Hugging Face's logo

qualcomm
/
FFNet-78S

Image Segmentation
PyTorch
LiteRT
ONNX
android
Model card Files Community

FFNet-78S: Optimized for Mobile Deployment

Semantic segmentation for automotive street scenes

FFNet-78S is a "fuss-free network" that segments street scene images with per-pixel classes like road, sidewalk, and pedestrian. Trained on the Cityscapes dataset.

This model is an implementation of FFNet-78S found here.

This repository provides scripts to run FFNet-78S on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

  • Model Type: Model_use_case.semantic_segmentation
  • Model Stats:
    • Model checkpoint: ffnet78S_dBBB_cityscapes_state_dict_quarts
    • Input resolution: 2048x1024
    • Number of output classes: 19
    • Number of parameters: 27.5M
    • Model size (float): 105 MB
    • Model size (w8a8): 26.7 MB
Model Precision Device Chipset Target Runtime Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit Target Model
FFNet-78S float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 200.224 ms 3 - 79 MB NPU FFNet-78S.tflite
FFNet-78S float QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 183.562 ms 24 - 84 MB NPU FFNet-78S.dlc
FFNet-78S float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 66.658 ms 2 - 128 MB NPU FFNet-78S.tflite
FFNet-78S float QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 76.883 ms 24 - 78 MB NPU FFNet-78S.dlc
FFNet-78S float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 49.402 ms 2 - 20 MB NPU FFNet-78S.tflite
FFNet-78S float QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 40.272 ms 24 - 46 MB NPU FFNet-78S.dlc
FFNet-78S float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 69.12 ms 2 - 78 MB NPU FFNet-78S.tflite
FFNet-78S float QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 59.063 ms 24 - 84 MB NPU FFNet-78S.dlc
FFNet-78S float SA7255P ADP Qualcomm® SA7255P TFLITE 200.224 ms 3 - 79 MB NPU FFNet-78S.tflite
FFNet-78S float SA7255P ADP Qualcomm® SA7255P QNN_DLC 183.562 ms 24 - 84 MB NPU FFNet-78S.dlc
FFNet-78S float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 49.84 ms 2 - 21 MB NPU FFNet-78S.tflite
FFNet-78S float SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 40.097 ms 24 - 51 MB NPU FFNet-78S.dlc
FFNet-78S float SA8295P ADP Qualcomm® SA8295P TFLITE 75.319 ms 2 - 74 MB NPU FFNet-78S.tflite
FFNet-78S float SA8295P ADP Qualcomm® SA8295P QNN_DLC 64.251 ms 17 - 76 MB NPU FFNet-78S.dlc
FFNet-78S float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 49.826 ms 2 - 26 MB NPU FFNet-78S.tflite
FFNet-78S float SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 40.647 ms 24 - 49 MB NPU FFNet-78S.dlc
FFNet-78S float SA8775P ADP Qualcomm® SA8775P TFLITE 69.12 ms 2 - 78 MB NPU FFNet-78S.tflite
FFNet-78S float SA8775P ADP Qualcomm® SA8775P QNN_DLC 59.063 ms 24 - 84 MB NPU FFNet-78S.dlc
FFNet-78S float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile TFLITE 49.685 ms 2 - 23 MB NPU FFNet-78S.tflite
FFNet-78S float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile QNN_DLC 39.978 ms 24 - 49 MB NPU FFNet-78S.dlc
FFNet-78S float Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile ONNX 35.342 ms 30 - 166 MB NPU FFNet-78S.onnx
FFNet-78S float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 33.847 ms 2 - 128 MB NPU FFNet-78S.tflite
FFNet-78S float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 28.257 ms 24 - 84 MB NPU FFNet-78S.dlc
FFNet-78S float Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 25.199 ms 26 - 112 MB NPU FFNet-78S.onnx
FFNet-78S float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile TFLITE 34.221 ms 2 - 79 MB NPU FFNet-78S.tflite
FFNet-78S float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile QNN_DLC 22.099 ms 28 - 99 MB NPU FFNet-78S.dlc
FFNet-78S float Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile ONNX 23.1 ms 30 - 85 MB NPU FFNet-78S.onnx
FFNet-78S float Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 42.923 ms 24 - 24 MB NPU FFNet-78S.dlc
FFNet-78S float Snapdragon X Elite CRD Snapdragon® X Elite ONNX 37.031 ms 31 - 31 MB NPU FFNet-78S.onnx
FFNet-78S w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) TFLITE 29.503 ms 0 - 49 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 QCS8275 (Proxy) Qualcomm® QCS8275 (Proxy) QNN_DLC 39.591 ms 6 - 71 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) TFLITE 12.28 ms 1 - 77 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 QCS8450 (Proxy) Qualcomm® QCS8450 (Proxy) QNN_DLC 20.922 ms 6 - 93 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) TFLITE 10.612 ms 1 - 13 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 QCS8550 (Proxy) Qualcomm® QCS8550 (Proxy) QNN_DLC 16.593 ms 6 - 31 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) TFLITE 11.055 ms 0 - 50 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 QCS9075 (Proxy) Qualcomm® QCS9075 (Proxy) QNN_DLC 17.09 ms 6 - 72 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 RB3 Gen 2 (Proxy) Qualcomm® QCS6490 (Proxy) TFLITE 68.688 ms 0 - 110 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 RB3 Gen 2 (Proxy) Qualcomm® QCS6490 (Proxy) QNN_DLC 70.386 ms 6 - 213 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 RB5 (Proxy) Qualcomm® QCS8250 (Proxy) TFLITE 336.474 ms 1 - 3 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 SA7255P ADP Qualcomm® SA7255P TFLITE 29.503 ms 0 - 49 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 SA7255P ADP Qualcomm® SA7255P QNN_DLC 39.591 ms 6 - 71 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) TFLITE 10.625 ms 1 - 17 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 SA8255 (Proxy) Qualcomm® SA8255P (Proxy) QNN_DLC 16.584 ms 6 - 29 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 SA8295P ADP Qualcomm® SA8295P TFLITE 16.848 ms 1 - 51 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 SA8295P ADP Qualcomm® SA8295P QNN_DLC 23.454 ms 6 - 73 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) TFLITE 10.637 ms 1 - 13 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 SA8650 (Proxy) Qualcomm® SA8650P (Proxy) QNN_DLC 16.645 ms 6 - 28 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 SA8775P ADP Qualcomm® SA8775P TFLITE 11.055 ms 0 - 50 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 SA8775P ADP Qualcomm® SA8775P QNN_DLC 17.09 ms 6 - 72 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile TFLITE 10.588 ms 1 - 12 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile QNN_DLC 16.603 ms 8 - 32 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 Samsung Galaxy S23 Snapdragon® 8 Gen 2 Mobile ONNX 18.641 ms 1 - 73 MB NPU FFNet-78S.onnx
FFNet-78S w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile TFLITE 7.569 ms 1 - 79 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile QNN_DLC 11.805 ms 6 - 96 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 Samsung Galaxy S24 Snapdragon® 8 Gen 3 Mobile ONNX 13.462 ms 6 - 214 MB NPU FFNet-78S.onnx
FFNet-78S w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile TFLITE 7.399 ms 0 - 54 MB NPU FFNet-78S.tflite
FFNet-78S w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile QNN_DLC 9.659 ms 6 - 82 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 Snapdragon 8 Elite QRD Snapdragon® 8 Elite Mobile ONNX 13.562 ms 14 - 482 MB NPU FFNet-78S.onnx
FFNet-78S w8a8 Snapdragon X Elite CRD Snapdragon® X Elite QNN_DLC 18.576 ms 108 - 108 MB NPU FFNet-78S.dlc
FFNet-78S w8a8 Snapdragon X Elite CRD Snapdragon® X Elite ONNX 22.384 ms 22 - 22 MB NPU FFNet-78S.onnx

Installation

Install the package via pip:

pip install "qai-hub-models[ffnet-78s]"

Configure Qualcomm® AI Hub to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.ffnet_78s.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.ffnet_78s.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

  • Performance check on-device on a cloud-hosted device
  • Downloads compiled assets that can be deployed on-device for Android.
  • Accuracy check between PyTorch and on-device outputs.
python -m qai_hub_models.models.ffnet_78s.export

Profiling Results
------------------------------------------------------------
FFNet-78S
Device                          : cs_8275 (ANDROID 14)                 
Runtime                         : TFLITE                               
Estimated inference time (ms)   : 200.2                                
Estimated peak memory usage (MB): [3, 79]                              
Total # Ops                     : 151                                  
Compute Unit(s)                 : npu (151 ops) gpu (0 ops) cpu (0 ops)

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.ffnet_78s import Model

# Load the model
torch_model = Model.from_pretrained()

# Device
device = hub.Device("Samsung Galaxy S24")

# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()

pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])

# Compile model on a specific device
compile_job = hub.submit_compile_job(
    model=pt_model,
    device=device,
    input_specs=torch_model.get_input_spec(),
)

# Get target model to run on-device
target_model = compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After compiling models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.

profile_job = hub.submit_profile_job(
    model=target_model,
    device=device,
)

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.

input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
    model=target_model,
    device=device,
    inputs=input_data,
)
    on_device_output = inference_job.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access.

Run demo on a cloud-hosted device

You can also run the demo on-device.

python -m qai_hub_models.models.ffnet_78s.demo --eval-mode on-device

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.ffnet_78s.demo -- --eval-mode on-device

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

  • TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.

  • QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on FFNet-78S's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

  • The license for the original implementation of FFNet-78S can be found here.
  • The license for the compiled assets for on-device deployment can be found here

References

Community

Downloads last month
58
Inference Providers NEW
Image Segmentation
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support