enable get_params alias for transforms v2

[pmeier]

After #7135, this is trivial to achieve. If available, we simply alias get_params from the v1 transformation to v2.

Although this doesn't sound like much, it is the final step of getting v2 BC with v1. Meaning, leaving bugs aside, v2 should now be a proper drop-in replacement for v1 🎉

cc @vfdev-5 @bjuncek

[NicolasHug]

As an alternative to __init_subclass__() (of which I keep forgetting the existence and the purpose), would this work too?

@staticmethod
def get_params(cls):
   if cls._v1_transform_cls is not None and hasattr(cls._v1_transform_cls, "get_params"):
        return cls._v1_transform_cls.get_params()
    else:
        raise AttributeError(
            "cls {cls} has no get_params method. You probably don't need one anymore
             as the same RNG is applied to all images, bboxes and masks in the same transform call.
             If what you need is a way to transform different batches with the same RNG,
             please reach out at #1234567 (the feedback issue.
         ")

[pmeier]

Two problems here with JIT:

1. get_params takes parameters in v1, e.g.

   vision/torchvision/transforms/transforms.py

   Lines 617 to 618 in 7cf0f4c

   	@staticmethod
   	def get_params(img: Tensor, output_size: Tuple[int, int]) -> Tuple[int, int, int, int]:

   Since I'm not aware of an arbitrary parameter passthrough like (*args: Any, **kwargs: Any) that works with JIT, we would have to implement this manually on every class that needs it.

2. You used the @staticmethod decorator, but take cls as first input. Since you need cls, we need to switch to @classmethod to make it work. That should work in eager mode, but for JIT this now means that the whole class needs to be scriptable. For example, manually defining get_params for RandomCrop as explained in 1. and trying to script it, leads to

   RuntimeError: 
   'Tensor (inferred)' object has no attribute or method '_v1_transform_cls'.:
     File "/home/philip/git/pytorch/vision/torchvision/prototype/transforms/_geometry.py", line 437
       @classmethod
       def get_params(cls, img: torch.Tensor, output_size: Tuple[int, int]) -> Tuple[int, int, int, int]:
           return cls._v1_transform_cls.get_params(img, output_size)
                  ~~~~~~~~~~~~~~~~~~~~~ <--- HERE
   

   It seems it can't infer the type of cls and uses torch.Tensor as fallback. Annotation cls: Type[RandomCrop] yields

   RuntimeError: 
   Unknown type constructor Type:
   

I agree using __init_subclass__ is unconventional, but it seems like the cleanest solution here. Since we actually alias the function, we avoid all of the JIT crazyness that we would have to deal with otherwise. If you can find a working solution, I'm happy to adopt it though.

[NicolasHug]

Thanks for the details. My last try would be to declare get_params as a @property but... let me guess... JIT doesn't support it?

[pmeier]

You guessed right. But even if did, it would wouldn't work for us here. @property needs an instance.

[NicolasHug]

Thanks Philip, LGTM, just some minor comments / Qs before approving

[NicolasHug]

Thanks a lot Philip