fix: support decollate for numpy scalars

[arthurdjn]

Description

This PR supports numpy scalars (e.g. in the form of np.array(1) ) in the decollate_batch function (fix issue #8471).

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests --disttests.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[KumoLiu]

Thanks for the pr! Do you think it might be beneficial to convert the array into a tensor? This way, the data could be handled more consistently.

[arthurdjn]

We could, I think it does not matter for my use cases. As long as the function handles numpy scalars in the form of an array it is good for me!

I will add this change and convert it as a tensor there (L629) if you prefer :)

[KumoLiu]

Thanks for the quick fix!
May I ask the reason for only convert to tensor when batch.ndim == 0 here?

[arthurdjn]

I noticed a different behavior when using the decollate_batch function on torch tensors vs numpy arrays (see discussion #8472) so I don't want to convert numpy arrays to torch tensors as it will introduce some breaking changes

This PR only address the issue #8471 as I think it was not expected and should be supported (?).

[ericspod]

Could we consider a more complete solution? The issue it seems is that 0-d arrays are iterable but can't be iterated over. We already check for non-iterable things in decollate_batch here. Can we modeify this to correctly pick up when the batch is a 0-d array and just return it in that case? Or return its contents?

[arthurdjn]

Thanks for the feedback. The initial PR was:

if isinstance(batch, (float, int, str, bytes)) or (
    type(batch).__module__ == "numpy" and not isinstance(batch, Iterable)
):
    return batch
if isinstance(batch, np.ndarray) and batch.ndim == 0:
    return batch.item() if detach else batch
# rest ...

Is this something that you find more complete?

Note

I refactored the PR to convert from numpy array to torch tensor as suggested by @KumoLiu.

[ericspod]

Is this something that you find more complete?

What I had in mind was more of the following change:

...
    if batch is None or isinstance(batch, (float, int, str, bytes)):
        return batch
    if getattr(batch, "ndim", -1) == 0:  # assumes only Numpy objects and Pytorch tensors have ndim
        return batch.item() if detach else batch
    if isinstance(batch, torch.Tensor):
        if detach:
            batch = batch.detach()
# REMOVE
#      if batch.ndim == 0:
#          return batch.item() if detach else batch
...

[arthurdjn]

Thanks! I will update the PR to include these changes.