Fix bug on autocontrast when min==max

[datumbox]

Fixes issue reported by @ain-soph at #4995 (comment)

cc @vfdev-5 @datumbox

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit f00737a (more details on the Dr. CI page):

---

• 2/2 failures introduced in this PR

---

2 failures not recognized by patterns:

Job	Step	Action
binary_linux_wheel_py3.7_rocm4.2	packaging/build_wheel.sh	🔁 rerun
binary_linux_conda_py3.9_cu111	packaging/build_conda.sh	🔁 rerun

---

1 job timed out:

• binary_linux_conda_py3.9_cu111

---

This comment was automatically generated by Dr. CI (expand for details).

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[ain-soph]

Thanks for your response, but this fix doesn't seem to be correct?

You are replacing NAN with 1.0, which is not expected.

Constant across channel should still keep the value after autocontrast, rather than replacing them to 1.0.

Am I understanding autocontrast correctly?

[ain-soph]

Just to confirm, current fix seems to implement (channel-wise) normalization as I illustrated in #4995 (comment). Is this expected? Or shall we use (image-wise)?

I'm not sure which behavior is we expect... Not quite familiar with this operation.

[datumbox]

@ain-soph I believe the correct is to do channel-wise.

I've added one more test that shows/confirms this. Below we are comparing the output of PIL vs TorchVision implementation for an image that has constant values across channels:

vision/test/test_transforms.py

Lines 576 to 582 in 35de04c

	def test_autocontrast_equal_minmax():
	img_tensor = torch.tensor([[[10]], [[128]], [[245]]], dtype=torch.uint8).expand(3, 32, 32)
	img_pil = F.to_pil_image(img_tensor)
	img_tensor = F.autocontrast(img_tensor)
	img_pil = F.autocontrast(img_pil)
	torch.testing.assert_close(img_tensor, F.pil_to_tensor(img_pil))

[ain-soph]

That sounds nice.

[fmassa]

LGTM, thanks!