Unified Pad and F.pad opertion for PIL and Tensor inputs

[vfdev-5]

Addresses #2292

Description:

• Followed initial code from https://github.com/fmassa/vision-1/tree/pad-tensor by @fmassa.
• Updated Pad and F.pad methods to accept PIL Image and torch Tensors.
  • F.pad accepts padding as tuple or list and is casted internally to required type.
• Updated docstring
• Added new tests

[fmassa]

Looks great, thanks a lot!

I have some minor comments, could you open a new issue so that we can track supporting other types of padding_mode in the future for Tensor?

[fmassa]

doc should be updated, as we only support List[int] in torchscript

[vfdev-5]

Actually, we can support both: tuple and list as tuple is casted to list below (L361). Annotation is kept as List for torchscript to pass. In the tests test_pad I check both types: [4, 4] and (2, 2, 2, 2).
Do you want me add some note about that ?

[fmassa]

yeah, I think we should be careful with how we describe our support.
If we say we support tuple but it fails in torchscript, that's a bad user experience.

I would say it might be better to say we only support List, and as a bonus we support Tuple in non-torchscript mode.

Thoughts?

[vfdev-5]

Yes, we need to provide correct information about what is supported and when. In case of pad in torchscript or non-torchscipt mode, both types (tuple, list) are supported :

>>> from torchvision.transforms.functional_tensor import pad
>>> import torch
>>> tensor = torch.randint(0, 255, (3, 10, 12), dtype=torch.uint8)
>>> r1 = pad(tensor, (2, 2), fill=0, padding_mode="constant")
>>> r2 = pad(tensor, [2, 2], fill=0, padding_mode="constant")
>>> r1.eq(r2).all()
tensor(True)
>>> script_pad = torch.jit.script(pad)
>>> r1 = script_pad(tensor, (2, 2), fill=0, padding_mode="constant")
>>> r2 = script_pad(tensor, [2, 2], fill=0, padding_mode="constant")
>>> r1.eq(r2).all()
tensor(True)

Anyway, I'll update the docstring to make it crystal clear :)

[vfdev-5]

@fmassa I updated 3 of 4 comments you mentioned.

Also, it would be good to add support for some of the other methods (which are already implemented in PyTorch btw). One of the reasons why I didn't add it in my initial POC PR was because they have hard-requirements on the input tensor being 4d, so we would need to either unsqueeze / squeeze for now, or extend them in PyTorch so that they support 3d inputs as well (but that should be done with care).

It is for for another PR, right ?

[fmassa]

It is for for another PR, right ?

Yes, doing it in another PR is fine, but it would be good to have an issue open to track it down so that we don't forget (and if you could work on it it would be even better :-))

[vfdev-5]

Yes, doing it in another PR is fine, but it would be good to have an issue open to track it down so that we don't forget (and if you could work on it it would be even better :-))

Yep, done here : #2350

[vfdev-5]

@fmassa I updated docstring of Pad, functional.pad, functional_pil.pad and functional_tensor.pad about what is supported for padding argument trying to unify the API between all functions.
I added compatibility support for padding as [value, ] for functional_pil.pad such that user can call

F.pad(x, padding=[1, ])
F_pil.pad(x, padding=[1, ])
F_t.pad(x, padding=[1, ])
fn = torch.jit.script(F.pad)
fn(x, padding=[1, ])

without checking if x is PIL Image or Tensor.
What do you think ?

[fmassa]

Looks great, thanks a lot @vfdev-5 !