CropOrPad self.bounds_parameters is set only for first image. #757
Closed
Description
Is there an existing issue for this?
- I have searched the existing issues
Bug summary
When using a mask without target shape CropOrPad will compute the target shape only for the first image used in the transform. The following images will have an incorrect output shape.
Code for reproduction
import numpy as np
import torchio as tio
shape = (1, 5, 5, 5)
mask_a = np.zeros(shape)
mask_a[0, 2, 2, 2] = 1
mask_b = np.zeros(shape)
mask_b[0, 1:4, 1:4, 1:4] = 1
image_a = tio.ScalarImage(tensor=np.ones(shape), affine=np.eye(4))
mask_a = tio.LabelMap(tensor=mask_a, affine=np.eye(4))
subject_a = tio.Subject(image=image_a, mask=mask_a)
image_b = tio.ScalarImage(tensor=np.ones(shape), affine=np.eye(4))
mask_b = tio.LabelMap(tensor=mask_b, affine=np.eye(4))
subject_b = tio.Subject(image=image_b, mask=mask_b)
# Works
shape_a = tio.CropOrPad(mask_name="mask")(subject_a).image.shape
shape_b = tio.CropOrPad(mask_name="mask")(subject_b).image.shape
assert shape_a != shape_b
# Fails
crop = tio.CropOrPad(mask_name="mask")
assert crop(subject_a).image.shape != crop(subject_b).image.shapeActual outcome
The transform outputs an image of shape (1, 1, 1, 1) for subject_a and of (1, 1, 1, 1) for subject_b.
Error messages
No response
Expected outcome
The transform should output an image of shape (1, 1, 1, 1) for subject_a and of (1, 3, 3, 3) for subject_b.
System info
Platform: macOS-10.16-x86_64-i386-64bit
TorchIO: 0.18.62
PyTorch: 1.10.0
SimpleITK: 2.1.1 (ITK 5.2)
NumPy: 1.20.3
Python: 3.8.11 (default, Aug 6 2021, 08:56:27)
[Clang 10.0.0 ]