RandomGrayscale fails for grayscale tensor inputs

#1505 added rgb_to_grayscale to tensor images and #2586 aligned it with the PIL op. It was missed that the PIL op is a no-op if the input is already grayscale, whereas the tensor op fails:

>>> image_tensor = torch.rand(1, 16, 16)
>>> F.rgb_to_grayscale(image_tensor)
TypeError: Input image tensor permitted channel values are [3], but found 1
>>> image_pil = F.to_pil_image(image_tensor)
>>> F.rgb_to_grayscale(image_pil)
<PIL.Image.Image image mode=L size=16x16 at 0x7F8181384A10>

As the name as well as the docstring

https://github.com/pytorch/vision/blob/d0dede0e09d6d72253080cb366742a270f0ea8cd/torchvision/transforms/functional.py#L1231-L1232

imply, this transform should only handle RGB inputs so that shouldn’t be an issue.

The problem is that the RandomGrayscale transform relies on the no-op behavior:

https://github.com/pytorch/vision/blob/d0dede0e09d6d72253080cb366742a270f0ea8cd/torchvision/transforms/transforms.py#L1613-L1616

>>> transform = transforms.RandomGrayscale(p=1.0)
>>> transform(image_tensor)
TypeError: Input image tensor permitted channel values are [3], but found 1
>>> transform(image_pil)
<PIL.Image.Image image mode=L size=16x16 at 0x7F23E5B4E850>

Note the docstring is conflicting here:

https://github.com/pytorch/vision/blob/d0dede0e09d6d72253080cb366742a270f0ea8cd/torchvision/transforms/transforms.py#L1586-L1587

https://github.com/pytorch/vision/blob/d0dede0e09d6d72253080cb366742a270f0ea8cd/torchvision/transforms/transforms.py#L1595-L1596

BC compatible fix would be to patch RandomGrayscale to be an explicit no-op for grayscale inputs.

cc @vfdev-5 @datumbox

[vfdev-5]

PIL does a copy as well if modes are same: https://github.com/python-pillow/Pillow/blob/92c26a77ca53a2bfbd8804f009c6c8755d0e5a43/src/PIL/Image.py#L945-L946

[pmeier]

Are we going to copy or not?

I would, yes. Given that the op currently won’t return the input tensor, users might rely on this. Given that no one raised an issue (that I know of) about this before, I’m guessing this an edge case anyway. Thus, the potential extra copy shouldn’t be a problem here.