Datasets:

amaye15

/

stanford-dogs

like 1

No worries, I'm on holidays right now so I'll correct this in two weeks.

Thanks !

Just had a look. If you can put the mapping here I'll do the rest.

Hi !

In the end, I uploaded the dataset to the hub myself, here is the script I used to have the right labels and class names:

def sdogs_to_hfds(data_dir: Path, to: Path, hub_id: str):
    """Converts the StanfordDogs dataset to the hugginface dataset format.

    Assumes that the data is downloaded `data_dir`. The original data can be
    found at https://vision.stanford.edu/aditya86/ImageNetDogs/.

    The folder pointed by `data_dir` should contain
    - The images (http://vision.stanford.edu/aditya86/ImageNetDogs/images.tar) extracted in the "Images" folder.
    - `train_list.mat` and `test_list.mat` (http://vision.stanford.edu/aditya86/ImageNetDogs/lists.tar) at its root.

    An example of `hub_id` is `myname/my-dataset`
    """

    splits = {
        "train": loadmat(data_dir / "train_list.mat"),
        "test": loadmat(data_dir / "test_list.mat"),
    }

    classes: dict[str, int] = {}
    ordered_classnames: list[str] = [None] * 120
    metadata: dict[str, list[dict]] = {"train": [], "test": []}

    for split_name, split in splits.items():
        for (file_name,), label in zip(split["file_list"][:, 0], split["labels"][:, 0]):
            file_name = Path(file_name)
            label = int(label.item()) - 1

            # read the class label
            class_name = file_name.parent.name
            if class_name in classes:
                assert classes[class_name] == label
            else:
                classes[class_name] = label
                ordered_classnames[label] = class_name

            # copy the files to the right folder
            dest = to / split_name / file_name
            dest.parent.mkdir(parents=True, exist_ok=True)
            shutil.copy(data_dir / "Images" / file_name, dest)

            metadata[split_name].append({"file_name": str(file_name), "label": label})

        pl.from_records(metadata[split_name]).write_csv(
            to / split_name / "metadata.csv"
        )

    # Print the class mapping to include it in the readme
    for class_name, class_idx in classes.items():
        print(f"'{class_idx}': {class_name}")

    # Declare the features of the dataset
    features = Features(
        {
            "image": Image(),
            "label": ClassLabel(
                num_classes=120,
                names=ordered_classnames,
            ),
        }
    )

    dataset = load_dataset("imagefolder", data_dir=to, features=features)
    dataset.push_to_hub(hub_id)

Cheers,
Augustin