[Datasets] Add ImageFolderDatasource

[bveeramani]

Why are these changes needed?

Popular datasets like ImageNet and Tiny ImageNet are arranged in a specific layout like this:

root/dog/xxx.png
root/dog/xxy.png
root/dog/[...]/xxz.png

root/cat/123.png
root/cat/nsdf3.png
root/cat/[...]/asd932_.png

This PR adds a datasource that reads such datasets.

Related issue number

Closes #23977

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[richardliaw]

For the CUJ that amog posted can we add it in as an example to test in ci?

[bveeramani]

For the CUJ that amog posted can we add it in as an example to test in ci?

No. There are issues with applying TorchVision transformations that are completely unrelated to this PR.

To make this code snippet work, you need to add several workarounds

def preprocess(df):
    preprocess = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    df["image"] = df["image"].map(preprocess)
    return df

def preprocess(df):
    preprocess = transforms.Compose([
        lambda ray_tensor: ray_tensor.to_numpy(),
        transforms.ToTensor(),
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        lambda torch_tensor: torch_tensor.numpy().astype(np.float32)
    ])
    df["image"] = TensorArray([preprocess(image) for image in df["image"]])
    return df

[richardliaw]

@bveeramani sorry, what I mean is we should have an end-to-end test such as Amog's example (with whatever modifications you want to make).

Can you please do that before merging?

[bveeramani]

Can you please do that before merging?

@richardliaw Added an E2E test. Wasn't sure what to test other than that there are no errors

[clarkzinzow]

LGTM, great work!

[jiaodong]

do we have any docs / past discussion about this part? Basically we're assuming we get the label based on user file path, which has to be structured in certain way in order to get the correct one without knobs needed to pass in custom label file or join ?

For example, if i read a s3 bucket with filenames of "dog.jpg", "dog_2.jpg" my dataloader will end up getting these string values by default.

[bveeramani]

Basically we're assuming we get the label based on user file path, which has to be structured in certain way in order to get the correct one without knobs needed to pass in custom label file or join ?

Yeah, that's right. The datasource assumes that the layout is structured in the same way as ImageNet. The functionality of the datasource is based on that of TorchVision's ImageFolder.

For example, if i read a s3 bucket with filenames of "dog.jpg", "dog_2.jpg" my dataloader will end up getting these string values by default.

Yeah, you're right. We don't validate that the label corresponds to a directory. In this case, we could raise an error stating that the folder isn't structured correctly.

Alternatively, if images aren't stored in a directory, we could set the label to None.

If images aren't stored in a sub-directory, then the image's label will be set to `None`.

.. code-block::

    root/dog/xxx.png  # Label is 'dog'
    root/123.jpg.     # Label is `None`