transforms v2 rotate can create different sizes between tensors and PIL with expand=True

[pmeier]

import torch
from torchvision.transforms.v2 import functional as F

input_tensor = torch.rand(3, 17, 11)
input_pil = F.to_pil_image(input_tensor)

kwargs = dict(angle=-10.9, center=(1.2, 4.9), expand=True)

output_tensor = F.rotate(input_tensor, **kwargs)
output_pil = F.rotate(input_pil, **kwargs)

shape_tensor = F.get_image_size(output_tensor)
shape_pil = F.get_image_size(output_pil)

assert shape_tensor == shape_pil, f"{shape_tensor} != {shape_pil}"

AssertionError: [15, 20] != [15, 19]

Replacing the import with from torchvision.transforms import functional as F, i.e. using transforms v1, does not have this problem.

[pmeier]

We had some subtle issues here before, which supposedly should have been fixed by #7271.

[pmeier]

PIL kernel is the culprit:

import torch
from torchvision.transforms.v2 import functional as F

input_tensor = torch.rand(3, 17, 11)
input_pil = F.to_pil_image(input_tensor)

kwargs = dict(angle=-10.9, expand=True)

output_w_center = F.rotate(input_tensor, **kwargs, center=(1.2, 4.9))
output_wo_center = F.rotate(input_pil, **kwargs, center=None)

shape_w_center = F.get_image_size(output_w_center)
shape_wo_center = F.get_image_size(output_wo_center)

assert shape_w_center == shape_wo_center, f"{shape_w_center} != {shape_wo_center}"

AssertionError: [15, 20] != [15, 19]

Meaning, #7271 did its job, but we need to expand it to the PIL kernel as well.

The issue is that center theoretically should have no effect on the output size if expand=True is given. Practically, rounding effects take place which lead the off-by-one errors. The difference between v1 and v2 is that we not only warn users about setting both parameters

vision/torchvision/transforms/v2/functional/_geometry.py

Lines 922 to 923 in 71968bc

	if center is not None and expand:
	warnings.warn("The provided center argument has no effect on the result if expand is True")

we also explicitly unset center in this case

vision/torchvision/transforms/v2/functional/_geometry.py

Line 924 in 71968bc

center = None