Clean up the documentation of transforms

[datumbox]

📚 Documentation

The methods in functional_pil.py and functional_tensor.py are private but contain a lot of duplicate documentation that is not currently used. It's worth cleaning this up and updating the documentation on functional.py to highlight potential differences or limitations between the two backends.

Here is an example of how we typically highlight backend differences in TorchVision:

vision/torchvision/transforms/transforms.py

Lines 332 to 333 in 1b83f46

	padding_mode (str): Type of padding. Should be: constant, edge, reflect or symmetric.
	Default is constant. Mode symmetric is not yet supported for Tensor inputs.

Moreover as per @voldemortX's comment (see #3071 (comment)), some of the limitations listed in the pydocs are inaccurate and they have already been resolved. These need to be corrected.

[voldemortX]

Hi! In the cleanup process, if you guys find functional limitations of tensor transforms (i.e. something PIL can do that tensors can't), maybe I could work on them?

[datumbox]

Hi @voldemortX, you are very welcome to do so! I think we already mention these limitations on the transforms.py pydoc linked above, if you want to have a go. :)

[voldemortX]

Hi @voldemortX, you are very welcome to do so! I think we already mention these limitations on the transforms.py pydoc linked above, if you want to have a go. :)

Cool! I'll investigate and summarize what can be done soon.

[voldemortX]

@datumbox I checked the current documentation in transforms.py and found the following tensor limitations:

1. Symmetric padding mode support: Pad, RandomCrop (seems already supported in Negative padding for functional_tensor symmetric #2749, Added symmetric padding mode for Tensors #2373 but forget to update the doc).
2. JIT unsupported for Lambda, RandomOrder, RandomChoice (seems not addressable yet).
3. Tensor interpolation mode only support nearest, bilinear, bicubic: Resize, RandomResizedCrop (limited by pytorch interpolate()).
4. Tensor interpolation mode only support nearest, bilinear: RandomPerspective, RandomRotation, RandomAffine (limited by pytorch grid_sample()).

Maybe I can try something for 3 and 4, after some thorough investigation about interpolation modes.

EDIT:
It seems pytorch now supports bicubic in gird_sample() here. I'm not sure whether its wise to wait for the other modes to be supported in pytorch or should I find a workaround.

[datumbox]

@voldemortX Thanks for the deep dive!

Great work flagging the issue on point 1, I'll update this ticket's description to include it. Unfortunately JIT still does not support Lambdas. Given that points 3 & 4 are quite big and not related to this very specific issue, I would recommend opening a RFC issue where you can outline the potential solutions and request feedback.

Finally, if you are interested in fixing the problems listed in this issue, feel free to send a PR.

[voldemortX]

Unfortunately JIT still does not support Lambdas. Given that points 3 & 4 are quite big and not related to this very specific issue, I would recommend opening a RFC issue where you can outline the potential solutions and request feedback.

At the moment I don't have a elegant workaround for interpolations modes, I think maybe if me or somebody want to do this, maybe it's better done in pytorch than in torchvision.

Finally, if you are interested in fixing the problems listed in this issue, feel free to send a PR.

You mean clean-up the docs? I could do that in a bit.