Random Anisotropy transform changes the shape of the subject when copy=False

[c-winder]

Is there an existing issue for this?

• I have searched the existing issues

Bug summary

Random Anisotropy transform changes the shape of the subject when copy=False.

Code for reproduction

import torchio as tio
import torch

subject = tio.Subject(
    scan=tio.ScalarImage(tensor=(torch.ones([1, 256, 256, 256])))
)

transform_copy = tio.RandomAnisotropy(
    downsampling=2.5,
    p=1,
    copy=True,
)

transform_no_copy = tio.RandomAnisotropy(
    downsampling=2.5,
    p=1,
    copy=False,
)

shape_original = subject.shape
shape_after_copy_transform = transform_copy(subject).shape
shape_after_no_copy_transform = transform_no_copy(subject).shape

print(f'{shape_original=}')
print(f'{shape_after_copy_transform=}')
print(f'{shape_after_no_copy_transform=}')

Actual outcome

It reduces one of the dimensions, with the dimension changing at random, e.g.

input shape = [1, 256, 256, 256]
output shape may be [1, 103, 256, 256], [1, 256, 103, 256], or [1, 256, 256, 103]

This worked previously when I was using TorchIO: 0.19.7 and PyTorch: 2.0.1.
Error started when I upgraded to TorchIO: 0.20.4 and PyTorch: 2.6

Error messages

shape_original=(1, 256, 256, 256)
shape_after_copy_transform=(1, 256, 256, 256)
shape_after_no_copy_transform=(1, 256, 256, 103)

or 

shape_original=(1, 256, 256, 256)
shape_after_copy_transform=(1, 256, 256, 256)
shape_after_no_copy_transform=(1, 256, 103, 256)

or 

shape_original=(1, 256, 256, 256)
shape_after_copy_transform=(1, 256, 256, 256)
shape_after_no_copy_transform=(1, 103, 256, 256)

Expected outcome

Shape should not change

shape_original=(1, 256, 256, 256)
shape_after_copy_transform=(1, 256, 256, 256)
shape_after_no_copy_transform=(1, 256, 256, 256)

System info

Platform:   Windows-10-10.0.19045-SP0
TorchIO:    0.20.4
PyTorch:    2.6.0+cu126
SimpleITK:  2.4.1 (ITK 5.4)
NumPy:      2.2.2
Python:     3.12.8 | packaged by conda-forge | (main, Dec  5 2024, 14:06:27) [MSC v.1942 64 bit (AMD64)]