subpixel_crop_2d not returning square shape

[McHaillet]

Hi Alister,

I ran into the following issue with subpixel cropping.

Versions

• torch-fourier-rescale version: 0.0.2
• Python version: 3.11.7
• Operating System: Ubuntu 22.04

Description

The output shape of subpixel_crop_2d is not square. This should highlight and reproduce the issue:

import torch
import einops
from torch_subpixel_crop import subpixel_crop_2d

x = torch.zeros((35, 465, 480))
center = torch.tensor([232,240])
center = einops.repeat(center, "yx -> b yx", b=len(x))
print(subpixel_crop_2d(x, positions=center, sidelength=465).shape)

>>> torch.Size([35, 465, 464])

while the expected shape is (35, 465, 465).

Perhaps a problem with uneven dimensions?

[alisterburt]

Huh, thanks for the report! Yes I would assume this is the odd box size...

I'm actually not sure there is a good way to be consistent with FFTs of odd dimensions, do you think it's worth trying to support?

At a minimum we should explicitly disallow odd dims it if we don't support it

[McHaillet]

The way it can be done is to use the s (shape) argument of irfftn. That way you can ask it to return the original shape.

So you think this issue originates from the fourier_shift_image_2d() function?

At a minimum we should explicitly disallow odd dims it if we don't support it

For sure!

BTW, the shape I started with (35, 465, 480) was output from fourier_rescale.