[bugfix] Make CUB200 labels 0-indexed.

[kdexd]

Bug description:

CUB200 dataset in torchvision.prototype.datasets module formed labels using file paths. This resulted in labels being 1-indexed (1-200) instead of 0-indexed (0-199). A related issue from the past was with Flowers102 (torchvision.datasets module, #5766).

This PR simply shifts the labels and makes them zero-indexed for consistency with other datasets.

[pmeier]

Thanks @kdexd!

[kdexd]

The pleasure is mine! On a related note, my ongoing project requires evaluating large vision models on a suite of ~25 image classification datasets, specifically a big subset of datasets used to evaluate CLIP models (see Figure below).

Some datasets are not yet implemented in torchvision.prototype.datasets, but are present in torchvision.datasets. In my development code I simply added a wrapper class to have a consistent API with new datasets (IterDataPipe), like so:

class TorchvisionWrapIterDataPipe(dpi.IterDataPipe):
    def __init__(self, name: str, root: str | Path, split: str, **kwargs):
        # Get the dataset from `torchvision.datasets` module.
        DatasetClass = getattr(torchvision.datasets, name)
        self._inner = DatasetClass(str(root), split, download=True, **kwargs)

        # Wrap the dataset as a MapDataPipe, then convert to iterable.
        _dp = MapToIterConverter(SequenceWrapper(self._inner))
        self._dp = hint_sharding(hint_shuffling(_dp))

    def __len__(self):
        return len(self._inner)

    def __iter__(self):
        for image, label in self._dp:
            yield {"image": image, "label": label}


FGVCAircraft = partial(TorchvisionWrapIterDataPipe, "FGVCAircraft")
Flowers102 = partial(TorchvisionWrapIterDataPipe, "Flowers102")
STL10 = partial(TorchvisionWrapIterDataPipe, "STL10", folds=0)
RenderedSST2 = partial(TorchvisionWrapIterDataPipe, "RenderedSST2")

Some other datasets like RESISC45 and HatefulMemes are not present in either modules at all. I believe that these datasets will eventually be commonly used to evaluate CLIP-like models and large SSL models. Torchvision users would find these useful if they are ready for evaluation out-of-the-box. If there is sufficient interest, would you accept pull requests for these datasets? In that case I will actually re-implement a consistent API for these datasets rather than a quick wrapper like shown above.

[pmeier]

If there is sufficient interest, would you accept pull requests for these datasets? In that case I will actually re-implement a consistent API for these datasets rather than a quick wrapper like shown above.

Thanks a lot for the interest in contributing. Could you open an issue so this won't get lost? As of now, our new API is not finalized and thus it makes little sense to port more datasets now. However, we are working hard to get this done and would notify you when we can start adding new datasets.