fix deprecation warning for datapipes

[pmeier]

Sister PR to pytorch/pytorch#74685. Fixes #322. CI will fail until the change is merged upstream.

[NivekT]

Let's add the "earliest removal version/date" here too:
pytorch/pytorch#74685 (comment)

Aside from that, thanks for doing this!

[pmeier]

Cross posting pytorch/pytorch#74685 (comment)

Note that according to BC policy of PyTorch core, we can only remove functionality if the warning stating the concrete deprecation version was emitted for two releases. Thus, if we deprecate with the next releases torch==1.12 / torchdata==0.5, we can only remove in torch==1.14 / torchdata==0.7.

[NivekT]

Our next release is 0.4 instead of 0.5. Otherwise, it looks good. Thanks!

[NivekT]

The other PR should be landed soon, so this can be merged next week

[pmeier]

We get a bunch of "unclosed file" warnings for Python >= 3.8. They are valid, but unrelated to this PR. See pytorch/vision#5801.

[ejguan]

We get a bunch of "unclosed file" warnings for Python >= 3.8. They are valid, but unrelated to this PR. See pytorch/vision#5801.

Took a quick look on the warning. It seems the warning comes from https://github.com/pytorch/vision/blob/c399c3f4957d997691561b51b0eca182ec7c3cac/torchvision/datasets/sbu.py#L51

Curious about when does this warning start to pop up? Based on git blame, the changes were made 3 years ago. To eliminate it, you may want to explicitly close the opened files in zip.

[ejguan]

It seems we have mypy Error from deprecation_warning on TorchData CI. IIUC, the Error should be eliminated after this PR is landed.

[facebook-github-bot]

@NivekT has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[pmeier]

Took a quick look on the warning. It seems the warning comes from https://github.com/pytorch/vision/blob/c399c3f4957d997691561b51b0eca182ec7c3cac/torchvision/datasets/sbu.py#L51

Nope. You linked an old dataset that is not tested in your CI. SBU with datapipes is not even landed yet (pytorch/vision#5683), so not sure where you got that from.

In any case, this is not linked to a specifc dataset as far as I can tell. See pytorch/vision#5801 for more failures that popped up when we also tested Python >= 3.8 in our CI. Let's discuss there.