allow len 1 sequences for fill with PIL

[pmeier]

The length 1 sequences for fill were just a workaround back when JIT didn't support Union. Since PIL is not restricted by JIT, we never added support for it.

This means however, that for example F.rotate(img_tensor, fill=[127]) works, while F.rotate(img_pil, fill=[127]) doesn't.

That makes parametrized tests harder, e.g.

• vision/test/transforms_v2_dispatcher_infos.py

  Lines 135 to 138 in a3ba01a

  	xfails_pil_if_fill_sequence_needs_broadcast = xfails_pil(
  	"PIL kernel doesn't support sequences of length 1 for `fill` if the number of color channels is larger.",
  	condition=fill_sequence_needs_broadcast,
  	)

• vision/test/test_transforms_v2_refactored.py

  Lines 2584 to 2585 in a3ba01a

  	if isinstance(input, PIL.Image.Image) and isinstance(value, (tuple, list)) and len(value) == 1:
  	pytest.xfail("F._pad_image_pil does not support sequences of length 1 for fill.")

This PR adds support for length 1 sequences of fill for the PIL backend. Note, that I've implemented the fix in v1, since all our v2 PIL ops with a fill parameter are thin wrappers around the v1 kernel. LMK if that is ok. Otherwise we can also do it in a more verbose way in v2.

cc @vfdev-5

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/7928

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 17 New Failures

As of commit 81471c8 with merge base 439c5e3 ():

NEW FAILURES - The following jobs have failed:

• linux (linux.g5.4xlarge.nvidia.gpu, cuda, 11.8) / linux-job (gh)
• macos (macos-12) / macos-job (gh)
• macos (macos-m1-12) / macos-job (gh)
• unittests-linux (3.8, linux.g5.4xlarge.nvidia.gpu, cuda, 11.8) / linux-job (gh)
• unittests-macos (3.10, macos-12) / macos-job (gh)
• unittests-macos (3.11, macos-12) / macos-job (gh)
• unittests-macos (3.8, macos-12) / macos-job (gh)
• unittests-macos (3.8, macos-m1-12) / macos-job (gh)
• unittests-macos (3.9, macos-12) / macos-job (gh)
• unittests-prototype (3.8, linux.g5.4xlarge.nvidia.gpu, cuda, 11.8) / linux-job (gh)
• unittests-windows (3.10, windows.4xlarge, cpu) / windows-job (gh)
• unittests-windows (3.11, windows.4xlarge, cpu) / windows-job (gh)
• unittests-windows (3.8, windows.4xlarge, cpu) / windows-job (gh)
• unittests-windows (3.8, windows.g5.4xlarge.nvidia.gpu, cuda, 11.8) / windows-job (gh)
• unittests-windows (3.9, windows.4xlarge, cpu) / windows-job (gh)
• windows (windows.4xlarge, cpu) / windows-job (gh)
• windows (windows.g5.4xlarge.nvidia.gpu, cuda, 11.8) / windows-job (gh)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[NicolasHug]

LGTM if green, thanks. Let's just make sure we don't need to update any docstring before merging

[pmeier]

Docstrings for v1 look this:

vision/torchvision/transforms/transforms.py

Lines 771 to 772 in a3ba01a

	fill (sequence or number): Pixel fill value for the area outside the transformed
	image. Default is ``0``. If given a number, the value is used for all bands respectively.

vision/torchvision/transforms/functional.py

Lines 736 to 741 in a3ba01a

	fill (sequence or number, optional): Pixel fill value for the area outside the transformed
	image. If given a number, the value is used for all bands respectively.
	.. note::
	In torchscript mode single int/float value is not supported, please use a sequence
	of length 1: ``[value, ]``.

• Only the functional mentions the sequence of length 1 format at all
• Neither the transform nor the functional mention that one can use it for eager as well

IMO, we don't need to update the docs here. The fact that PIL also supports length 1 sequences is more a QoL thing. I don't expect users to actively use it. But if they do, it is nice that the PIL and Tensor backend have the same behavior.