Update cuCIM for consistency with scikit-image 0.26 by grlee77 · Pull Request #1038 · rapidsai/cucim

grlee77 · 2026-03-04T14:12:23Z

This PR ports recent API and implementation changes from scikit-image 0.26.

Key Changes

There is one new keyword argument, boundary_mode added to binary_blobs.
There was some minor refactoring in regionprops and addition of one new property, "intensity_median".
There are some new deprecations / removals
- Previously deprecated shift_x and shift_y arguments to erosion and dilation are now removed
- For remove_small_objects, the parameter min_size is deprecated in favor of max_size
- For remove_small_holes, the parameter area_threshold is deprecated in favor of max_size
many misc docstring fixes
large refactor of geometric transforms code to match the scikit-image 0.26 behavior/API
- most changes to _shared/utils.py are additional helpers related to this refactor (also ported from scikit-image 0.26)
cucim.skimage.segmentation also now has a separate __init__.pyi for lazy loading
some existing test cases now load the local copies of test data from the repository instead of relying on fetch (fetch no longer works for these files in scikit-image 0.26).

Possibly follow-up tasks

provide batch implementation for intensity_median in _regionprops_gpu.py
make decision on deprecating cucim.skimage.morphology.binary_* functions
- These are deprecated now in scikit-image 0.26.
- For cuCIM, the binary variants are around 2x faster so I kept them for now. A follow-up task would be to update the grayscale variants to automatically dispatch to the faster binary kernel when possible. Then we could remove the separate binary_* functions without losing performance.

Directory layout change in scikit-image 0.26

scikit-image has switched to separate src/skimage and tests/skimage folder layout.
I have currently kept the existing folder layout in cuCIM (tests subfolder within each cucim.skimage module)

warn and clamp blob size if too small add new boundary_mode kwarg

One point of difference is that I did not yet mark binary_* morphology functions as deprecated. For cuCIM there is around 2x performance benefit to using these over the grayscale ones.

some data is now loaded from local files instead of via fetch

…it-image 0.26

bump minimum version requirement to scikit-image 0.23 (released april 2024)

these checks are non-critical and were passing locally but failing on CI

gigony

Thanks @grlee77 ,
I found minor issues and typos.
Otherwise, it looks good to me.

python/cucim/src/cucim/skimage/measure/_regionprops_gpu_convex.py

python/cucim/src/cucim/skimage/exposure/exposure.py

python/cucim/src/cucim/skimage/morphology/isotropic.py

python/cucim/src/cucim/skimage/measure/_regionprops.py

python/cucim/src/cucim/skimage/morphology/isotropic.py

python/cucim/src/cucim/skimage/transform/_geometric.py

python/cucim/src/cucim/skimage/_shared/utils.py

python/cucim/src/cucim/skimage/measure/_regionprops.py

python/cucim/src/cucim/skimage/morphology/misc.py

Co-authored-by: Gigon Bae <gigony@gmail.com>

…ve_small_holes and remove_small_objects is used

Co-authored-by: Gigon Bae <gigony@gmail.com>

…timate. Also fix existing bug in EssentialMatrixTransform (missing forward of xp kwarg to parent class constructor)

… failure

grlee77 · 2026-03-17T03:20:49Z

/merge

This PR ports recent API and implementation changes from scikit-image 0.26. ### Key Changes - There is one new keyword argument, `boundary_mode` added to `binary_blobs`. - There was some minor refactoring in `regionprops` and addition of one new property, "intensity_median". - There are some new deprecations / removals - Previously deprecated `shift_x` and `shift_y` arguments to `erosion` and `dilation` are now removed - For `remove_small_objects`, the parameter `min_size` is deprecated in favor of `max_size` - For `remove_small_holes`, the parameter `area_threshold` is deprecated in favor of `max_size` - many misc docstring fixes - large refactor of geometric transforms code to match the scikit-image 0.26 behavior/API - most changes to `_shared/utils.py` are additional helpers related to this refactor (also ported from scikit-image 0.26) - `cucim.skimage.segmentation` also now has a separate `__init__.pyi` for lazy loading - some existing test cases now load the local copies of test data from the repository instead of relying on `fetch` (fetch no longer works for these files in scikit-image 0.26). ### Possibly follow-up tasks - provide batch implementation for `intensity_median` in `_regionprops_gpu.py` - make decision on deprecating `cucim.skimage.morphology.binary_*` functions - These are deprecated now in scikit-image 0.26. - For cuCIM, the binary variants are around 2x faster so I kept them for now. A follow-up task would be to update the grayscale variants to automatically dispatch to the faster binary kernel when possible. Then we could remove the separate `binary_*` functions without losing performance. ### Directory layout change in scikit-image 0.26 - scikit-image has switched to separate `src/skimage` and `tests/skimage` folder layout. - I have currently kept the existing folder layout in cuCIM (`tests` subfolder within each `cucim.skimage` module) Authors: - Gregory Lee (https://github.com/grlee77) Approvers: - Gil Forsyth (https://github.com/gforsyth) - Gigon Bae (https://github.com/gigony) URL: rapidsai#1038

Changes merged into `main` in 20fec39 that are needed for the 26.04 release if we want Python 3.14 support. Cherry-picked to avoid the 26.06 tag on the `main` branch. Authors: - Gil Forsyth (https://github.com/gforsyth) - Gregory Lee (https://github.com/grlee77) Approvers: - Gregory Lee (https://github.com/grlee77) - James Lamb (https://github.com/jameslamb) URL: #1047

jakirkham

Thanks Greg! 🙏

Had looked over this previously and had a few questions. Finished out this review. Please let me know if you have questions on any of this

jakirkham · 2026-03-04T21:06:56Z

python/cucim/src/cucim/skimage/_shared/tests/test_utils.py

+# SPDX-FileCopyrightText: Copyright (c) 2021-2026, NVIDIA CORPORATION. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0 AND BSD-3-Clause
-
+import re


Thinks we should keep this spaced out from the license header

Suggested change

import re

import re

jakirkham · 2026-03-04T21:08:06Z

python/cucim/src/cucim/skimage/_shared/compat.py

 # check CuPy instead of SciPy
 CUPY_LT_14 = parse(cp.__version__) < parse("14.0.0a1")

+
 # Starting in SciPy v1.12, 'scipy.sparse.linalg.cg' keyword argument `tol` is


Just out-of-curiosity was this to address a linter issue or something?

jakirkham · 2026-03-04T21:10:19Z

python/cucim/src/cucim/skimage/color/tests/test_delta_e.py

-    path = fetch("color/tests/ciede2000_test_data.txt")
-    return np.loadtxt(path, dtype=dtype)


Why do we no longer need to fetch this test data? Is that done somewhere else?

jakirkham · 2026-03-04T21:12:37Z

python/cucim/src/cucim/skimage/exposure/exposure.py

    image : array
        Input image.
-    source_range : string
+    source_range : {'image', 'dtype'}


So is this a dict now?

jakirkham · 2026-03-04T21:14:36Z

python/cucim/src/cucim/skimage/exposure/exposure.py

    Examples
    --------
-    >>> from skimage import data
+    >>> from skimage as ski
    >>> from cucim.skimage import exposure, img_as_float
-    >>> image = img_as_float(cp.array(data.moon()))
+    >>> image = img_as_float(cp.array(ski.data.moon()))


Was scikit-image's layout changed recently?

jakirkham · 2026-03-17T16:28:24Z

python/cucim/src/cucim/skimage/morphology/tests/test_gray.py

-@pytest.mark.parametrize("func", [morphology.erosion, morphology.dilation])
-@pytest.mark.parametrize("name", ["shift_x", "shift_y"])
-@pytest.mark.parametrize("value", [True, False, None])
-def test_deprecated_shift(func, name, value):
-    img = cp.ones(10)
-    func(img)  # Shouldn't warn
-
-    regex = "`shift_x` and `shift_y` are deprecated"
-    with pytest.warns(FutureWarning, match=regex) as record:
-        func(img, **{name: value})
-    assert_stacklevel(record)


Guessing the deprecation happened hence we can drop this test

jakirkham · 2026-03-17T16:29:58Z

python/cucim/src/cucim/skimage/morphology/gray.py

    is_footprint_sequence = _footprint_is_sequence(footprint)
-    footprint = _shift_footprints(footprint, shift_x, shift_y)
-
-    # if isinstance(footprint, tuple) and not is_footprint_sequence:
-    #     if len(footprint) != image.ndim:
-    #         raise ValueError(
-    #             "footprint.ndim={len(footprint)}, image.ndim={image.ndim}"
-    #         )
-    #     if image.ndim == 2 and any(s % 2 == 0 for s in footprint):
-    #         # only odd-shaped footprints are properly handled for tuples
-    #         footprint = cp.ones(footprint, dtype=bool)
-    #     else:
-    #         ndi.grey_erosion(image, size=footprint, output=out)
-    #         return out


This is dropped because it has been unused?

Why did we not need to use it previously?

jakirkham · 2026-03-17T16:31:43Z

python/cucim/src/cucim/skimage/transform/tests/test_geometric.py

-def test_3d_similarity_estimation():
-    # Note: some choices of seed will produce poorly conditioned values,
-    #       leading to NaN in the estimate. We fix the seed here to avoid this.
-    seed = 10
-    rng = cp.random.default_rng(seed)
-    src_points = rng.random(size=(1000, 3))


Why was this test dropped?

jakirkham · 2026-03-17T16:32:39Z

python/cucim/src/cucim/skimage/transform/tests/test_geometric.py

-@pytest.mark.parametrize("xp", [np, cp])
-def test_similarity_transform_params(xp):
-    with pytest.raises(ValueError):
-        _ = SimilarityTransform(
-            translation=(4, 5, 6, 7), dimensionality=4, xp=xp
-        )
-    tf = SimilarityTransform(scale=4, dimensionality=3, xp=xp)
-    assert_array_equal(tf(xp.asarray([[1, 1, 1]])), xp.asarray([[4, 4, 4]]))


Why was this test dropped?

jakirkham · 2026-03-17T16:35:31Z

conda/environments/all_cuda-131_arch-x86_64.yaml

 - pywavelets>=1.6
 - recommonmark
- scikit-image>=0.19.0,<0.26.0
+- scikit-image>=0.23.2,<0.27.0


Looks like this was missed in the Conda recipe. We are handling this in PR: #1049

xref: #1049 (comment)

grlee77 added 18 commits March 4, 2026 09:10

update binary_blobs

b761d36

warn and clamp blob size if too small add new boundary_mode kwarg

docstring fixes in skimage.color module

8d15e8d

docstring and style updates to skimage.feature

4a935b0

regionprops updates for consisitency with skimage v0.26

e3082c8

update _shared.utils to be consistent with skimage 0.26

f87598a

update morphology module for consistency with skimage v0.26

261d96d

One point of difference is that I did not yet mark binary_* morphology functions as deprecated. For cuCIM there is around 2x performance benefit to using these over the grayscale ones.

update restoration and registration modules

4e0d324

some data is now loaded from local files instead of via fetch

add __init__.pyi to segmentation module

5cf9d5a

add __init__.pyi to segmentation module

463276d

update _shared module

71ba599

minor cleanup in transform module

aff4e92

update util module

9404e21

overhaul geometric transform implementation for consistency with scik…

55678de

…it-image 0.26

update _shared.utils tests

7097797

udpate file loading for test_delta_e.py

046b8f0

fix failing test cases

bf01210

bump allowed scikit-image version

6b17bef

typo fix

83b2288

grlee77 added this to the v26.04.00 milestone Mar 4, 2026

grlee77 requested review from a team as code owners March 4, 2026 14:12

grlee77 added the breaking Introduces a breaking change label Mar 4, 2026

grlee77 requested a review from AyodeAwe March 4, 2026 14:12

grlee77 added the improvement Improves an existing functionality label Mar 4, 2026

grlee77 added this to cucim Mar 4, 2026

revert unintended change to benchmarks file

20bd853

grlee77 mentioned this pull request Mar 4, 2026

Add support for Python 3.14 #1033

Closed

grlee77 added 2 commits March 4, 2026 10:03

fix data fetch across scikit-image versions

f3040ed

bump scikit-image minimum version requirement

e711db0

bump minimum version requirement to scikit-image 0.23 (released april 2024)

update warnings filters in test_deprecated_properties

2095929

grlee77 force-pushed the grelee/skimage-0.26-update branch from f315bed to 2095929 Compare March 4, 2026 19:59

remove assert_stacklevel checks

6185c32

these checks are non-critical and were passing locally but failing on CI

gforsyth approved these changes Mar 11, 2026

View reviewed changes

gforsyth removed the request for review from AyodeAwe March 11, 2026 17:00

gigony assigned grlee77 Mar 16, 2026

gigony approved these changes Mar 16, 2026

View reviewed changes

grlee77 and others added 2 commits March 16, 2026 20:55

Apply suggestions from code review

46ce831

Co-authored-by: Gigon Bae <gigony@gmail.com>

Restore desired legacy behavior when the deprecated parameter to remo…

b8df259

…ve_small_holes and remove_small_objects is used

grlee77 mentioned this pull request Mar 17, 2026

[FEA] add intensity_median support to regionprops_table #1045

Open

grlee77 and others added 3 commits March 16, 2026 22:17

Apply suggestions from code review

a93f2f4

Co-authored-by: Gigon Bae <gigony@gmail.com>

preserve expected array type (numpy vs cupy) for identity and from_es…

daf3a51

…timate. Also fix existing bug in EssentialMatrixTransform (missing forward of xp kwarg to parent class constructor)

use fixed seed for random generator to avoid possible stochastic test…

d45173b

… failure

rapids-bot bot merged commit 20fec39 into rapidsai:main Mar 17, 2026
59 checks passed

github-project-automation bot moved this to Done in cucim Mar 17, 2026

gforsyth mentioned this pull request Mar 17, 2026

python 3.14 rev2 #1046

Closed

jakirkham reviewed Mar 17, 2026

View reviewed changes

jakirkham mentioned this pull request Mar 17, 2026

Add support for Python 3.14 #1049

Merged

		path = fetch("color/tests/ciede2000_test_data.txt")
		return np.loadtxt(path, dtype=dtype)

Conversation

grlee77 commented Mar 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Key Changes

Possibly follow-up tasks

Directory layout change in scikit-image 0.26

Uh oh!

gigony left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

grlee77 commented Mar 17, 2026

Uh oh!

Uh oh!

jakirkham left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jakirkham Mar 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

grlee77 commented Mar 4, 2026 •

edited

Loading

jakirkham Mar 17, 2026 •

edited

Loading