CenterCrop is incompatible with torch.jit.trace

CenterCrop is not traceable

from torchvision.transforms import CenterCrop

c = CenterCrop(size=(112, 112))
torch.jit.trace(c, example_inputs=torch.ones((200, 200)))

TypeError: type Tensor doesn't define __round__ method

The reason being these 2 lines in torchvision.transforms.functional.py (line 590)

crop_top = int(round((image_height - crop_height) / 2.0))
crop_left = int(round((image_width - crop_width) / 2.0))

rewriting this with the val // 2 operator instead of int(round(val / 2.0)) will have exactly the same output, be simpler and traceable.

crop_top = (image_height - crop_height) // 2
crop_left = (image_width - crop_width) // 2

proof:
There are 2 situations:

• val is even
• val is odd

if even the result will be the same because val / 2.0 == something.0 and if odd val / 2.0 == something.5, round rounds down in that case so it will be the same too.

>>> round(2.5)
2

Start in torchvision.transforms.functional.py around line 590, where CenterCrop computes crop_top and crop_left, and reproduce the issue with the torch.jit.trace example from the report. Verify that the revised calculation allows tracing to complete and preserves the expected crop behavior for even and odd differences.