add sequence fill support for ElasticTransform

[pmeier]

According to our docstring we already have it

vision/torchvision/transforms/transforms.py

Lines 2058 to 2059 in 455eda6

	fill (sequence or number): Pixel fill value for the area outside the transformed
	image. Default is ``0``. If given a number, the value is used for all bands respectively.

However, we do

vision/torchvision/transforms/transforms.py

Lines 2107 to 2109 in 455eda6

	if not isinstance(fill, (int, float)):
	raise TypeError(f"fill should be int or float. Got {type(fill)}")
	self.fill = fill

which will fail for sequences.

cc @vfdev-5

[pmeier]

Drive-by since I was looking into the fill support. This seems to be a copy-paste error. Internally we just convert PIL images to tensors

vision/torchvision/transforms/functional.py

Lines 1558 to 1562 in 59dc938

	t_img = img
	if not isinstance(img, torch.Tensor):
	if not F_pil._is_pil_image(img):
	raise TypeError(f"img should be PIL Image or Tensor. Got {type(img)}")
	t_img = pil_to_tensor(img)

and then call the tensor kernel:

vision/torchvision/transforms/functional.py

Lines 1573 to 1578 in 59dc938

	output = F_t.elastic_transform(
	t_img,
	displacement,
	interpolation=interpolation.value,
	fill=fill,
	)

Meaning, there is no difference between both types.

[NicolasHug]

Thanks Philip, some minor comments below. Should we add a non-regression test?

[NicolasHug]

Do we actually need to convert to float?

[pmeier]

Unfortunately, we do due to JIT 🥲 Removing the float conversion from L2110 gives us

import torch.jit

from torchvision import transforms

torch.jit.script(transforms.ElasticTransform(fill=[1]))

[...]
Expected a value of type 'Optional[List[float]]' for argument 'fill' but instead found type 'List[int]'.
[...]

I know this is ugly AF and far from being Pythonic, but given that it is on v1 I really don't want to deal with this any more than I have to.

[pmeier]

I just realized that we don't have any JIT tests for ElasticTransform. Adding some, revealed that ElasticTransform is not scriptable at all.

1. If you use anything that is not a scalar int or float, the regular constructor will fail.
2. If you use an int (default) or float, JIT will scream at you:

import torch.jit

from torchvision import transforms

torch.jit.script(transforms.ElasticTransform(fill=0.0))

[...]
Expected a value of type 'Optional[List[float]]' for argument 'fill' but instead found type 'float'.
[...]

[NicolasHug]

Thanks Philip, just nits, LGTM anyway