face-obfuscated-imagenet.py

import datasets
import os
import tarfile
import json
import urllib.request
from PIL import Image
import io


class FaceObfuscatedImagenet(datasets.GeneratorBasedBuilder):
    """Comprehensive ImageNet dataset collection with face obfuscation support.
    This dataset includes multiple variants such as the full ILSVRC2012 ImageNet (1k),
    ImageNet-100, Imagenette (10), and ImageNet-Mini subsets, alongside their corresponding
    face-obfuscated (blurred) versions. This dataset supports privacy-preserving research
    by providing both original and face-obfuscated images for each subset.
    The face-obfuscated versions are designed to protect privacy while maintaining the
    utility of the dataset for training and evaluation purposes.
    """

    # This could be a more descriptive name, but we keep it simple for now
    VERSION = datasets.Version("1.0.0")

    # URLs for automatic download
    _INET_TRAIN_URL = (
        "https://www.image-net.org/data/ILSVRC/2012/ILSVRC2012_img_train.tar"
    )
    _INET_VAL_URL = "https://www.image-net.org/data/ILSVRC/2012/ILSVRC2012_img_val.tar"
    _INET_DEVKIT_URL = (
        "https://www.image-net.org/data/ILSVRC/2012/ILSVRC2012_devkit_t12.tar.gz"
    )
    _NOFACE_TRAIN_URL = "https://image-net.org/data/ILSVRC/blurred/train_blurred.tar.gz"
    _NOFACE_VAL_URL = "https://image-net.org/data/ILSVRC/blurred/val_blurred.tar.gz"
    _CLASS_INDEX_URL = "https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json"

    # this allows us to use the same class and cache for different "versions" of the dataset"
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(
            name="1k",
            version=VERSION,
            description=(
                "Standard ILSVRC2012 ImageNet dataset containing approximately 1.2 million "
                "training images across 1000 object categories. This widely used benchmark "
                "dataset is designed for image classification and object recognition tasks."
            ),
        ),
        datasets.BuilderConfig(
            name="100",
            version=VERSION,
            description=(
                "ImageNet-100 is a curated subset of the full ImageNet dataset consisting of 100 "
                "carefully selected classes. This smaller dataset enables faster experimentation "
                "while maintaining a diverse set of categories."
            ),
        ),
        datasets.BuilderConfig(
            name="10",
            version=VERSION,
            description=(
                "Imagenette is a subset of 10 easily classified ImageNet classes intended for "
                "quick prototyping and educational purposes. It provides a lightweight alternative "
                "for rapid development."
            ),
        ),
        datasets.BuilderConfig(
            name="mini",
            version=VERSION,
            description=(
                "ImageNet-Mini is a compact subset of ImageNet containing 100 classes with fewer "
                "images per class. It is intended for faster training and benchmarking while "
                "retaining representative class diversity."
            ),
        ),
        datasets.BuilderConfig(
            name="noface-1k",
            version=VERSION,
            description=(
                "Face-obfuscated variant of the standard ILSVRC2012 ImageNet dataset, where all "
                "human faces in images have been blurred or obscured to protect privacy, enabling "
                "privacy-preserving research without compromising classification utility."
            ),
        ),
        datasets.BuilderConfig(
            name="noface-100",
            version=VERSION,
            description=(
                "Face-obfuscated ImageNet-100 subset with the same 100 classes as the standard "
                "ImageNet-100, but with all human faces blurred or masked. Designed to support "
                "privacy-conscious applications and studies."
            ),
        ),
        datasets.BuilderConfig(
            name="noface-10",
            version=VERSION,
            description=(
                "Face-obfuscated Imagenette subset with 10 classes, where images have undergone "
                "face blurring to safeguard personal identities. Suitable for quick prototyping "
                "in privacy-sensitive scenarios."
            ),
        ),
        # datasets.BuilderConfig(
        #     name="noface-mini",
        #     version=VERSION,
        #     description=(
        #         "Face-obfuscated ImageNet-Mini subset with 100 classes, preserving the original "
        #         "class diversity while anonymizing human faces for privacy preservation during "
        #         "model training and evaluation."
        #     ),
        # ),
    ]

    def _info(self):
        # depending on the config, we will set the class names
        with urllib.request.urlopen(self._CLASS_INDEX_URL) as response:
            full_mapping = json.load(response)  # { "0": ["n01440764", "tench"], ... }

        # Build mapping from wnid to class name
        wnid_to_name = {value[0]: value[1] for key, value in full_mapping.items()}

        # Use the mapping to get formatted class names (e.g., "n01440764: tench")
        wnids = self._get_class_names()
        formatted_class_names = [
            f"{wnid}: {wnid_to_name.get(wnid, wnid)}" for wnid in wnids
        ]

        # generic description for the dataset -- specific subsets are indicated by the config name
        return datasets.DatasetInfo(
            description=(
                "Comprehensive ImageNet dataset collection including multiple variants such as "
                "the full ILSVRC2012 ImageNet (1k), ImageNet-100, Imagenette (10), and ImageNet-Mini subsets, "
                "alongside their corresponding face-obfuscated (blurred) versions. "
                "This dataset supports privacy-preserving research by providing both original and "
                "face-obfuscated images for each subset."
            ),
            features=datasets.Features(
                {
                    "image": datasets.Image(),
                    "label": datasets.ClassLabel(names=formatted_class_names),
                    "is_one": datasets.Value("bool"),
                    "is_ten": datasets.Value("bool"),
                }
            ),
            supervised_keys=("image", "label"),  # type: ignore
            homepage="https://image-net.org/face-obfuscation/",
            citation="""@inproceedings{yang2022study,
              title={A Study of Face Obfuscation in ImageNet},
              author={Yang, Kevin and Yau, John and Fei-Fei, Li and Deng, Jia and Russakovsky, Olga},
              booktitle={International Conference on Machine Learning (ICML)},
              year={2022}
            }""",
        )

    def _split_generators(self, dl_manager):
        if dl_manager.is_streaming:
            raise NotImplementedError(
                "Streaming is not yet supported for this dataset."
            )

        # cache location for the dataset downloads
        inet_train_path = (
            "/gpfs/data/shared/imagenet/ILSVRC2012/ILSVRC2012_img_train.tar"
        )
        inet_val_path = "/gpfs/data/shared/imagenet/ILSVRC2012/ILSVRC2012_img_val.tar"
        inet_devkit_archive_path = (
            "/gpfs/data/shared/imagenet/ILSVRC2012/ILSVRC2012_devkit_t12.tar.gz"
        )
        noface_train_path = "/gpfs/data/shared/imagenet_blurred/train_blurred.tar.gz"
        noface_val_path = "/gpfs/data/shared/imagenet_blurred/val_blurred.tar.gz"

        # download if missing (this should only happen once)
        if not os.path.exists(inet_train_path):
            inet_train_path = dl_manager.download(self._INET_TRAIN_URL)
        if not os.path.exists(inet_val_path):
            inet_val_path = dl_manager.download(self._INET_VAL_URL)
        if not os.path.exists(inet_devkit_archive_path):
            inet_devkit_archive_path = dl_manager.download(self._INET_DEVKIT_URL)
        if not os.path.exists(noface_train_path):
            noface_train_path = dl_manager.download(self._NOFACE_TRAIN_URL)
        if not os.path.exists(noface_val_path):
            noface_val_path = dl_manager.download(self._NOFACE_VAL_URL)

        # Define extraction directories
        noface_train_extract_dir = noface_train_path + "_extracted"  # type: ignore
        noface_val_extract_dir = noface_val_path + "_extracted"  # type: ignore

        # Extract train if not already extracted
        if not os.path.exists(noface_train_extract_dir):
            print(
                "Extracting noface training archive... This will take a while, but will make the overall process much faster. "
                "Do not worry if this takes a long time. Be patient, this is a one-time operation. In the future, this will be cached."
            )
            with tarfile.open(noface_train_path, "r:gz") as tar:  # type: ignore
                tar.extractall(path=noface_train_extract_dir)
            print("Noface training extraction done.")

        # Extract val if not already extracted
        if not os.path.exists(noface_val_extract_dir):
            print(
                "Extracting noface validation archive... This may take a while, but will make the overall process much faster. "
                "Do not worry if this takes a long time. Be patient, this is a one-time operation. In the future, this will be cached."
            )
            with tarfile.open(noface_val_path, "r:gz") as tar:  # type: ignore
                tar.extractall(path=noface_val_extract_dir)
            print("Noface alidation extraction done.")

        return self._return_generated_splits(
            inet_train_path,
            inet_val_path,
            inet_devkit_archive_path,
            noface_train_extract_dir,
            noface_val_extract_dir,
        )

    def _generate_examples(self, archive_path, split, devkit_archive=None):
        config_name = self.config.name

        # Depending on the config name, we will generate examples for the appropriate subset
        if config_name == "1k":
            return self._generate_full_imagenet(
                archive_path, split, devkit_archive=devkit_archive
            )
        elif config_name == "100":
            return self._generate_nonblur_subset(
                archive_path,
                split,
                wnid_list=self._inet100_class_names(),
                devkit_archive=devkit_archive,
            )
        elif config_name == "mini":
            return self._generate_nonblur_subset(
                archive_path,
                split,
                wnid_list=self._inet_mini_class_names(),
                devkit_archive=devkit_archive,
            )
        elif config_name == "10":
            return self._generate_nonblur_subset(
                archive_path,
                split,
                wnid_list=self._imagenette_class_names(),
                devkit_archive=devkit_archive,
            )
        elif config_name == "noface-1k":
            train_metadata = "https://huggingface.co/datasets/randall-lab/face-obfuscated-imagenet/resolve/main/metadata/noface_1k/train.txt"
            val_metadata = "https://huggingface.co/datasets/randall-lab/face-obfuscated-imagenet/resolve/main/metadata/noface_1k/val.txt"
            return self._generate_noface(
                archive_path, split, train_metadata, val_metadata
            )
        elif config_name == "noface-100":
            train_metadata = "https://huggingface.co/datasets/randall-lab/face-obfuscated-imagenet/resolve/main/metadata/noface_100/train.txt"
            val_metadata = "https://huggingface.co/datasets/randall-lab/face-obfuscated-imagenet/resolve/main/metadata/noface_100/val.txt"
            return self._generate_noface(
                archive_path, split, train_metadata, val_metadata
            )
        elif config_name == "noface-10":
            train_metadata = "https://huggingface.co/datasets/randall-lab/face-obfuscated-imagenet/resolve/main/metadata/noface_10/train.txt"
            val_metadata = "https://huggingface.co/datasets/randall-lab/face-obfuscated-imagenet/resolve/main/metadata/noface_10/val.txt"
            return self._generate_noface(
                archive_path, split, train_metadata, val_metadata
            )
        # elif config_name == "noface-mini":
        #     # TODO: Update the URLs to point to the correct metadata files
        #     train_metadata = "/path/to/noface_mini/train.txt"
        #     val_metadata = "/path/to/noface_mini/val.txt"
        #     return self._generate_noface(
        #         archive_path, split, train_metadata, val_metadata
        #     )
        else:
            raise ValueError(
                f"Unknown config name: {config_name}"
            )  # this should never happen

    def _return_generated_splits(
        self,
        inet_train_path,
        inet_val_path,
        inet_devkit_path,
        noface_train_extract_dir,
        noface_val_extract_dir,
    ):
        if self.config.name.startswith("noface"):
            # For all blurred (noface) configs
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,  # type: ignore
                    gen_kwargs={
                        "archive_path": noface_train_extract_dir,
                        "split": "train",
                    },
                ),
                datasets.SplitGenerator(
                    name=datasets.Split.VALIDATION,  # type: ignore
                    gen_kwargs={
                        "archive_path": noface_val_extract_dir,
                        "split": "validation",
                    },
                ),
            ]
        else:
            # For all non-blurred (normal) configs
            return [
                datasets.SplitGenerator(
                    name=datasets.Split.TRAIN,  # type: ignore
                    gen_kwargs={"archive_path": inet_train_path, "split": "train"},
                ),
                datasets.SplitGenerator(
                    name=datasets.Split.VALIDATION,  # type: ignore
                    gen_kwargs={
                        "archive_path": inet_val_path,
                        "split": "validation",
                        "devkit_archive": inet_devkit_path,
                    },
                ),
            ]

    def _generate_full_imagenet(self, archive_path, split, devkit_archive=None):
        if split == "train":
            # Open the training tar archive.
            with tarfile.open(archive_path, "r:*") as train_tar:
                # The training archive contains many tar files (one per class).
                sub_tar_members = [
                    m
                    for m in train_tar.getmembers()
                    if m.isfile() and m.name.endswith(".tar")
                ]
                # Download and cache the class mapping (if not already done).
                if not hasattr(self, "_mapping"):
                    with urllib.request.urlopen(self._CLASS_INDEX_URL) as response:
                        self._mapping = json.load(response)
                mapping = self._mapping
                # Build a mapping from WordNet ID (wnid) to integer label.
                wnid_to_label = {mapping[str(i)][0]: i for i in range(1000)}

                example_idx = 0
                for m in sub_tar_members:
                    # Each member’s name is e.g. "n01440764.tar"; extract the wnid.
                    wnid = os.path.splitext(m.name)[0]
                    if wnid not in wnid_to_label:
                        # Skip any unexpected files.
                        continue
                    label = wnid_to_label[wnid]
                    sub_tar_file = train_tar.extractfile(m)
                    # Open the inner tar file containing images for this class.
                    with tarfile.open(fileobj=sub_tar_file, mode="r:*") as sub_tar:
                        for sub_m in sub_tar.getmembers():
                            if sub_m.isfile():
                                img_f = sub_tar.extractfile(sub_m)
                                image_bytes = img_f.read()
                                img = Image.open(io.BytesIO(image_bytes))
                                if img.mode != "RGB":
                                    img = img.convert("RGB")
                                yield example_idx, {
                                    "image": img,
                                    "label": label,
                                    "is_one": False,
                                    "is_ten": False,
                                }
                                example_idx += 1

        elif split == "validation":
            # For validation, open the validation tar archive.
            with tarfile.open(archive_path, "r:*") as val_tar:
                # Get all file members (each a validation image).
                members = [m for m in val_tar.getmembers() if m.isfile()]
                # Sort by filename to ensure the same order as in the ground truth file.
                members = sorted(members, key=lambda m: m.name)
                # Open the devkit archive to extract the ground truth file.
                with tarfile.open(devkit_archive, "r:*") as devkit_tar:
                    gt_member = None
                    for m in devkit_tar.getmembers():
                        if m.name.endswith("ILSVRC2012_validation_ground_truth.txt"):
                            gt_member = m
                            break
                    if gt_member is None:
                        raise ValueError(
                            "Could not find the ground truth file in the devkit archive."
                        )
                    gt_file = devkit_tar.extractfile(gt_member)
                    gt_lines = gt_file.read().decode("utf-8").strip().splitlines()
                if len(gt_lines) != len(members):
                    raise ValueError(
                        "Mismatch between the number of validation images and ground truth labels."
                    )
                for example_idx, (m, gt) in enumerate(zip(members, gt_lines)):
                    # Convert the ground truth label from 1-indexed to 0-indexed.
                    label = int(gt) - 1
                    img_f = val_tar.extractfile(m)
                    image_bytes = img_f.read()
                    img = Image.open(io.BytesIO(image_bytes))
                    if img.mode != "RGB":
                        img = img.convert("RGB")
                    yield example_idx, {
                        "image": img,
                        "label": label,
                        "is_one": False,
                        "is_ten": False,
                    }

    def _generate_nonblur_subset(
        self, archive_path, split, wnid_list, devkit_archive=None
    ):
        wnid_to_new_label = {wnid: idx for idx, wnid in enumerate(wnid_list)}

        if split == "train":
            with tarfile.open(archive_path, "r:*") as train_tar:
                # The training tar archive contains many tar files (one per class)
                sub_tar_members = [
                    m
                    for m in train_tar.getmembers()
                    if m.isfile() and m.name.endswith(".tar")
                ]
                example_idx = 0
                for m in sub_tar_members:
                    # Each member’s name is, e.g., "n01440764.tar"; extract the wnid.
                    wnid = os.path.splitext(m.name)[0]
                    # Only process if the wnid is in our selected subset.
                    if wnid not in wnid_to_new_label:
                        continue
                    new_label = wnid_to_new_label[wnid]
                    sub_tar_file = train_tar.extractfile(m)
                    # Open the inner tar file containing images for this class.
                    with tarfile.open(fileobj=sub_tar_file, mode="r:*") as sub_tar:
                        for sub_m in sub_tar.getmembers():
                            if sub_m.isfile():
                                img_f = sub_tar.extractfile(sub_m)
                                image_bytes = img_f.read()
                                img = Image.open(io.BytesIO(image_bytes))
                                if img.mode != "RGB":
                                    img = img.convert("RGB")
                                yield example_idx, {
                                    "image": img,
                                    "label": new_label,
                                    "is_one": False,
                                    "is_ten": False,
                                }
                                example_idx += 1

        elif split == "validation":
            # For validation, filter examples based on their wnid.
            with tarfile.open(archive_path, "r:*") as val_tar:
                members = [m for m in val_tar.getmembers() if m.isfile()]
                # Sort by filename to match the ground truth order.
                members = sorted(members, key=lambda m: m.name)
                # Open the devkit archive to extract the ground truth file.
                with tarfile.open(devkit_archive, "r:*") as devkit_tar:
                    gt_member = None
                    for m in devkit_tar.getmembers():
                        if m.name.endswith("ILSVRC2012_validation_ground_truth.txt"):
                            gt_member = m
                            break
                    if gt_member is None:
                        raise ValueError(
                            "Could not find the ground truth file in the devkit archive."
                        )
                    gt_file = devkit_tar.extractfile(gt_member)
                    gt_lines = gt_file.read().decode("utf-8").strip().splitlines()

                # Download the full class mapping to convert the original label to wnid.
                with urllib.request.urlopen(self._CLASS_INDEX_URL) as response:
                    full_mapping = json.load(response)
                # Build a mapping from original index (0-indexed) to wnid.
                idx_to_wnid = {
                    int(key): value[0] for key, value in full_mapping.items()
                }

                example_idx = 0
                # Iterate over the validation tar members paired with the ground truth lines.
                for m, gt in zip(members, gt_lines):
                    original_label = int(gt) - 1  # original labels are 1-indexed
                    wnid = idx_to_wnid.get(original_label)
                    # Only include the example if its wnid is in our selected subset.
                    if wnid not in wnid_to_new_label:
                        continue
                    new_label = wnid_to_new_label[wnid]
                    img_f = val_tar.extractfile(m)
                    image_bytes = img_f.read()
                    img = Image.open(io.BytesIO(image_bytes))
                    if img.mode != "RGB":
                        img = img.convert("RGB")
                    yield example_idx, {
                        "image": img,
                        "label": new_label,
                        "is_one": False,
                        "is_ten": False,
                    }
                    example_idx += 1

    def _generate_noface(self, archive_path, split, train_txt_url, val_txt_url):
        def is_valid_file(path):
            return os.path.isfile(path) and os.path.getsize(path) > 0

        num_failed = 0

        if split == "train":
            with urllib.request.urlopen(train_txt_url) as f:
                train_ids = [line.decode("utf-8").strip() for line in f if line.strip()]

            if not hasattr(self, "_mapping"):
                with urllib.request.urlopen(self._CLASS_INDEX_URL) as response:
                    self._mapping = json.load(response)

            wnid_to_label = {
                wnid: idx for idx, wnid in enumerate(self._get_class_names())
            }

            valid_entries = []
            for file_id in train_ids:
                # inputs are like n01440764_2708-0-0
                # Split to get all parts: ["n01440764_2708", "0", "0"]
                parts = file_id.split("-")
                base_file_id = parts[0]  # "n01440764_2708"
                is_one = parts[1] == "1" if len(parts) > 1 else False
                is_ten = parts[2] == "1" if len(parts) > 2 else False

                wnid = base_file_id.split("_")[0]
                img_name = base_file_id
                found = False
                for ext in [".JPEG", ".jpg"]:
                    candidate_path = os.path.join(
                        archive_path, "train_blurred", wnid, img_name + ext
                    )
                    if is_valid_file(candidate_path):
                        valid_entries.append((candidate_path, wnid, is_one, is_ten))
                        found = True
                        break
                if not found:
                    num_failed += 1

            print(f"[train] Skipped {num_failed} invalid or zero-byte files.")

            example_idx = 0
            for img_path, wnid, is_one, is_ten in valid_entries:
                label = wnid_to_label[wnid]
                img = Image.open(img_path)
                if img.mode != "RGB":
                    img = img.convert("RGB")

                yield example_idx, {
                    "image": img,
                    "label": label,
                    "is_one": is_one,
                    "is_ten": is_ten,
                }
                example_idx += 1

        if split == "validation":
            with urllib.request.urlopen(val_txt_url) as f:
                val_ids = [line.decode("utf-8").strip() for line in f if line.strip()]

            if not hasattr(self, "_mapping"):
                with urllib.request.urlopen(self._CLASS_INDEX_URL) as response:
                    self._mapping = json.load(response)

            wnid_to_label = {
                wnid: idx for idx, wnid in enumerate(self._get_class_names())
            }

            valid_entries = []
            for file_id in val_ids:
                wnid = file_id.split("_")[0]
                img_name = file_id[len(wnid) + 1 :]
                found = False
                for ext in [".JPEG", ".jpg"]:
                    candidate_path = os.path.join(
                        archive_path, "val_blurred", wnid, img_name + ext
                    )
                    if is_valid_file(candidate_path):
                        valid_entries.append((candidate_path, wnid))
                        found = True
                        break
                if not found:
                    num_failed += 1

            print(f"[validation] Skipped {num_failed} invalid or zero-byte files.")

            example_idx = 0
            for img_path, wnid in valid_entries:
                label = wnid_to_label[wnid]
                img = Image.open(img_path)
                if img.mode != "RGB":
                    img = img.convert("RGB")

                yield example_idx, {
                    "image": img,
                    "label": label,
                    "is_one": False,
                    "is_ten": False,
                }
                example_idx += 1

    def _get_class_names(self):
        # Load the full ImageNet class index from the JSON
        with urllib.request.urlopen(self._CLASS_INDEX_URL) as response:
            full_mapping = json.load(response)  # { "0": ["n01440764", "tench"], ... }

        if self.config.name in {"1k", "noface-1k"}:
            return [full_mapping[str(i)][0] for i in range(1000)]

        if self.config.name in {"100", "noface-100"}:
            return self._inet100_class_names()
        elif self.config.name in {"10", "noface-10"}:
            return self._imagenette_class_names()
        elif self.config.name in {"mini"}:  # "noface-mini"
            return self._inet_mini_class_names()
        else:
            raise ValueError(f"Unknown config name: {self.config.name}")

        # Build list of formatted class names like "n01558993: robin"
        return [f"{wnid}: {wnid_to_name.get(wnid, wnid)}" for wnid in wnid_list]

    @staticmethod
    def _inet100_class_names():
        return [
            "n02869837",
            "n01749939",
            "n02488291",
            "n02107142",
            "n13037406",
            "n02091831",
            "n04517823",
            "n04589890",
            "n03062245",
            "n01773797",
            "n01735189",
            "n07831146",
            "n07753275",
            "n03085013",
            "n04485082",
            "n02105505",
            "n01983481",
            "n02788148",
            "n03530642",
            "n04435653",
            "n02086910",
            "n02859443",
            "n13040303",
            "n03594734",
            "n02085620",
            "n02099849",
            "n01558993",
            "n04493381",
            "n02109047",
            "n04111531",
            "n02877765",
            "n04429376",
            "n02009229",
            "n01978455",
            "n02106550",
            "n01820546",
            "n01692333",
            "n07714571",
            "n02974003",
            "n02114855",
            "n03785016",
            "n03764736",
            "n03775546",
            "n02087046",
            "n07836838",
            "n04099969",
            "n04592741",
            "n03891251",
            "n02701002",
            "n03379051",
            "n02259212",
            "n07715103",
            "n03947888",
            "n04026417",
            "n02326432",
            "n03637318",
            "n01980166",
            "n02113799",
            "n02086240",
            "n03903868",
            "n02483362",
            "n04127249",
            "n02089973",
            "n03017168",
            "n02093428",
            "n02804414",
            "n02396427",
            "n04418357",
            "n02172182",
            "n01729322",
            "n02113978",
            "n03787032",
            "n02089867",
            "n02119022",
            "n03777754",
            "n04238763",
            "n02231487",
            "n03032252",
            "n02138441",
            "n02104029",
            "n03837869",
            "n03494278",
            "n04136333",
            "n03794056",
            "n03492542",
            "n02018207",
            "n04067472",
            "n03930630",
            "n03584829",
            "n02123045",
            "n04229816",
            "n02100583",
            "n03642806",
            "n04336792",
            "n03259280",
            "n02116738",
            "n02108089",
            "n03424325",
            "n01855672",
            "n02090622",
        ]

    @staticmethod
    def _inet_mini_class_names():
        return [
            "n01532829",
            "n01558993",
            "n01704323",
            "n01749939",
            "n01770081",
            "n01843383",
            "n01855672",
            "n01910747",
            "n01930112",
            "n01981276",
            "n02074367",
            "n02089867",
            "n02091244",
            "n02091831",
            "n02099601",
            "n02101006",
            "n02105505",
            "n02108089",
            "n02108551",
            "n02108915",
            "n02110063",
            "n02110341",
            "n02111277",
            "n02113712",
            "n02114548",
            "n02116738",
            "n02120079",
            "n02129165",
            "n02138441",
            "n02165456",
            "n02174001",
            "n02219486",
            "n02443484",
            "n02457408",
            "n02606052",
            "n02687172",
            "n02747177",
            "n02795169",
            "n02823428",
            "n02871525",
            "n02950826",
            "n02966193",
            "n02971356",
            "n02981792",
            "n03017168",
            "n03047690",
            "n03062245",
            "n03075370",
            "n03127925",
            "n03146219",
            "n03207743",
            "n03220513",
            "n03272010",
            "n03337140",
            "n03347037",
            "n03400231",
            "n03417042",
            "n03476684",
            "n03527444",
            "n03535780",
            "n03544143",
            "n03584254",
            "n03676483",
            "n03770439",
            "n03773504",
            "n03775546",
            "n03838899",
            "n03854065",
            "n03888605",
            "n03908618",
            "n03924679",
            "n03980874",
            "n03998194",
            "n04067472",
            "n04146614",
            "n04149813",
            "n04243546",
            "n04251144",
            "n04258138",
            "n04275548",
            "n04296562",
            "n04389033",
            "n04418357",
            "n04435653",
            "n04443257",
            "n04509417",
            "n04515003",
            "n04522168",
            "n04596742",
            "n04604644",
            "n04612504",
            "n06794110",
            "n07584110",
            "n07613480",
            "n07697537",
            "n07747607",
            "n09246464",
            "n09256479",
            "n13054560",
            "n13133613",
        ]

    @staticmethod
    def _imagenette_class_names():
        return [
            "n01440764",
            "n02102040",
            "n02979186",
            "n03000684",
            "n03028079",
            "n03394916",
            "n03417042",
            "n03425413",
            "n03445777",
            "n03888257",
        ]