Repository navigation
NormalizeIntensity inverse fails on non-contiguous inputs with nonzero=True #9182
Description
Describe the bug
The text in this issue was written with AI assistance (Claude and GPT-6 Astra).
NormalizeIntensity(nonzero=True) accepts a transposed image, but inverse() on the result raises a stride error. NormalizeIntensityd and channel_wise=True fail the same way. This blocks restoring the original intensities after normalizing an image whose axes were reordered.
To Reproduce
import torch
from monai.transforms import NormalizeIntensity, NormalizeIntensityd
image = torch.tensor([[[0.0, 1.0], [2.0, 3.0]], [[0.0, 0.0], [0.0, 2.0]]]).transpose(1, 2)
for channel_wise in (False, True):
for dictionary in (False, True):
transform = (
NormalizeIntensityd("image", nonzero=True, channel_wise=channel_wise)
if dictionary
else NormalizeIntensity(nonzero=True, channel_wise=channel_wise)
)
data = {"image": image.clone()} if dictionary else image.clone()
normalized = transform(data)
label = f"{type(transform).__name__}(channel_wise={channel_wise})"
try:
restored = transform.inverse(normalized)
restored = restored["image"] if dictionary else restored
print(label, "round_trip=", torch.allclose(restored, image))
except RuntimeError as error:
print(label, type(error).__name__ + ":", str(error))All four cases print:
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.
Expected behavior
round_trip= True for all four cases, as for a contiguous input.
Environment
MONAI dev (6aa4abf) from source, PyTorch 2.14.1+cpu, NumPy 2.5.4, Python 3.12.13, Linux.
Additional context
The error comes from mask.view(-1)[zeroed_idx] = True in NormalizeIntensity.inverse (added in #8905): the boolean mask keeps the transposed layout. Making that mask contiguous fixes all four cases. I have the fix with regression tests and will open a PR.
Activity