scheme* as you will see in the next
paragraph.
</Tip>
### Symmetric and affine quantization schemes
The equation above is called the *affine quantization sheme* because the mapping from `[a, b]` to `int8` is an affine one.
A common special case of this scheme is the *symmetric quantization scheme*, where we consider a symmetric range of float values `[-a, a]`.
In this case the integer space is usally `[-127, 127]`, meaning that the `-128` is opted out of the regular `[-128, 127]` signed `int8` range.
The reason being that having both ranges symmetric allows to have `Z = 0`. While one value out of the 256 representable
values is lost, it can provide a speedup since a lot of addition operations can be skipped.
**Note**: To learn how the quantization parameters `S` and `Z` are computed, you can read the
[Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference](https://arxiv.org/abs/1712.05877)
paper, or [Lei Mao's blog post](https://leimao.github.io/article/Neural-Networks-Quantization) on the subject.
### Per-tensor and per-channel quantization
Depending on the accuracy / latency trade-off you are targetting you can play with the granularity of the quantization parameters:
- Quantization parameters can be computed on a *per-tensor* basis, meaning that one pair of `(S, Z)` will be used per
tensor.
- Quantization parameters can be computed on a *per-channel* basis, meaning that it is possible to store a pair of
`(S, Z)` per element along one of the dimensions of a tensor. For example for a tensor of shape `[N, C, H, W]`, having
*per-channel* quantization parameters for the second dimension would result in having `C` pairs of `(S, Z)`. While this
can give a better accuracy, it requires more memory.
### Calibration
The section above described how quantization from `float32` to `int8` works, but one question
remains: how is the `[a, b]` range of `float32` values determined? That is where calibration comes in to play.
Calibration is the step during quantization where the `float32` ranges are computed. For weights it is quite easy since
the actual range is known at *quantization-time*. But it is less clear for activations, and different approaches exist:
1. Post training **dynamic quantization**: the range for each activation is computed on the fly at *runtime*. While this
gives great results without too much work, it can be a bit slower than static quantization because of the overhead
introduced by computing the range each time.
It is also not an option on certain hardware.
2. Post training **static quantization**: the range for each activation is computed in advance at *quantization-time*,
typically by passing representative data through the model and recording the activation values. In practice, the steps are:
1. Observers are put on activations to record their values.
2. A certain number of forward passes on a calibration dataset is done (around `200` examples is enough).
3. The ranges for each computation are computed according to some *calibration technique*.
3. **Quantization aware training**: the range for each activation is computed at *training-time*, following the same idea
than post training static quantization. But "fake quantize" operators are used instead of observers: they record
values just as observers do, but they also simulate the error induced by quantization to let the model adapt to it.
For both post training static quantization and quantization aware training, it is necessary to define calibration
techniques, the most common are:
- Min-max: the computed range is `[min observed value, max observed value]`, this works well with weights.
- Moving average min-max: the computed range is `[moving average min observed value, moving average max observed value]`,
this works well with activations.
- Histogram: records a histogram of values along with min and max values, then chooses according to some criterion:
- Entropy: the range is computed as the one minimizing the error between the full-precision and the quantized data.
- Mean Square Error: the range is computed as the one minimizing the mean square error between the full-precision and
the quantized data.
- Percentile: the range is computed using a given percentile value `p` on the observed values. The idea is to try to have
`p%` of the observed values in the computed range. While this is possible when doing affine quantization, it is not always
possible to exactly match that when doing symmetric quantization. You can check [how it is done in ONNX
Runtime](https://github.com/microsoft/onnxruntime/blob/2cb12caf9317f1ded37f6db125cb03ba99320c40/onnxruntime/python/tools/quantization/calibrate.py#L698)
for more details.
### Pratical steps to follow to quantize a model to `int8`
To effectively quantize a model to `int8`, the steps to follow are:
1. Choose which operators to quantize. Good operators to quantize are the one dominating it terms of computation time,
for instance linear projections and matrix multiplications.
2. Try post-training dynamic quantization, if it is fast enough stop here, otherwise continue to step 3.
3. Try post-training static quantization which can be faster than dynamic quantization but often with a drop in terms of
accuracy. Apply observers to your models in places where you want to quantize.
4. Choose a calibration technique and perform it.
5. Convert the model to its quantized form: the observers are removed and the `float32` operators are converted to their `int8`
coutnerparts.
6. Evaluate the quantized model: is the accuracy good enough? If yes, stop here, otherwise start again at step 3 but
with quantization aware training this time.
## Supported tools to perform quantization in 🤗 Optimum
🤗 Optimum provides APIs to perform quantization using different tools for different targets:
- The `optimum.onnxruntime` package allows to
[quantize and run ONNX models](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/quantization) using the
ONNX Runtime tool.
- The `optimum.intel` package enables to [quantize](https://huggingface.co/docs/optimum/intel/optimization_inc) 🤗 Transformers
models while respecting accuracy and latency constraints.
- The `optimum.fx` package provides wrappers around the
[PyTorch quantization functions](https://pytorch.org/docs/stable/quantization-support.html#torch-quantization-quantize-fx)
to allow graph-mode quantization of 🤗 Transformers models in PyTorch. This is a lower-level API compared to the two
mentioned above, giving more flexibility, but requiring more work on your end.
- The `optimum.gptq` package allows to [quantize and run LLM models](../llm_quantization/usage_guides/quantization) with GPTQ.
## Going further: How do machines represent numbers?
<Tip>
The section is not fundamental to understand the rest. It explains in brief how numbers are represented in computers.
Since quantization is about going from one representation to another, it can be useful to have some basics, but it is
definitely not mandatory.
</Tip>
The most fundamental unit of representation for computers is the bit. Everything in computers is represented as a
sequence of bits, including numbers. But the representation varies whether the numbers in question are integers or
real numbers.
#### Integer representation
Integers are usually represented with the following bit lengths: `8`, `16`, `32`, `64`. When representing integers, two cases
are considered:
1. Unsigned (positive) integers: they are simply represented as a sequence of bits. Each bit corresponds to a power
of two (from `0` to `n-1` where `n` is the bit-length), and the resulting number is the sum of those powers of two.
Example: `19` is represented as an unsigned int8 as `00010011` because :
```
19 = 0 x 2^7 + 0 x 2^6 + 0 x 2^5 + 1 x 2^4 + 0 x 2^3 + 0 x 2^2 + 1 x 2^1 + 1 x 2^0
```
2. Signed integers: it is less straightforward to represent signed integers, and multiple approachs exist, the most
common being the *two's complement*. For more information, you can check the
[Wikipedia page](https://en.wikipedia.org/wiki/Signed_number_representations) on the subject.
#### Real numbers representation
Real numbers are usually represented with the following bit lengths: `16`, `32`, `64`.
The two main ways of representing real numbers are:
1. Fixed-point: there are fixed number of digits reserved for representing the integer part and the fractional part.
2. Floating-point: the number of digits for representing the integer and the fractional parts can vary.
The floating-point representation can represent bigger ranges of values, and this is the one we will be focusing on
since it is the most commonly used. There are three components in the floating-point representation:
1. The sign bit: this is the bit specifying the sign of the number.
2. The exponent part
- 5 bits in `float16`
- 8 bits in `bfloat16`
- 8 bits in `float32`
- 11 bits in `float64`
2. The mantissa
- 11 bits in `float16` (10 explictly stored)
- 8 bits in `bfloat16` (7 explicitly stored)
- 24 bits in `float32` (23 explictly stored)
- 53 bits in `float64` (52 explictly stored)
For more information on the bits allocation for each data type, check the nice illustration on the Wikipedia page about
the [bfloat16 floating-point format](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format).
For a real number `x` we have:
```
x = sign x mantissa x (2^exponent)
```
## References
- The
[Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference](https://arxiv.org/abs/1712.05877) paper
- The [Basics of Quantization in Machine Learning (ML) for Beginners](https://iq.opengenus.org/basics-of-quantization-in-ml/)
blog post
- The [How to accelerate and compress neural networks with quantization](https://tivadardanka.com/blog/neural-networks-quantization)
blog post
- The Wikipedia pages on integers representation [here](https://en.wikipedia.org/wiki/Integer_(computer_science)) and
[here](https://en.wikipedia.org/wiki/Signed_number_representations)
- The Wikipedia pages on
- [bfloat16 floating-point format](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format)
- [Half-precision floating-point format](https://en.wikipedia.org/wiki/Half-precision_floating-point_format)
- [Single-precision floating-point format](https://en.wikipedia.org/wiki/Single-precision_floating-point_format)
- [Double-precision floating-point format](https://en.wikipedia.org/wiki/Double-precision_floating-point_format)
| huggingface/optimum/blob/main/docs/source/concept_guides/quantization.mdx |
!--Copyright 2021 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# VisualBERT
## Overview
The VisualBERT model was proposed in [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
VisualBERT is a neural network trained on a variety of (image, text) pairs.
The abstract from the paper is the following:
*We propose VisualBERT, a simple and flexible framework for modeling a broad range of vision-and-language tasks.
VisualBERT consists of a stack of Transformer layers that implicitly align elements of an input text and regions in an
associated input image with self-attention. We further propose two visually-grounded language model objectives for
pre-training VisualBERT on image caption data. Experiments on four vision-and-language tasks including VQA, VCR, NLVR2,
and Flickr30K show that VisualBERT outperforms or rivals with state-of-the-art models while being significantly
simpler. Further analysis demonstrates that VisualBERT can ground elements of language to image regions without any
explicit supervision and is even sensitive to syntactic relationships, tracking, for example, associations between
verbs and image regions corresponding to their arguments.*
This model was contributed by [gchhablani](https://huggingface.co/gchhablani). The original code can be found [here](https://github.com/uclanlp/visualbert).
## Usage tips
1. Most of the checkpoints provided work with the [`VisualBertForPreTraining`] configuration. Other
checkpoints provided are the fine-tuned checkpoints for down-stream tasks - VQA ('visualbert-vqa'), VCR
('visualbert-vcr'), NLVR2 ('visualbert-nlvr2'). Hence, if you are not working on these downstream tasks, it is
recommended that you use the pretrained checkpoints.
2. For the VCR task, the authors use a fine-tuned detector for generating visual embeddings, for all the checkpoints.
We do not provide the detector and its weights as a part of the package, but it will be available in the research
projects, and the states can be loaded directly into the detector provided.
VisualBERT is a multi-modal vision and language model. It can be used for visual question answering, multiple choice,
visual reasoning and region-to-phrase correspondence tasks. VisualBERT uses a BERT-like transformer to prepare
embeddings for image-text pairs. Both the text and visual features are then projected to a latent space with identical
dimension.
To feed images to the model, each image is passed through a pre-trained object detector and the regions and the
bounding boxes are extracted. The authors use the features generated after passing these regions through a pre-trained
CNN like ResNet as visual embeddings. They also add absolute position embeddings, and feed the resulting sequence of
vectors to a standard BERT model. The text input is concatenated in the front of the visual embeddings in the embedding
layer, and is expected to be bound by [CLS] and a [SEP] tokens, as in BERT. The segment IDs must also be set
appropriately for the textual and visual parts.
The [`BertTokenizer`] is used to encode the text. A custom detector/image processor must be used
to get the visual embeddings. The following example notebooks show how to use VisualBERT with Detectron-like models:
- [VisualBERT VQA demo notebook](https://github.com/huggingface/transformers/tree/main/examples/research_projects/visual_bert) : This notebook
contains an example on VisualBERT VQA.
- [Generate Embeddings for VisualBERT (Colab Notebook)](https://colab.research.google.com/drive/1bLGxKdldwqnMVA5x4neY7-l_8fKGWQYI?usp=sharing) : This notebook contains
an example on how to generate visual embeddings.
The following example shows how to get the last hidden state using [`VisualBertModel`]:
```python
>>> import torch
>>> from transformers import BertTokenizer, VisualBertModel
>>> model = VisualBertModel.from_pretrained("uclanlp/visualbert-vqa-coco-pre")
>>> tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
>>> inputs = tokenizer("What is the man eating?", return_tensors="pt")
>>> # this is a custom function that returns the visual embeddings given the image path
>>> visual_embeds = get_visual_embeddings(image_path)
>>> visual_token_type_ids = torch.ones(visual_embeds.shape[:-1], dtype=torch.long)
>>> visual_attention_mask = torch.ones(visual_embeds.shape[:-1], dtype=torch.float)
>>> inputs.update(
... {
... "visual_embeds": visual_embeds,
... "visual_token_type_ids": visual_token_type_ids,
... "visual_attention_mask": visual_attention_mask,
... }
... )
>>> outputs = model(**inputs)
>>> last_hidden_state = outputs.last_hidden_state
```
## VisualBertConfig
[[autodoc]] VisualBertConfig
## VisualBertModel
[[autodoc]] VisualBertModel
- forward
## VisualBertForPreTraining
[[autodoc]] VisualBertForPreTraining
- forward
## VisualBertForQuestionAnswering
[[autodoc]] VisualBertForQuestionAnswering
- forward
## VisualBertForMultipleChoice
[[autodoc]] VisualBertForMultipleChoice
- forward
## VisualBertForVisualReasoning
[[autodoc]] VisualBertForVisualReasoning
- forward
## VisualBertForRegionToPhraseAlignment
[[autodoc]] VisualBertForRegionToPhraseAlignment
- forward
| huggingface/transformers/blob/main/docs/source/en/model_doc/visual_bert.md |
Datasets server - reverse proxy
> Reverse-proxy in front of the API
See [docker-compose-datasets-server.yml](../../tools/docker-compose-datasets-server.yml) for usage.
Note that the template configuration is located in [chart/nginx-templates/](../../chart/nginx-templates/) in order to be reachable by the Helm chart to deploy on Kubernetes.
The reverse proxy uses nginx:
- it serves the static assets directly (the API also serves them if required, but it's unnecessary to go through starlette for this, and it generates errors in Safari, see [1](https://github.com/encode/starlette/issues/950) and [2](https://developer.apple.com/library/archive/documentation/AppleApplications/Reference/SafariWebContent/CreatingVideoforSafarioniPhone/CreatingVideoforSafarioniPhone.html#//apple_ref/doc/uid/TP40006514-SW6))
- it serves the OpenAPI specification
- it proxies the other requests to the API
It takes various environment variables, all of them are mandatory:
- `ASSETS_DIRECTORY`: the directory that contains the static assets, eg `/assets`
- `CACHED_ASSETS_DIRECTORY`: the directory that contains the static cached assets, eg `/cached-assets`
- `OPENAPI_FILE`: the path to the OpenAPI file, eg `docs/source/openapi.json`
- `HOST`: domain of the reverse proxy, eg `localhost`
- `PORT`: port of the reverse proxy, eg `80`
- `URL_ADMIN`= URL of the admin, eg `http://admin:8081`
- `URL_API`= URL of the API, eg `http://api:8080`
- `URL_ROWS`= URL of the rows service, eg `http://rows:8082`
- `URL_SEARCH`= URL of the search service, eg `http://search:8083`
- `URL_SSE_API`= URL of the SSE API service, eg `http://sse-api:8085`
The image requires three directories to be mounted (from volumes):
- `$ASSETS_DIRECTORY` (read-only): the directory that contains the static assets.
- `/etc/nginx/templates` (read-only): the directory that contains the nginx configuration template ([templates](./templates/))
| huggingface/datasets-server/blob/main/services/reverse-proxy/README.md |
Datasets
<img class="float-left !m-0 !border-0 !dark:border-0 !shadow-none !max-w-lg w-[150px]" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/datasets/datasets_logo.png"/>
🤗 Datasets is a library for easily accessing and sharing datasets for Audio, Computer Vision, and Natural Language Processing (NLP) tasks.
Load a dataset in a single line of code, and use our powerful data processing methods to quickly get your dataset ready for training in a deep learning model. Backed by the Apache Arrow format, process large datasets with zero-copy reads without any memory constraints for optimal speed and efficiency. We also feature a deep integration with the [Hugging Face Hub](https://huggingface.co/datasets), allowing you to easily load and share a dataset with the wider machine learning community.
Find your dataset today on the [Hugging Face Hub](https://huggingface.co/datasets), and take an in-depth look inside of it with the live viewer.
<div class="mt-10">
<div class="w-full flex flex-col space-y-4 md:space-y-0 md:grid md:grid-cols-2 md:gap-y-4 md:gap-x-5">
<a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./tutorial"
><div class="w-full text-center bg-gradient-to-br from-blue-400 to-blue-500 rounded-lg py-1.5 font-semibold mb-5 text-white text-lg leading-relaxed">Tutorials</div>
<p class="text-gray-700">Learn the basics and become familiar with loading, accessing, and processing a dataset. Start here if you are using 🤗 Datasets for the first time!</p>
</a>
<a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./how_to"
><div class="w-full text-center bg-gradient-to-br from-indigo-400 to-indigo-500 rounded-lg py-1.5 font-semibold mb-5 text-white text-lg leading-relaxed">How-to guides</div>
<p class="text-gray-700">Practical guides to help you achieve a specific goal. Take a look at these guides to learn how to use 🤗 Datasets to solve real-world problems.</p>
</a>
<a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./about_arrow"
><div class="w-full text-center bg-gradient-to-br from-pink-400 to-pink-500 rounded-lg py-1.5 font-semibold mb-5 text-white text-lg leading-relaxed">Conceptual guides</div>
<p class="text-gray-700">High-level explanations for building a better understanding about important topics such as the underlying data format, the cache, and how datasets are generated.</p>
</a>
<a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="./package_reference/main_classes"
><div class="w-full text-center bg-gradient-to-br from-purple-400 to-purple-500 rounded-lg py-1.5 font-semibold mb-5 text-white text-lg leading-relaxed">Reference</div>
<p class="text-gray-700">Technical descriptions of how 🤗 Datasets classes and methods work.</p>
</a>
</div>
</div>
| huggingface/datasets/blob/main/docs/source/index.mdx |
--
title: "Chat Templates: An End to the Silent Performance Killer"
thumbnail: /blog/assets/chat-templates/thumbnail.png
authors:
- user: rocketknight1
---
# Chat Templates
> *A spectre is haunting chat models - the spectre of incorrect formatting!*
## tl;dr
Chat models have been trained with very different formats for converting conversations into a single tokenizable string. Using a format different from the format a model was trained with will usually cause severe, silent performance degradation, so matching the format used during training is extremely important! Hugging Face tokenizers now have a `chat_template` attribute that can be used to save the chat format the model was trained with. This attribute contains a Jinja template that converts conversation histories into a correctly formatted string. Please see the [technical documentation](https://huggingface.co/docs/transformers/main/en/chat_templating) for information on how to write and apply chat templates in your code.
## Introduction
If you're familiar with the 🤗 Transformers library, you've probably written code like this:
```python
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModel.from_pretrained(checkpoint)
```
By loading the tokenizer and model from the same checkpoint, you ensure that inputs are tokenized
in the way the model expects. If you pick a tokenizer from a different model, the input tokenization
might be completely different, and the result will be that your model's performance will be seriously damaged. The term for this is a **distribution shift** - the model has been learning data from one distribution (the tokenization it was trained with), and suddenly it has shifted to a completely different one.
Whether you're fine-tuning a model or using it directly for inference, it's always a good idea to minimize these distribution shifts and keep the input you give it as similar as possible to the input it was trained on. With regular language models, it's relatively easy to do that - simply load your tokenizer and model from the same checkpoint, and you're good to go.
With chat models, however, it's a bit different. This is because "chat" is not just a single string of text that can be straightforwardly tokenized - it's a sequence of messages, each of which contains a `role` as well as `content`, which is the actual text of the message. Most commonly, the roles are "user" for messages sent by the user, "assistant" for responses written by the model, and optionally "system" for high-level directives given at the start of the conversation.
If that all seems a bit abstract, here's an example chat to make it more concrete:
```python
[
{"role": "user", "content": "Hi there!"},
{"role": "assistant", "content": "Nice to meet you!"}
]
```
This sequence of messages needs to be converted into a text string before it can be tokenized and used as input to a model. The problem, though, is that there are many ways to do this conversion! You could, for example, convert the list of messages into an "instant messenger" format:
```
User: Hey there!
Bot: Nice to meet you!
```
Or you could add special tokens to indicate the roles:
```
[USER] Hey there! [/USER]
[ASST] Nice to meet you! [/ASST]
```
Or you could add tokens to indicate the boundaries between messages, but insert the role information as a string:
```
<|im_start|>user
Hey there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
```
There are lots of ways to do this, and none of them is obviously the best or correct way to do it. As a result, different models have been trained with wildly different formatting. I didn't make these examples up; they're all real and being used by at least one active model! But once a model has been trained with a certain format, you really want to ensure that future inputs use the same format, or else you could get a performance-destroying distribution shift.
## Templates: A way to save format information
Right now, if you're lucky, the format you need is correctly documented somewhere in the model card. If you're unlucky, it isn't, so good luck if you want to use that model. In extreme cases, we've even put the whole prompt format in [a blog post](https://huggingface.co/blog/llama2#how-to-prompt-llama-2) to ensure that users don't miss it! Even in the best-case scenario, though, you have to locate the template information and manually code it up in your fine-tuning or inference pipeline. We think this is an especially dangerous issue because using the wrong chat format is a **silent error** - you won't get a loud failure or a Python exception to tell you something is wrong, the model will just perform much worse than it would have with the right format, and it'll be very difficult to debug the cause!
This is the problem that **chat templates** aim to solve. Chat templates are [Jinja template strings](https://jinja.palletsprojects.com/en/3.1.x/) that are saved and loaded with your tokenizer, and that contain all the information needed to turn a list of chat messages into a correctly formatted input for your model. Here are three chat template strings, corresponding to the three message formats above:
```jinja
{% for message in messages %}
{% if message['role'] == 'user' %}
{{ "User : " }}
{% else %}
{{ "Bot : " }}
{{ message['content'] + '\n' }}
{% endfor %}
```
```jinja
{% for message in messages %}
{% if message['role'] == 'user' %}
{{ "[USER] " + message['content'] + " [/USER]" }}
{% else %}
{{ "[ASST] " + message['content'] + " [/ASST]" }}
{{ message['content'] + '\n' }}
{% endfor %}
```
```jinja
"{% for message in messages %}"
"{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}"
"{% endfor %}"
```
If you're unfamiliar with Jinja, I strongly recommend that you take a moment to look at these template strings, and their corresponding template outputs, and see if you can convince yourself that you understand how the template turns a list of messages into a formatted string! The syntax is very similar to Python in a lot of ways.
## Why templates?
Although Jinja can be confusing at first if you're unfamiliar with it, in practice we find that Python programmers can pick it up quickly. During development of this feature, we considered other approaches, such as a limited system to allow users to specify per-role prefixes and suffixes for messages. We found that this could become confusing and unwieldy, and was so inflexible that hacky workarounds were needed for several models. Templating, on the other hand, is powerful enough to cleanly support all of the message formats that we're aware of.
## Why bother doing this? Why not just pick a standard format?
This is an excellent idea! Unfortunately, it's too late, because multiple important models have already been trained with very different chat formats.
However, we can still mitigate this problem a bit. We think the closest thing to a 'standard' for formatting is the [ChatML format](https://github.com/openai/openai-python/blob/main/chatml.md) created by OpenAI. If you're training a new model for chat, and this format is suitable for you, we recommend using it and adding special `<|im_start|>` and `<|im_end|>` tokens to your tokenizer. It has the advantage of being very flexible with roles, as the role is just inserted as a string rather than having specific role tokens. If you'd like to use this one, it's the third of the templates above, and you can set it with this simple one-liner:
```py
tokenizer.chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}"
```
There's also a second reason not to hardcode a standard format, though, beyond the proliferation of existing formats - we expect that templates will be broadly useful in preprocessing for many types of models, including those that might be doing very different things from standard chat. Hardcoding a standard format limits the ability of model developers to use this feature to do things we haven't even thought of yet, whereas templating gives users and developers maximum freedom. It's even possible to encode checks and logic in templates, which is a feature we don't use extensively in any of the default templates, but which we expect to have enormous power in the hands of adventurous users. We strongly believe that the open-source ecosystem should enable you to do what you want, not dictate to you what you're permitted to do.
## How do templates work?
Chat templates are part of the **tokenizer**, because they fulfill the same role as tokenizers do: They store information about how data is preprocessed, to ensure that you feed data to the model in the same format that it saw during training. We have designed it to be very easy to add template information to an existing tokenizer and save it or upload it to the Hub.
Before chat templates, chat formatting information was stored at the **class level** - this meant that, for example, all LLaMA checkpoints would get the same chat formatting, using code that was hardcoded in `transformers` for the LLaMA model class. For backward compatibility, model classes that had custom chat format methods have been given **default chat templates** instead.
Default chat templates are also set at the class level, and tell classes like `ConversationPipeline` how to format inputs when the model does not have a chat template. We're doing this **purely for backwards compatibility** - we highly recommend that you explicitly set a chat template on any chat model, even when the default chat template is appropriate. This ensures that any future changes or deprecations in the default chat template don't break your model. Although we will be keeping default chat templates for the foreseeable future, we hope to transition all models to explicit chat templates over time, at which point the default chat templates may be removed entirely.
For information about how to set and apply chat templates, please see the [technical documentation](https://huggingface.co/docs/transformers/main/en/chat_templating).
## How do I get started with templates?
Easy! If a tokenizer has the `chat_template` attribute set, it's ready to go. You can use that model and tokenizer in `ConversationPipeline`, or you can call `tokenizer.apply_chat_template()` to format chats for inference or training. Please see our [developer guide](https://huggingface.co/docs/transformers/main/en/chat_templating) or the [apply_chat_template documentation](https://huggingface.co/docs/transformers/main/en/internal/tokenization_utils#transformers.PreTrainedTokenizerBase.apply_chat_template) for more!
If a tokenizer doesn't have a `chat_template` attribute, it might still work, but it will use the default chat template set for that model class. This is fragile, as we mentioned above, and it's also a source of silent bugs when the class template doesn't match what the model was actually trained with. If you want to use a checkpoint that doesn't have a `chat_template`, we recommend checking docs like the model card to verify what the right format is, and then adding a correct `chat_template`for that format. We recommend doing this even if the default chat template is correct - it future-proofs the model, and also makes it clear that the template is present and suitable.
You can add a `chat_template` even for checkpoints that you're not the owner of, by opening a [pull request](https://huggingface.co/docs/hub/repositories-pull-requests-discussions). The only change you need to make is to set the `tokenizer.chat_template` attribute to a Jinja template string. Once that's done, push your changes and you're ready to go!
If you'd like to use a checkpoint for chat but you can't find any documentation on the chat format it used, you should probably open an issue on the checkpoint or ping the owner! Once you figure out the format the model is using, please open a pull request to add a suitable `chat_template`. Other users will really appreciate it!
## Conclusion: Template philosophy
We think templates are a very exciting change. In addition to resolving a huge source of silent, performance-killing bugs, we think they open up completely new approaches and data modalities. Perhaps most importantly, they also represent a philosophical shift: They take a big function out of the core `transformers` codebase and move it into individual model repos, where users have the freedom to do weird and wild and wonderful things. We're excited to see what uses you find for them!
| huggingface/blog/blob/main/chat-templates.md |
--
language:
- en
license:
- bsd-3-clause
annotations_creators:
- crowdsourced
- expert-generated
language_creators:
- found
multilinguality:
- monolingual
size_categories:
- n<1K
task_categories:
- image-segmentation
task_ids:
- semantic-segmentation
pretty_name: Sample Segmentation
---
# Dataset Card for Sample Segmentation
This is a sample dataset card for a semantic segmentation dataset. | huggingface/huggingface_hub/blob/main/tests/fixtures/cards/sample_datasetcard_simple.md |
!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Nyströmformer
## Overview
The Nyströmformer model was proposed in [*Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention*](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn
Fung, Yin Li, and Vikas Singh.
The abstract from the paper is the following:
*Transformers have emerged as a powerful tool for a broad range of natural language processing tasks. A key component
that drives the impressive performance of Transformers is the self-attention mechanism that encodes the influence or
dependence of other tokens on each specific token. While beneficial, the quadratic complexity of self-attention on the
input sequence length has limited its application to longer sequences -- a topic being actively studied in the
community. To address this limitation, we propose Nyströmformer -- a model that exhibits favorable scalability as a
function of sequence length. Our idea is based on adapting the Nyström method to approximate standard self-attention
with O(n) complexity. The scalability of Nyströmformer enables application to longer sequences with thousands of
tokens. We perform evaluations on multiple downstream tasks on the GLUE benchmark and IMDB reviews with standard
sequence length, and find that our Nyströmformer performs comparably, or in a few cases, even slightly better, than
standard self-attention. On longer sequence tasks in the Long Range Arena (LRA) benchmark, Nyströmformer performs
favorably relative to other efficient self-attention methods. Our code is available at this https URL.*
This model was contributed by [novice03](https://huggingface.co/novice03). The original code can be found [here](https://github.com/mlpen/Nystromformer).
## Resources
- [Text classification task guide](../tasks/sequence_classification)
- [Token classification task guide](../tasks/token_classification)
- [Question answering task guide](../tasks/question_answering)
- [Masked language modeling task guide](../tasks/masked_language_modeling)
- [Multiple choice task guide](../tasks/multiple_choice)
## NystromformerConfig
[[autodoc]] NystromformerConfig
## NystromformerModel
[[autodoc]] NystromformerModel
- forward
## NystromformerForMaskedLM
[[autodoc]] NystromformerForMaskedLM
- forward
## NystromformerForSequenceClassification
[[autodoc]] NystromformerForSequenceClassification
- forward
## NystromformerForMultipleChoice
[[autodoc]] NystromformerForMultipleChoice
- forward
## NystromformerForTokenClassification
[[autodoc]] NystromformerForTokenClassification
- forward
## NystromformerForQuestionAnswering
[[autodoc]] NystromformerForQuestionAnswering
- forward
| huggingface/transformers/blob/main/docs/source/en/model_doc/nystromformer.md |
Gradio Demo: kitchen_sink
```
!pip install -q gradio
```
```
# Downloading files from the demo repo
import os
os.mkdir('files')
!wget -q -O files/cantina.wav https://github.com/gradio-app/gradio/raw/main/demo/kitchen_sink/files/cantina.wav
!wget -q -O files/cheetah1.jpg https://github.com/gradio-app/gradio/raw/main/demo/kitchen_sink/files/cheetah1.jpg
!wget -q -O files/lion.jpg https://github.com/gradio-app/gradio/raw/main/demo/kitchen_sink/files/lion.jpg
!wget -q -O files/logo.png https://github.com/gradio-app/gradio/raw/main/demo/kitchen_sink/files/logo.png
!wget -q -O files/time.csv https://github.com/gradio-app/gradio/raw/main/demo/kitchen_sink/files/time.csv
!wget -q -O files/titanic.csv https://github.com/gradio-app/gradio/raw/main/demo/kitchen_sink/files/titanic.csv
!wget -q -O files/tower.jpg https://github.com/gradio-app/gradio/raw/main/demo/kitchen_sink/files/tower.jpg
!wget -q -O files/world.mp4 https://github.com/gradio-app/gradio/raw/main/demo/kitchen_sink/files/world.mp4
```
```
import os
import json
import numpy as np
import gradio as gr
CHOICES = ["foo", "bar", "baz"]
JSONOBJ = """{"items":{"item":[{"id": "0001","type": null,"is_good": false,"ppu": 0.55,"batters":{"batter":[{ "id": "1001", "type": "Regular" },{ "id": "1002", "type": "Chocolate" },{ "id": "1003", "type": "Blueberry" },{ "id": "1004", "type": "Devil's Food" }]},"topping":[{ "id": "5001", "type": "None" },{ "id": "5002", "type": "Glazed" },{ "id": "5005", "type": "Sugar" },{ "id": "5007", "type": "Powdered Sugar" },{ "id": "5006", "type": "Chocolate with Sprinkles" },{ "id": "5003", "type": "Chocolate" },{ "id": "5004", "type": "Maple" }]}]}}"""
def fn(
text1,
text2,
num,
slider1,
slider2,
single_checkbox,
checkboxes,
radio,
dropdown,
multi_dropdown,
im1,
# im2,
# im3,
im4,
video,
audio1,
audio2,
file,
df1,
):
return (
(text1 if single_checkbox else text2)
+ ", selected:"
+ ", ".join(checkboxes), # Text
{
"positive": num / (num + slider1 + slider2),
"negative": slider1 / (num + slider1 + slider2),
"neutral": slider2 / (num + slider1 + slider2),
}, # Label
(audio1[0], np.flipud(audio1[1]))
if audio1 is not None
else os.path.join(os.path.abspath(''), "files/cantina.wav"), # Audio
np.flipud(im1)
if im1 is not None
else os.path.join(os.path.abspath(''), "files/cheetah1.jpg"), # Image
video
if video is not None
else os.path.join(os.path.abspath(''), "files/world.mp4"), # Video
[
("The", "art"),
("quick brown", "adj"),
("fox", "nn"),
("jumped", "vrb"),
("testing testing testing", None),
("over", "prp"),
("the", "art"),
("testing", None),
("lazy", "adj"),
("dogs", "nn"),
(".", "punc"),
]
+ [(f"test {x}", f"test {x}") for x in range(10)], # HighlightedText
# [("The testing testing testing", None), ("quick brown", 0.2), ("fox", 1), ("jumped", -1), ("testing testing testing", 0), ("over", 0), ("the", 0), ("testing", 0), ("lazy", 1), ("dogs", 0), (".", 1)] + [(f"test {x}", x/10) for x in range(-10, 10)], # HighlightedText
[
("The testing testing testing", None),
("over", 0.6),
("the", 0.2),
("testing", None),
("lazy", -0.1),
("dogs", 0.4),
(".", 0),
]
+ [(f"test", x / 10) for x in range(-10, 10)], # HighlightedText
json.loads(JSONOBJ), # JSON
"<button style='background-color: red'>Click Me: "
+ radio
+ "</button>", # HTML
os.path.join(os.path.abspath(''), "files/titanic.csv"),
df1, # Dataframe
np.random.randint(0, 10, (4, 4)), # Dataframe
)
demo = gr.Interface(
fn,
inputs=[
gr.Textbox(value="Lorem ipsum", label="Textbox"),
gr.Textbox(lines=3, placeholder="Type here..", label="Textbox 2"),
gr.Number(label="Number", value=42),
gr.Slider(10, 20, value=15, label="Slider: 10 - 20"),
gr.Slider(maximum=20, step=0.04, label="Slider: step @ 0.04"),
gr.Checkbox(label="Checkbox"),
gr.CheckboxGroup(label="CheckboxGroup", choices=CHOICES, value=CHOICES[0:2]),
gr.Radio(label="Radio", choices=CHOICES, value=CHOICES[2]),
gr.Dropdown(label="Dropdown", choices=CHOICES),
gr.Dropdown(
label="Multiselect Dropdown (Max choice: 2)",
choices=CHOICES,
multiselect=True,
max_choices=2,
),
gr.Image(label="Image"),
# gr.Image(label="Image w/ Cropper", tool="select"),
# gr.Image(label="Sketchpad", source="canvas"),
gr.Image(label="Webcam", sources=["webcam"]),
gr.Video(label="Video"),
gr.Audio(label="Audio"),
gr.Audio(label="Microphone", sources=["microphone"]),
gr.File(label="File"),
gr.Dataframe(label="Dataframe", headers=["Name", "Age", "Gender"]),
],
outputs=[
gr.Textbox(label="Textbox"),
gr.Label(label="Label"),
gr.Audio(label="Audio"),
gr.Image(label="Image"),
gr.Video(label="Video"),
gr.HighlightedText(
label="HighlightedText", color_map={"punc": "pink", "test 0": "blue"}
),
gr.HighlightedText(label="HighlightedText", show_legend=True),
gr.JSON(label="JSON"),
gr.HTML(label="HTML"),
gr.File(label="File"),
gr.Dataframe(label="Dataframe"),
gr.Dataframe(label="Numpy"),
],
examples=[
[
"the quick brown fox",
"jumps over the lazy dog",
10,
12,
4,
True,
["foo", "baz"],
"baz",
"bar",
["foo", "bar"],
os.path.join(os.path.abspath(''), "files/cheetah1.jpg"),
# os.path.join(os.path.abspath(''), "files/cheetah1.jpg"),
# os.path.join(os.path.abspath(''), "files/cheetah1.jpg"),
os.path.join(os.path.abspath(''), "files/cheetah1.jpg"),
os.path.join(os.path.abspath(''), "files/world.mp4"),
os.path.join(os.path.abspath(''), "files/cantina.wav"),
os.path.join(os.path.abspath(''), "files/cantina.wav"),
os.path.join(os.path.abspath(''), "files/titanic.csv"),
[[1, 2, 3, 4], [4, 5, 6, 7], [8, 9, 1, 2], [3, 4, 5, 6]],
]
]
* 3,
title="Kitchen Sink",
description="Try out all the components!",
article="Learn more about [Gradio](http://gradio.app)",
cache_examples=True,
)
if __name__ == "__main__":
demo.launch()
```
| gradio-app/gradio/blob/main/demo/kitchen_sink/run.ipynb |
`@gradio/utils`
General functions for handling events in Gradio Svelte components
```javascript
export async function uploadToHuggingFace(
data: string,
type: "base64" | "url"
): Promise<string>
export function copy(node: HTMLDivElement): ActionReturn
``` | gradio-app/gradio/blob/main/js/utils/README.md |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# SAM
## Overview
SAM (Segment Anything Model) was proposed in [Segment Anything](https://arxiv.org/pdf/2304.02643v1.pdf) by Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alex Berg, Wan-Yen Lo, Piotr Dollar, Ross Girshick.
The model can be used to predict segmentation masks of any object of interest given an input image.
The abstract from the paper is the following:
*We introduce the Segment Anything (SA) project: a new task, model, and dataset for image segmentation. Using our efficient model in a data collection loop, we built the largest segmentation dataset to date (by far), with over 1 billion masks on 11M licensed and privacy respecting images. The model is designed and trained to be promptable, so it can transfer zero-shot to new image distributions and tasks. We evaluate its capabilities on numerous tasks and find that its zero-shot performance is impressive -- often competitive with or even superior to prior fully supervised results. We are releasing the Segment Anything Model (SAM) and corresponding dataset (SA-1B) of 1B masks and 11M images at [https://segment-anything.com](https://segment-anything.com) to foster research into foundation models for computer vision.*
Tips:
- The model predicts binary masks that states the presence or not of the object of interest given an image.
- The model predicts much better results if input 2D points and/or input bounding boxes are provided
- You can prompt multiple points for the same image, and predict a single mask.
- Fine-tuning the model is not supported yet
- According to the paper, textual input should be also supported. However, at this time of writing this seems to be not supported according to [the official repository](https://github.com/facebookresearch/segment-anything/issues/4#issuecomment-1497626844).
This model was contributed by [ybelkada](https://huggingface.co/ybelkada) and [ArthurZ](https://huggingface.co/ArthurZ).
The original code can be found [here](https://github.com/facebookresearch/segment-anything).
Below is an example on how to run mask generation given an image and a 2D point:
```python
import torch
from PIL import Image
import requests
from transformers import SamModel, SamProcessor
device = "cuda" if torch.cuda.is_available() else "cpu"
model = SamModel.from_pretrained("facebook/sam-vit-huge").to(device)
processor = SamProcessor.from_pretrained("facebook/sam-vit-huge")
img_url = "https://huggingface.co/ybelkada/segment-anything/resolve/main/assets/car.png"
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB")
input_points = [[[450, 600]]] # 2D location of a window in the image
inputs = processor(raw_image, input_points=input_points, return_tensors="pt").to(device)
with torch.no_grad():
outputs = model(**inputs)
masks = processor.image_processor.post_process_masks(
outputs.pred_masks.cpu(), inputs["original_sizes"].cpu(), inputs["reshaped_input_sizes"].cpu()
)
scores = outputs.iou_scores
```
Resources:
- [Demo notebook](https://github.com/huggingface/notebooks/blob/main/examples/segment_anything.ipynb) for using the model.
- [Demo notebook](https://github.com/huggingface/notebooks/blob/main/examples/automatic_mask_generation.ipynb) for using the automatic mask generation pipeline.
- [Demo notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/SAM/Run_inference_with_MedSAM_using_HuggingFace_Transformers.ipynb) for inference with MedSAM, a fine-tuned version of SAM on the medical domain.
- [Demo notebook](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/SAM/Fine_tune_SAM_(segment_anything)_on_a_custom_dataset.ipynb) for fine-tuning the model on custom data.
## SamConfig
[[autodoc]] SamConfig
## SamVisionConfig
[[autodoc]] SamVisionConfig
## SamMaskDecoderConfig
[[autodoc]] SamMaskDecoderConfig
## SamPromptEncoderConfig
[[autodoc]] SamPromptEncoderConfig
## SamProcessor
[[autodoc]] SamProcessor
## SamImageProcessor
[[autodoc]] SamImageProcessor
## SamModel
[[autodoc]] SamModel
- forward
## TFSamModel
[[autodoc]] TFSamModel
- call
| huggingface/transformers/blob/main/docs/source/en/model_doc/sam.md |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# Contribute a community pipeline
<Tip>
💡 Take a look at GitHub Issue [#841](https://github.com/huggingface/diffusers/issues/841) for more context about why we're adding community pipelines to help everyone easily share their work without being slowed down.
</Tip>
Community pipelines allow you to add any additional features you'd like on top of the [`DiffusionPipeline`]. The main benefit of building on top of the `DiffusionPipeline` is anyone can load and use your pipeline by only adding one more argument, making it super easy for the community to access.
This guide will show you how to create a community pipeline and explain how they work. To keep things simple, you'll create a "one-step" pipeline where the `UNet` does a single forward pass and calls the scheduler once.
## Initialize the pipeline
You should start by creating a `one_step_unet.py` file for your community pipeline. In this file, create a pipeline class that inherits from the [`DiffusionPipeline`] to be able to load model weights and the scheduler configuration from the Hub. The one-step pipeline needs a `UNet` and a scheduler, so you'll need to add these as arguments to the `__init__` function:
```python
from diffusers import DiffusionPipeline
import torch
class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
def __init__(self, unet, scheduler):
super().__init__()
```
To ensure your pipeline and its components (`unet` and `scheduler`) can be saved with [`~DiffusionPipeline.save_pretrained`], add them to the `register_modules` function:
```diff
from diffusers import DiffusionPipeline
import torch
class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
def __init__(self, unet, scheduler):
super().__init__()
+ self.register_modules(unet=unet, scheduler=scheduler)
```
Cool, the `__init__` step is done and you can move to the forward pass now! 🔥
## Define the forward pass
In the forward pass, which we recommend defining as `__call__`, you have complete creative freedom to add whatever feature you'd like. For our amazing one-step pipeline, create a random image and only call the `unet` and `scheduler` once by setting `timestep=1`:
```diff
from diffusers import DiffusionPipeline
import torch
class UnetSchedulerOneForwardPipeline(DiffusionPipeline):
def __init__(self, unet, scheduler):
super().__init__()
self.register_modules(unet=unet, scheduler=scheduler)
+ def __call__(self):
+ image = torch.randn(
+ (1, self.unet.config.in_channels, self.unet.config.sample_size, self.unet.config.sample_size),
+ )
+ timestep = 1
+ model_output = self.unet(image, timestep).sample
+ scheduler_output = self.scheduler.step(model_output, timestep, image).prev_sample
+ return scheduler_output
```
That's it! 🚀 You can now run this pipeline by passing a `unet` and `scheduler` to it:
```python
from diffusers import DDPMScheduler, UNet2DModel
scheduler = DDPMScheduler()
unet = UNet2DModel()
pipeline = UnetSchedulerOneForwardPipeline(unet=unet, scheduler=scheduler)
output = pipeline()
```
But what's even better is you can load pre-existing weights into the pipeline if the pipeline structure is identical. For example, you can load the [`google/ddpm-cifar10-32`](https://huggingface.co/google/ddpm-cifar10-32) weights into the one-step pipeline:
```python
pipeline = UnetSchedulerOneForwardPipeline.from_pretrained("google/ddpm-cifar10-32", use_safetensors=True)
output = pipeline()
```
## Share your pipeline
Open a Pull Request on the 🧨 Diffusers [repository](https://github.com/huggingface/diffusers) to add your awesome pipeline in `one_step_unet.py` to the [examples/community](https://github.com/huggingface/diffusers/tree/main/examples/community) subfolder.
Once it is merged, anyone with `diffusers >= 0.4.0` installed can use this pipeline magically 🪄 by specifying it in the `custom_pipeline` argument:
```python
from diffusers import DiffusionPipeline
pipe = DiffusionPipeline.from_pretrained(
"google/ddpm-cifar10-32", custom_pipeline="one_step_unet", use_safetensors=True
)
pipe()
```
Another way to share your community pipeline is to upload the `one_step_unet.py` file directly to your preferred [model repository](https://huggingface.co/docs/hub/models-uploading) on the Hub. Instead of specifying the `one_step_unet.py` file, pass the model repository id to the `custom_pipeline` argument:
```python
from diffusers import DiffusionPipeline
pipeline = DiffusionPipeline.from_pretrained(
"google/ddpm-cifar10-32", custom_pipeline="stevhliu/one_step_unet", use_safetensors=True
)
```
Take a look at the following table to compare the two sharing workflows to help you decide the best option for you:
| | GitHub community pipeline | HF Hub community pipeline |
|----------------|------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------|
| usage | same | same |
| review process | open a Pull Request on GitHub and undergo a review process from the Diffusers team before merging; may be slower | upload directly to a Hub repository without any review; this is the fastest workflow |
| visibility | included in the official Diffusers repository and documentation | included on your HF Hub profile and relies on your own usage/promotion to gain visibility |
<Tip>
💡 You can use whatever package you want in your community pipeline file - as long as the user has it installed, everything will work fine. Make sure you have one and only one pipeline class that inherits from `DiffusionPipeline` because this is automatically detected.
</Tip>
## How do community pipelines work?
A community pipeline is a class that inherits from [`DiffusionPipeline`] which means:
- It can be loaded with the [`custom_pipeline`] argument.
- The model weights and scheduler configuration are loaded from [`pretrained_model_name_or_path`].
- The code that implements a feature in the community pipeline is defined in a `pipeline.py` file.
Sometimes you can't load all the pipeline components weights from an official repository. In this case, the other components should be passed directly to the pipeline:
```python
from diffusers import DiffusionPipeline
from transformers import CLIPImageProcessor, CLIPModel
model_id = "CompVis/stable-diffusion-v1-4"
clip_model_id = "laion/CLIP-ViT-B-32-laion2B-s34B-b79K"
feature_extractor = CLIPImageProcessor.from_pretrained(clip_model_id)
clip_model = CLIPModel.from_pretrained(clip_model_id, torch_dtype=torch.float16)
pipeline = DiffusionPipeline.from_pretrained(
model_id,
custom_pipeline="clip_guided_stable_diffusion",
clip_model=clip_model,
feature_extractor=feature_extractor,
scheduler=scheduler,
torch_dtype=torch.float16,
use_safetensors=True,
)
```
The magic behind community pipelines is contained in the following code. It allows the community pipeline to be loaded from GitHub or the Hub, and it'll be available to all 🧨 Diffusers packages.
```python
# 2. Load the pipeline class, if using custom module then load it from the Hub
# if we load from explicit class, let's use it
if custom_pipeline is not None:
pipeline_class = get_class_from_dynamic_module(
custom_pipeline, module_file=CUSTOM_PIPELINE_FILE_NAME, cache_dir=custom_pipeline
)
elif cls != DiffusionPipeline:
pipeline_class = cls
else:
diffusers_module = importlib.import_module(cls.__module__.split(".")[0])
pipeline_class = getattr(diffusers_module, config_dict["_class_name"])
```
| huggingface/diffusers/blob/main/docs/source/en/using-diffusers/contribute_pipeline.md |
--
title: "Make your llama generation time fly with AWS Inferentia2"
thumbnail: /blog/assets/inferentia-llama2/thumbnail.png
authors:
- user: dacorvo
---
# Make your llama generation time fly with AWS Inferentia2
In a [previous post on the Hugging Face blog](https://huggingface.co/blog/accelerate-transformers-with-inferentia2), we introduced [AWS Inferentia2](https://aws.amazon.com/ec2/instance-types/inf2/), the second-generation AWS Inferentia accelerator, and explained how you could use [optimum-neuron](https://huggingface.co/docs/optimum-neuron/index) to quickly deploy Hugging Face models for standard text and vision tasks on AWS Inferencia 2 instances.
In a further step of integration with the [AWS Neuron SDK](https://github.com/aws-neuron/aws-neuron-sdk), it is now possible to use 🤗 [optimum-neuron](https://huggingface.co/docs/optimum-neuron/index) to deploy LLM models for text generation on AWS Inferentia2.
And what better model could we choose for that demonstration than [Llama 2](https://huggingface.co/meta-llama/Llama-2-13b-hf), one of the most popular models on the [Hugging Face hub](https://huggingface.co/models).
## Setup 🤗 optimum-neuron on your Inferentia2 instance
Our recommendation is to use the [Hugging Face Neuron Deep Learning AMI](https://aws.amazon.com/marketplace/pp/prodview-gr3e6yiscria2) (DLAMI). The DLAMI comes with all required libraries pre-packaged for you, including the Optimum Neuron, Neuron Drivers, Transformers, Datasets, and Accelerate.
Alternatively, you can use the [Hugging Face Neuron SDK DLC](https://github.com/aws/deep-learning-containers/releases?q=hf&expanded=true) to deploy on Amazon SageMaker.
*Note: stay tuned for an upcoming post dedicated to SageMaker deployment.*
Finally, these components can also be installed manually on a fresh Inferentia2 instance following the `optimum-neuron` [installation instructions](https://huggingface.co/docs/optimum-neuron/installation).
## Export the Llama 2 model to Neuron
As explained in the [optimum-neuron documentation](https://huggingface.co/docs/optimum-neuron/guides/export_model#why-compile-to-neuron-model), models need to be compiled and exported to a serialized format before running them on Neuron devices.
Fortunately, 🤗 `optimum-neuron` offers a [very simple API](https://huggingface.co/docs/optimum-neuron/guides/models#configuring-the-export-of-a-generative-model) to export standard 🤗 [transformers models](https://huggingface.co/docs/transformers/index) to the Neuron format.
```
>>> from optimum.neuron import NeuronModelForCausalLM
>>> compiler_args = {"num_cores": 24, "auto_cast_type": 'fp16'}
>>> input_shapes = {"batch_size": 1, "sequence_length": 2048}
>>> model = NeuronModelForCausalLM.from_pretrained(
"meta-llama/Llama-2-7b-hf",
export=True,
**compiler_args,
**input_shapes)
```
This deserves a little explanation:
- using `compiler_args`, we specify on how many cores we want the model to be deployed (each neuron device has two cores), and with which precision (here `float16`),
- using `input_shape`, we set the static input and output dimensions of the model. All model compilers require static shapes, and neuron makes no exception. Note that the
`sequence_length` not only constrains the length of the input context, but also the length of the KV cache, and thus, the output length.
Depending on your choice of parameters and inferentia host, this may take from a few minutes to more than an hour.
Fortunately, you will need to do this only once because you can save your model and reload it later.
```
>>> model.save_pretrained("a_local_path_for_compiled_neuron_model")
```
Even better, you can push it to the [Hugging Face hub](https://huggingface.co/models).
```
>>> model.push_to_hub(
"a_local_path_for_compiled_neuron_model",
repository_id="aws-neuron/Llama-2-7b-hf-neuron-latency")
```
## Generate Text using Llama 2 on AWS Inferentia2
Once your model has been exported, you can generate text using the transformers library, as it has been described in [detail in this previous post](https://huggingface.co/blog/how-to-generate).
```
>>> from optimum.neuron import NeuronModelForCausalLM
>>> from transformers import AutoTokenizer
>>> model = NeuronModelForCausalLM.from_pretrained('aws-neuron/Llama-2-7b-hf-neuron-latency')
>>> tokenizer = AutoTokenizer.from_pretrained("aws-neuron/Llama-2-7b-hf-neuron-latency")
>>> inputs = tokenizer("What is deep-learning ?", return_tensors="pt")
>>> outputs = model.generate(**inputs,
max_new_tokens=128,
do_sample=True,
temperature=0.9,
top_k=50,
top_p=0.9)
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
['What is deep-learning ?\nThe term “deep-learning” refers to a type of machine-learning
that aims to model high-level abstractions of the data in the form of a hierarchy of multiple
layers of increasingly complex processing nodes.']
```
*Note: when passing multiple input prompts to a model, the resulting token sequences must be padded to the left with an end-of-stream token.
The tokenizers saved with the exported models are configured accordingly.*
The following generation strategies are supported:
- greedy search,
- multinomial sampling with top-k and top-p (with temperature).
Most logits pre-processing/filters (such as repetition penalty) are supported.
## All-in-one with optimum-neuron pipelines
For those who like to keep it simple, there is an even simpler way to use an LLM model on AWS inferentia 2 using [optimum-neuron pipelines](https://huggingface.co/docs/optimum-neuron/guides/pipelines).
Using them is as simple as:
```
>>> from optimum.neuron import pipeline
>>> p = pipeline('text-generation', 'aws-neuron/Llama-2-7b-hf-neuron-budget')
>>> p("My favorite place on earth is", max_new_tokens=64, do_sample=True, top_k=50)
[{'generated_text': 'My favorite place on earth is the ocean. It is where I feel most
at peace. I love to travel and see new places. I have a'}]
```
## Benchmarks
But how much efficient is text-generation on Inferentia2? Let's figure out!
We have uploaded on the hub pre-compiled versions of the LLama 2 7B and 13B models with different configurations:
| Model type | num cores | batch_size | Hugging Face Hub model |
|----------------------------|-----------|------------|-------------------------------------------|
| Llama2 7B - B (budget) | 2 | 1 |[aws-neuron/Llama-2-7b-hf-neuron-budget](https://huggingface.co/aws-neuron/Llama-2-7b-hf-neuron-budget) |
| Llama2 7B - L (latency) | 24 | 1 |[aws-neuron/Llama-2-7b-hf-neuron-latency](https://huggingface.co/aws-neuron/Llama-2-7b-hf-neuron-latency) |
| Llama2 7B - T (throughput) | 24 | 4 |[aws-neuron/Llama-2-7b-hf-neuron-throughput](https://huggingface.co/aws-neuron/Llama-2-7b-hf-neuron-throughput) |
| Llama2 13B - L (latency) | 24 | 1 |[aws-neuron/Llama-2-13b-hf-neuron-latency](https://huggingface.co/aws-neuron/Llama-2-13b-hf-neuron-latency) |
| Llama2 13B - T (throughput)| 24 | 4 |[aws-neuron/Llama-2-13b-hf-neuron-throughput](https://huggingface.co/aws-neuron/Llama-2-13b-hf-neuron-throughput)|
*Note: all models are compiled with a maximum sequence length of 2048.*
The `llama2 7B` "budget" model is meant to be deployed on `inf2.xlarge` instance that has only one neuron device, and enough `cpu` memory to load the model.
All other models are compiled to use the full extent of cores available on the `inf2.48xlarge` instance.
*Note: please refer to the [inferentia2 product page](https://aws.amazon.com/ec2/instance-types/inf2/) for details on the available instances.*
We created two "latency" oriented configurations for the `llama2 7B` and `llama2 13B` models that can serve only one request at a time, but at full speed.
We also created two "throughput" oriented configurations to serve up to four requests in parallel.
To evaluate the models, we generate tokens up to a total sequence length of 1024, starting from
256 input tokens (i.e. we generate 256, 512 and 768 tokens).
*Note: the "budget" model numbers are reported but not included in the graphs for better readability.*
### Encoding time
The encoding time is the time required to process the input tokens and generate the first output token.
It is a very important metric, as it corresponds to the latency directly perceived by the user when streaming generated tokens.
We test the encoding time for increasing context sizes, 256 input tokens corresponding roughly to a typical Q/A usage,
while 768 is more typical of a Retrieval Augmented Generation (RAG) use-case.
The "budget" model (`Llama2 7B-B`) is deployed on an `inf2.xlarge` instance while other models are deployed on an `inf2.48xlarge` instance.
Encoding time is expressed in **seconds**.
| input tokens | Llama2 7B-L | Llama2 7B-T | Llama2 13B-L | Llama2 13B-T | Llama2 7B-B |
|-----------------|----------------|----------------|-----------------|-----------------|----------------|
| 256 | 0.5 | 0.9 | 0.6 | 1.8 | 0.3 |
| 512 | 0.7 | 1.6 | 1.1 | 3.0 | 0.4 |
| 768 | 1.1 | 3.3 | 1.7 | 5.2 | 0.5 |
We can see that all deployed models exhibit excellent response times, even for long contexts.
### End-to-end Latency
The end-to-end latency corresponds to the total time to reach a sequence length of 1024 tokens.
It therefore includes the encoding and generation time.
The "budget" model (`Llama2 7B-B`) is deployed on an `inf2.xlarge` instance while other models are deployed on an `inf2.48xlarge` instance.
Latency is expressed in **seconds**.
| new tokens | Llama2 7B-L | Llama2 7B-T | Llama2 13B-L | Llama2 13B-T | Llama2 7B-B |
|---------------|----------------|----------------|-----------------|-----------------|----------------|
| 256 | 2.3 | 2.7 | 3.5 | 4.1 | 15.9 |
| 512 | 4.4 | 5.3 | 6.9 | 7.8 | 31.7 |
| 768 | 6.2 | 7.7 | 10.2 | 11.1 | 47.3 |
All models deployed on the high-end instance exhibit a good latency, even those actually configured to optimize throughput.
The "budget" deployed model latency is significantly higher, but still ok.
### Throughput
We adopt the same convention as other benchmarks to evaluate the throughput, by dividing the end-to-end
latency by the sum of both input and output tokens.
In other words, we divide the end-to-end latency by `batch_size * sequence_length` to obtain the number of generated tokens per second.
The "budget" model (`Llama2 7B-B`) is deployed on an `inf2.xlarge` instance while other models are deployed on an `inf2.48xlarge` instance.
Throughput is expressed in **tokens/second**.
| new tokens | Llama2 7B-L | Llama2 7B-T | Llama2 13B-L | Llama2 13B-T | Llama2 7B-B |
|---------------|----------------|----------------|-----------------|-----------------|----------------|
| 256 | 227 | 750 | 145 | 504 | 32 |
| 512 | 177 | 579 | 111 | 394 | 24 |
| 768 | 164 | 529 | 101 | 370 | 22 |
Again, the models deployed on the high-end instance have a very good throughput, even those optimized for latency.
The "budget" model has a much lower throughput, but still ok for a streaming use-case, considering that an average reader reads around 5 words per-second.
## Conclusion
We have illustrated how easy it is to deploy `llama2` models from the [Hugging Face hub](https://huggingface.co/models) on
[AWS Inferentia2](https://aws.amazon.com/ec2/instance-types/inf2/) using 🤗 [optimum-neuron](https://huggingface.co/docs/optimum-neuron/index).
The deployed models demonstrate very good performance in terms of encoding time, latency and throughput.
Interestingly, the deployed models latency is not too sensitive to the batch size, which opens the way for their deployment on inference endpoints
serving multiple requests in parallel.
There is still plenty of room for improvement though:
- in the current implementation, the only way to augment the throughput is to increase the batch size, but it is currently limited by the device memory.
Alternative options such as pipelining are currently integrated,
- the static sequence length limits the model ability to encode long contexts. It would be interesting to see if attention sinks might be a valid option to address this.
| huggingface/blog/blob/main/inferentia-llama2.md |
!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
-->
# DDIM
[Denoising Diffusion Implicit Models](https://huggingface.co/papers/2010.02502) (DDIM) by Jiaming Song, Chenlin Meng and Stefano Ermon.
The abstract from the paper is:
*Denoising diffusion probabilistic models (DDPMs) have achieved high quality image generation without adversarial training, yet they require simulating a Markov chain for many steps to produce a sample. To accelerate sampling, we present denoising diffusion implicit models (DDIMs), a more efficient class of iterative implicit probabilistic models with the same training procedure as DDPMs. In DDPMs, the generative process is defined as the reverse of a Markovian diffusion process. We construct a class of non-Markovian diffusion processes that lead to the same training objective, but whose reverse process can be much faster to sample from. We empirically demonstrate that DDIMs can produce high quality samples 10× to 50× faster in terms of wall-clock time compared to DDPMs, allow us to trade off computation for sample quality, and can perform semantically meaningful image interpolation directly in the latent space.*
The original codebase can be found at [ermongroup/ddim](https://github.com/ermongroup/ddim).
## DDIMPipeline
[[autodoc]] DDIMPipeline
- all
- __call__
## ImagePipelineOutput
[[autodoc]] pipelines.ImagePipelineOutput
| huggingface/diffusers/blob/main/docs/source/en/api/pipelines/ddim.md |