spatial_resample raises TypeError when spatial_size is None and spatial_rank is 1

Describe the bug

spatial_resample passes lambda x: x >= 0 to fall_back_tuple, but that helper's default predicate is lambda x: x and x > 0, whose x and short-circuits on None. The override does not, so a None element raises TypeError instead of falling back to the default — which is the documented behaviour of fall_back_tuple, and its own docstring says so:

>>> fall_back_tuple((-1, None), (32, 32))
(32, 32)

monai/transforms/spatial/functional.py:159:

spatial_size = torch.tensor(
    fall_back_tuple(ensure_tuple(spatial_size)[:spatial_rank], in_spatial_size, lambda x: x >= 0)
)

spatial_size reaches that line as None whenever the caller did not supply one and spatial_rank <= 1: the branch above it, elif spatial_size is None and spatial_rank > 1, is the only thing that replaces None, so with a rank of 1 the None survives into ensure_tuple(None) → (None,) → None >= 0.

To Reproduce

import numpy as np
from monai.data.image_writer import NibabelWriter

w = NibabelWriter()
w.set_data_array(np.random.rand(3, 5), channel_dim=None)
w.set_metadata({"affine": np.diag([1, 1, 1]), "original_affine": np.diag([1.4, 1, 1])})

File "monai/data/image_writer.py", line 604, in set_metadata
File "monai/data/image_writer.py", line 273, in resample_if_needed
File "monai/transforms/spatial/array.py", line 229, in __call__
File "monai/transforms/spatial/functional.py", line 159, in <lambda>
    fall_back_tuple(ensure_tuple(spatial_size)[:spatial_rank], in_spatial_size, lambda x: x >= 0)
TypeError: '>=' not supported between instances of 'NoneType' and 'int'

Instrumenting spatial_resample confirms it is entered with spatial_size=None and img.shape=(1, 3, 5).

Expected behavior

None means "no size given for this axis" and should fall back to the corresponding in_spatial_size entry, exactly as fall_back_tuple's docstring describes.

Screenshots / test impact

This is not a corner case reachable only by hand — it fails 8 tests on current dev:

tests/data/test_nifti_rw.py — test_write_2d, test_write_3d
tests/data/test_image_rw.py — 4 failures + 4 errors across TestRegRes / writer round-trips

Reproduced on unmodified dev (c1240a2d4) in two independent environments:


Python 3.12 / torch 2.13.0 / numpy 2.x	fails
Python 3.10 / torch 2.11.0 (highest version CI tests) / numpy 2.2.6	fails identically

So it is not a new-dependency artifact. git log -L159,159 dates that line to #6068 (Feb 2023).

Environment

Ensuring you use the relevant python executable, please paste the output of:

MONAI version: 1.6.0rc1+58.gc1240a2d4
Python: 3.10 and 3.12 (both affected)
PyTorch: 2.11.0 and 2.13.0 (both affected)
numpy: 2.2.6
nibabel: 5.4.2

Additional context

The narrowest fix is to make the predicate None-safe at the call site, matching the helper's default:

lambda x: x is not None and x >= 0

Worth checking the other fall_back_tuple callers that pass an explicit func for the same hazard — any predicate that does not short-circuit on None inherits it.

Nice catch — the fix you already pinned in the issue body (lambda x: x is not None and x >= 0 at functional.py:159) matches fall_back_tuple's own default short-circuit exactly, so it's a drop-in. I'll put up a PR with that change plus regression tests covering the spatial_rank == 1/spatial_size=None path and the writer round-trip from your repro. I'll also sweep the other fall_back_tuple call sites for the same non-short-circuiting-predicate hazard you flagged and report back what I find.

fix(transforms): guard None spatial_size in spatial_resample (Project…

docs(transforms): update spatial_resample docstrings for None spatial…

docs(transforms): add Returns and Raises sections to spatial_resample…

fix(transforms): guard None spatial_size in spatial_resample (Project…

docs(transforms): update spatial_resample docstrings for None spatial…

docs(transforms): add Returns and Raises sections to spatial_resample…

Some extra findings on the root cause, in case they are useful for the fix already in progress (#9080). I'm not planning to open a competing PR.

Why spatial_rank is 1 here: this looks like a regression from #8765 (spatial_ndim tracking in MetaTensor). In ImageWriter.resample_if_needed (monai/data/image_writer.py), the affine is set on data_array before the channel dim is added:

data_array = convert_to_tensor(data_array, track_meta=True)        # shape (2, 3), no channel yet
data_array.affine = ...                                             # setter clamps spatial_ndim to ndim - 1 = 1
output_array = resampler(data_array[None], ...)                     # now (1, 2, 3), but cached spatial_ndim stays 1

So spatial_resample sees spatial_rank == 1 for a 2D image (debug print inside spatial_resample: shape (1, 2, 3), affine (3, 3), get_spatial_ndim == 1), skips the compute_shape_offset branch, and reaches fall_back_tuple with None.

Bisect (test_nifti_rw.py -k "write_2d or write_3d"):

31419ab (parent of Add explicit spatial_ndim tracking to MetaTensor #8765): 2 passed
26326b5 (Add explicit spatial_ndim tracking to MetaTensor #8765): 2 failed (test_write_2d TypeError, test_write_3d assertion)

Consequence for the None-safe predicate alone: it removes the TypeError, but with spatial_rank == 1 no resampling happens, so the output keeps the input affine (test_write_2d then fails with affine 1.0 vs expected 1.4), and test_write_3d still fails (shape (1, 1, 5) vs expected (1, 1, 3)).

Possible fix (add the channel dim before setting the affine):

-        data_array = convert_to_tensor(data_array, track_meta=True)
+        # add the channel dim before setting the affine, so that the cached spatial_ndim is not clamped
+        data_array = convert_to_tensor(data_array, track_meta=True)[None]
         if affine is not None:
             data_array.affine = convert_to_tensor(affine, track_meta=False)  # type: ignore
         resampler = SpatialResample(mode=mode, padding_mode=padding_mode, align_corners=align_corners, dtype=dtype)
         output_array = resampler(
-            data_array[None], dst_affine=target_affine, spatial_size=output_spatial_shape  # type: ignore
+            data_array, dst_affine=target_affine, spatial_size=output_spatial_shape  # type: ignore
         )

Results on current dev (macOS arm64, Python 3.12.14, torch 2.14.0, nibabel 5.4.2, itk 5.4.7):

test_nifti_rw.py: 39 passed / 2 failed → 41 passed
test_image_rw.py (TestLoadSaveNifti, 8 cases): all failing → all passing
whole tests/data: 10 more tests pass, no new failures (the only remaining failure, test_lmdbdataset_dist, also fails before Add explicit spatial_ndim tracking to MetaTensor #8765 on my machine and looks macOS-multiprocessing related)
test_save_image.py, test_save_imaged.py, test_itk_writer.py, test_png_rw.py: unchanged, all passing

Other code paths that set an affine on a tensor before adding a channel dim might be affected the same way, but I only checked the writer.

Happy for this to be folded into #9080 or used however is most convenient.

(Investigation done with the help of an AI assistant; I reviewed the results.)

fix(transforms): resolve TypeError when spatial_size is None (Project…

Navigation Menu

Repository navigation

spatial_resample raises TypeError when spatial_size is None and spatial_rank is 1 #9068

Describe the bug

To Reproduce

Expected behavior

Screenshots / test impact

Environment

Additional context

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

spatial_resample raises TypeError when spatial_size is None and spatial_rank is 1 #9068

Description

Describe the bug

To Reproduce

Expected behavior

Screenshots / test impact

Environment

Additional context

Activity

MDSALMANSHAMS commented on Aug 27, 2026

markov12 commented on Sep 12, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions