update

.gitignore

@@ -0,0 +1,164 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+data

.pre-commit-config.yaml

@@ -0,0 +1,14 @@
+repos:
+  - repo: https://github.com/pycqa/isort
+    rev: 5.12.0
+    hooks:
+      - id: isort
+        args: ["--profile", "black", "--filter-files"]
+  - repo: https://github.com/psf/black
+    rev: 22.6.0
+    hooks:
+      - id: black-jupyter
+  - repo: https://github.com/kynan/nbstripout
+    rev: 0.5.0
+    hooks:
+      - id: nbstripout

README.md

@@ -12,4 +12,5 @@ Supported datasets:
 
 Supported models:
 
-* [open_clip](https://github.com/mlfoundations/open_clip): [[ViT-B/32]](https://huggingface.co/laion/CLIP-ViT-B-32-laion2B-s34B-b79K)
+* [CLIP](https://github.com/openai/CLIP): ViT-B/32, ViT-L/14
+* [OpenCLIP](https://github.com/mlfoundations/open_clip): [ViT-B-32](https://huggingface.co/laion/CLIP-ViT-B-32-laion2B-s34B-b79K), [ViT-L-14](https://huggingface.co/laion/CLIP-ViT-L-14-laion2B-s32B-b82K)

dataset_lib/datasets.py

@@ -0,0 +1,62 @@
+import os
+
+from PIL import Image
+from torchvision.datasets import VisionDataset
+from torchvision.datasets.folder import find_classes, make_dataset
+
+SUPPORTED_DATASETS = []
+workdir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+with open(os.path.join(workdir, "supported_datasets.txt"), "r") as f:
+    for line in f:
+        SUPPORTED_DATASETS.append(line.strip())
+
+data_dir = os.path.join(workdir, "data")
+
+
+def get_dataset(dataset, transform=None):
+    if dataset == "imagenette":
+        root_dir = os.path.join(data_dir, "imagenette2")
+        train_dataset = Imagenette(root_dir, train=True, transform=transform)
+        val_dataset = Imagenette(root_dir, train=False, transform=transform)
+        return {"train": train_dataset, "val": val_dataset}
+
+
+class Imagenette(VisionDataset):
+    _WNID_TO_CLASS = {
+        "n01440764": ("tench", "Tinca tinca"),
+        "n02102040": ("English springer", "English springer spaniel"),
+        "n02979186": ("cassette player",),
+        "n03000684": ("chainsaw", "chain saw"),
+        "n03028079": ("church", "church building"),
+        "n03394916": ("French horn", "horn"),
+        "n03417042": ("garbage truck", "dustcart"),
+        "n03425413": ("gas pump", "gasoline pump", "petrol pump", "island dispenser"),
+        "n03445777": ("golf ball",),
+        "n03888257": ("parachute", "chute"),
+    }
+
+    def __init__(self, root, train=True, transform=None):
+        super().__init__(root, transform=transform)
+        self.train = train
+
+        self._split = "train" if train else "val"
+        self._image_root = os.path.join(root, self._split)
+
+        self.wnids, self.wnid_to_idx = find_classes(self._image_root)
+        self.classes = [self._WNID_TO_CLASS[wnid][0] for wnid in self.wnids]
+
+        self._samples = make_dataset(
+            self._image_root, self.wnid_to_idx, extensions=".jpeg"
+        )
+
+    def __len__(self):
+        return len(self._samples)
+
+    def __getitem__(self, idx):
+        path, label = self._samples[idx]
+        image = Image.open(path).convert("RGB")
+
+        if self.transform is not None:
+            image = self.transform(image)
+
+        return image, label

dataset_lib/encode.py

@@ -0,0 +1,89 @@
+import argparse
+import os
+import pickle
+
+import torch
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+
+from dataset_lib.datasets import SUPPORTED_DATASETS, get_dataset
+from dataset_lib.models import SUPPORTED_MODELS, get_transform_and_encoding_fn
+
+curr_dir = os.path.dirname(os.path.abspath(__file__))
+workdir = os.path.dirname(curr_dir)
+
+
+def model_type(model):
+    assert model in SUPPORTED_MODELS, NotImplementedError(
+        f"Model {model} not supported"
+    )
+    return model
+
+
+def dataset_type(dataset):
+    assert dataset in SUPPORTED_DATASETS, NotImplementedError(
+        f"Dataset {dataset} not supported"
+    )
+    return dataset
+
+
+def config():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model",
+        type=model_type,
+        help="Model to use for encoding dataset",
+        default="open_clip:ViT-B-32",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=dataset_type,
+        help="Dataset to encode",
+        default="imagenette",
+    )
+
+    return parser.parse_args()
+
+
+def encode(
+    model_name,
+    dataset_name,
+    device=torch.device("cuda" if torch.cuda.is_available() else "cpu"),
+):
+    preprocess, encode_fn = get_transform_and_encoding_fn(model_name, device=device)
+    datasets = get_dataset(dataset_name, transform=preprocess)
+
+    prefix = os.path.join(workdir, f"{dataset_name}_{model_name.replace('/', '-')}")
+
+    print(f"Encoding {dataset_name} dataset with {model_name} model")
+    for op, dataset in datasets.items():
+        loader = DataLoader(dataset, batch_size=64, num_workers=4)
+
+        data = []
+        for _, data in enumerate(tqdm(loader, desc=f"({op}) encoding: ")):
+            image, label = data
+
+            image = image.to(device)
+
+            h = encode_fn(image)
+
+            h = h.cpu().numpy()
+            label = label.numpy()
+
+            data.extend(list(zip(h, label)))
+
+        data_name = f"{prefix}_{op}.pkl"
+        with open(data_name, "wb") as f:
+            pickle.dump(data, f)
+
+
+def main(args):
+    model_name = args.model
+    dataset_name = args.dataset
+
+    encode(model_name, dataset_name)
+
+
+if __name__ == "__main__":
+    args = config()
+    main(args)

dataset_lib/models.py

@@ -0,0 +1,49 @@
+import os
+
+import clip
+import open_clip
+import torch
+import torch.amp
+
+SUPPORTED_MODELS = {}
+workdir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+with open(os.path.join(workdir, "supported_models.txt"), "r") as f:
+    for line in f:
+        model, model_path = line.strip().split(",")
+        SUPPORTED_MODELS[model] = model_path
+
+
+def get_transform_and_encoding_fn(
+    model_name, device=torch.device("cuda" if torch.cuda.is_available() else "cpu")
+):
+    if "open_clip" in model_name:
+        return open_clip_get_transform_and_encoding_fn(model_name, device=device)
+    if "clip" in model_name:
+        return clip_get_transform_and_encoding_fn(model_name, device=device)
+
+
+def open_clip_get_transform_and_encoding_fn(model_name, device):
+    model, _, preprocess = open_clip.create_model_and_transforms(
+        SUPPORTED_MODELS[model_name], device=device
+    )
+
+    @torch.no_grad()
+    @torch.cuda.amp.autocast()
+    def encode_fn(image):
+        image_features = model.encode_image(image)
+        return image_features / torch.linalg.norm(image_features, dim=-1, keepdim=True)
+
+    return preprocess, encode_fn
+
+
+def clip_get_transform_and_encoding_fn(model_name, device):
+    backbone_name = model_name.split(":")[-1]
+    model, preprocess = clip.load(backbone_name, device=device)
+
+    @torch.no_grad()
+    @torch.cuda.amp.autocast()
+    def encode_fn(image):
+        image_features = model.encode_image(image)
+        return image_features / torch.linalg.norm(image_features, dim=-1, keepdim=True)
+
+    return preprocess, encode_fn

imagenette_clip:ViT-B-32_train.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19f3c3f597d1aaf7579be3285f9bc83337b9eff2275c411fbfbd8ebd6efa7746
+size 36858462

imagenette_clip:ViT-B-32_val.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13c1c0e3e22f3446e6bcbcb33164ee290a23dbfedd5b30cc82148e4db6cc3d6a
+size 12689531

imagenette_clip:ViT-L-14_train.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5adb5c53898f50e957028da4314af87506165f4aadd74b062d271a9dcf5f2f4e
+size 36920933

imagenette_clip:ViT-L-14_val.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0c31fb9e428bf448f2c85e1596887e74448e48e656530289db5d020bd9327d6
+size 12711046

imagenette_open_clip:ViT-B-32_train.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c946ac593a1681cbae7fe475803183999b9085270e6c8555cfd8092caeba518
+size 36858458

imagenette_open_clip:ViT-B-32_val.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42f3a6c1e00468e22b65736ae9ce17db3299860454ca81c51c831c4147f55928
+size 12689529

imagenette_open_clip:ViT-L-14_train.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d8c47412b0cfbbd74175d695e997b7c8f2d032988c8466c625ee33d3c15e3f4
+size 36920931

imagenette_open_clip:ViT-L-14_val.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23b659cd74dfb48e5b79d633833ac3799734c4b653b0640dc2c3e5aea287cca0
+size 12711042

make_dataset.py

@@ -0,0 +1,9 @@
+from itertools import product
+
+from dataset_lib.datasets import SUPPORTED_DATASETS
+from dataset_lib.encode import encode
+from dataset_lib.models import SUPPORTED_MODELS
+
+if __name__ == "__main__":
+    for model, dataset in product(SUPPORTED_MODELS.keys(), SUPPORTED_DATASETS):
+        encode(model, dataset)

supported_datasets.txt

@@ -0,0 +1 @@
+imagenette

supported_models.txt

@@ -0,0 +1,4 @@
+open_clip:ViT-B-32,hf-hub:laion/CLIP-ViT-B-32-laion2B-s34B-b79K
+open_clip:ViT-L-14,hf-hub:laion/CLIP-ViT-L-14-laion2B-s32B-b82K
+clip:ViT-B/32,
+clip:ViT-L/14,