introduce temporary numpy 2.0 compatibility fixes

should be able to revert these changes after CuPy v13.3 is released

python/cucim/src/cucim/core/operations/morphology/tests/test_distance_transform.py

@@ -6,6 +6,7 @@
 import scipy.ndimage as ndi_cpu
 
 from cucim.core.operations.morphology import distance_transform_edt
+from cucim.skimage._shared.compat import _full
 
 
 def binary_image(shape, pct_true=50):
@@ -137,7 +138,7 @@ def test_distance_transform_edt_block_params_invalid(block_params):
 @pytest.mark.parametrize("ndim", [2, 3])
 def test_distance_transform_edt_uniform_valued(value, ndim):
  """ensure default block_params is robust to anisotropic shape."""
- img = cp.full((48,) * ndim, value, dtype=cp.uint8)
+ img = _full((48,) * ndim, value, dtype=cp.uint8)
  # ensure there is at least 1 pixel at background intensity
  img[(slice(24, 25),) * ndim] = 0
  out = distance_transform_edt(img)

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

@@ -1,5 +1,6 @@
 """Compatibility helpers for dependencies."""
 
+import cupy as cp
 import numpy as np
 from packaging.version import parse
 
@@ -28,3 +29,12 @@
 # deprecated in favor of `rtol`.
 # As of CuPy 13.0, it is still always using 'tol''
 SCIPY_CG_TOL_PARAM_NAME = "tol" # if CUPY_LT_14 else "rtol"
+
+
+def _full(shape, fill_value, dtype=None, order="C"):
+ if NUMPY_LT_2_0_0:
+ return cp.full(shape, fill_value, dtype, order)
+ else:
+ out = cp.empty(shape, dtype=dtype, order=order)
+ out[:] = fill_value
+ return out

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

@@ -743,7 +743,7 @@ def check_random_state(seed):
  if seed is None or seed is cp.random:
  return cp.random.mtrand._rand
  if isinstance(seed, (numbers.Integral, cp.integer)):
- return cp.random.RandomState(seed)
+ return cp.random.RandomState(cp.uint32(seed))
  if isinstance(seed, cp.random.RandomState):
  return seed
  raise ValueError(

python/cucim/src/cucim/skimage/_vendored/_ndimage_filters.py

@@ -5,6 +5,7 @@
 import cupy
 import numpy
 
+from cucim.skimage._shared.compat import _full
 from cucim.skimage._vendored import (
  _internal as internal,
  _ndimage_filters_core as _filters_core,
@@ -361,7 +362,7 @@ def uniform_filter1d(
  from SciPy due to floating-point rounding of intermediate results.
  """
  weights_dtype = cupy.promote_types(input.dtype, cupy.float32)
- weights = cupy.full(size, 1 / size, dtype=weights_dtype)
+ weights = _full(size, 1 / size, dtype=weights_dtype)
  return correlate1d(
  input, weights, axis, output, mode, cval, origin, algorithm=algorithm
  )
@@ -410,9 +411,7 @@ def uniform_filter(
 
  def get(size):
  return (
- None
- if size <= 1
- else cupy.full(size, 1 / size, dtype=weights_dtype)
+ None if size <= 1 else _full(size, 1 / size, dtype=weights_dtype)
  ) # noqa
 
  return _run_1d_correlates(

python/cucim/src/cucim/skimage/color/colorconv.py

@@ -55,6 +55,7 @@
 import numpy as np
 from scipy import linalg
 
+from .._shared.compat import _full
 from .._shared.utils import (
  _supported_float_type,
  channel_as_last_axis,
@@ -71,6 +72,9 @@
  from numpy.exceptions import AxisError
 
 
+using_numpy2 = np.__version__.split(".")[0] == 2
+
+
 def convert_colorspace(arr, fromspace, tospace, *, channel_axis=-1):
  """Convert an image array to a new color space.
 
@@ -1131,7 +1135,7 @@ def gray2rgba(image, alpha=None, *, channel_axis=-1, check_alpha=True):
  f'{image.dtype.name}',
  stacklevel=2
  )
- alpha_arr = cp.full(image.shape, alpha, dtype=image.dtype)
+ alpha_arr = _full(image.shape, alpha, dtype=image.dtype)
  else:
  alpha_arr = cp.asarray(alpha).astype(image.dtype, copy=False)
  if check_alpha and not cp.array_equal(alpha, alpha_arr):

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

@@ -6,9 +6,12 @@
 from cupy.testing import assert_array_almost_equal, assert_array_equal
 from numpy.testing import assert_no_warnings
 
+from cucim.skimage._shared.compat import _full
 from cucim.skimage._shared.testing import expected_warnings
 from cucim.skimage.color.colorlabel import hsv2rgb, label2rgb, rgb2hsv
 
+using_numpy2 = np.__version__.split(".")[0] == 2
+
 
 def test_shape_mismatch():
  image = cp.ones((3, 3))
@@ -213,7 +216,7 @@ def test_avg(channel_axis):
 
 def test_negative_intensity():
  labels = cp.arange(100).reshape(10, 10)
- image = cp.full((10, 10), -1, dtype="float64")
+ image = _full((10, 10), -1, dtype="float64")
  with pytest.warns(UserWarning):
  label2rgb(labels, image, bg_label=-1)
 

python/cucim/src/cucim/skimage/feature/tests/test_canny.py

@@ -51,7 +51,7 @@ def test_01_01_circle(self):
  def test_01_02_circle_with_noise(self):
  """Test that the Canny filter finds the circle outlines
  in a noisy image"""
- cp.random.seed(0)
+ cp.random.seed(cp.uint32(0))
  i, j = cp.mgrid[-200:200, -200:200].astype(float) / 200
  c = cp.abs(cp.sqrt(i * i + j * j) - 0.5) < 0.02
  cf = c.astype(float) * 0.5 + cp.random.uniform(size=c.shape) * 0.5

python/cucim/src/cucim/skimage/feature/tests/test_peak.py

@@ -8,6 +8,7 @@
 from scipy import ndimage as ndimage_cpu
 
 from cucim.skimage._shared._warnings import expected_warnings
+from cucim.skimage._shared.compat import _full
 from cucim.skimage.feature import peak
 
 np.random.seed(21)
@@ -51,7 +52,7 @@ def test_absolute_threshold(self):
  assert_array_almost_equal(peaks, [(3, 3)])
 
  def test_constant_image(self):
- image = cp.full((20, 20), 128, dtype=cp.uint8)
+ image = _full((20, 20), 128, dtype=cp.uint8)
  peaks = peak.peak_local_max(image, min_distance=1)
  assert len(peaks) == 0
 
@@ -491,7 +492,7 @@ def test_threshold_rel_default(self):
  )
 
  def test_peak_at_border(self):
- image = cp.full((10, 10), -2)
+ image = _full((10, 10), -2)
  image[2, 4] = -1
  image[3, 0] = -1
 

python/cucim/src/cucim/skimage/feature/tests/test_template.py

@@ -6,6 +6,7 @@
 from skimage import data
 
 from cucim.skimage import img_as_float
+from cucim.skimage._shared.compat import _full
 from cucim.skimage.feature import match_template, peak_local_max
 from cucim.skimage.morphology import diamond
 
@@ -55,7 +56,7 @@ def test_normalization():
  N = 20
  ipos, jpos = (2, 3)
  ineg, jneg = (12, 11)
- image = cp.full((N, N), 0.5)
+ image = _full((N, N), 0.5)
  image[ipos : ipos + n, jpos : jpos + n] = 1
  image[ineg : ineg + n, jneg : jneg + n] = 0
 

python/cucim/src/cucim/skimage/filters/tests/test_gabor.py

@@ -4,6 +4,7 @@
 from cupy.testing import assert_array_almost_equal
 from numpy.testing import assert_almost_equal
 
+from cucim.skimage._shared.compat import _full
 from cucim.skimage._shared.utils import _supported_float_type
 from cucim.skimage.filters._gabor import _sigma_prefactor, gabor, gabor_kernel
 
@@ -117,6 +118,6 @@ def test_gabor_float_dtype(dtype):
 
 @pytest.mark.parametrize("dtype", [cp.uint8, cp.int32, cp.intp])
 def test_gabor_int_dtype(dtype):
- image = cp.full((16, 16), 128, dtype=dtype)
+ image = _full((16, 16), 128, dtype=dtype)
  y = gabor(image, 0.3)
  assert all(arr.dtype == dtype for arr in y)

python/cucim/src/cucim/skimage/filters/tests/test_thresholding.py

@@ -17,6 +17,7 @@
 from cucim.skimage import util
 from cucim.skimage._shared._dependency_checks import has_mpl
 from cucim.skimage._shared._warnings import expected_warnings
+from cucim.skimage._shared.compat import _full
 from cucim.skimage._shared.utils import _supported_float_type
 from cucim.skimage.color import rgb2gray
 from cucim.skimage.exposure import histogram
@@ -388,7 +389,7 @@ def test_li_astro_image():
 
 
 def test_li_nan_image():
- image = cp.full((5, 5), cp.nan)
+ image = _full((5, 5), cp.nan)
  assert cp.isnan(threshold_li(image))
 
 
@@ -639,7 +640,7 @@ def test_mean():
 def test_triangle_uniform_images(dtype, kwargs):
  assert threshold_triangle(cp.zeros((10, 10), dtype=dtype), **kwargs) == 0
  assert threshold_triangle(cp.ones((10, 10), dtype=dtype), **kwargs) == 1
- assert threshold_triangle(cp.full((10, 10), 2, dtype=dtype), **kwargs) == 2
+ assert threshold_triangle(_full((10, 10), 2, dtype=dtype), **kwargs) == 2
 
 
 # also run cases with nbins > 100000 to also test CuPy-based code path.
@@ -775,7 +776,7 @@ def test_niblack_sauvola_pathological_image():
  # resulted in NaNs. Here we check that these are safely caught.
  # see https://github.com/scikit-image/scikit-image/issues/3007
  value = 0.03082192 + 2.19178082e-09
- src_img = cp.full((4, 4), value).astype(cp.float64)
+ src_img = _full((4, 4), value).astype(cp.float64)
  assert not cp.any(cp.isnan(threshold_niblack(src_img)))
 
 

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

@@ -3,6 +3,7 @@
 import cupy as cp
 import numpy as np
 
+from .._shared.compat import _full
 from .._shared.utils import _supported_float_type, check_nD
 from ._moments_analytical import moments_raw_to_central
 
@@ -420,7 +421,7 @@ def moments_normalized(mu, order=3, spacing=None):
  # compute using in a single kernel for the 2D or 3D cases
  unit_scale = scale == 1.0
  kernel = _get_normalize_kernel(mu.ndim, order, unit_scale)
- nu = cp.full(mu.shape, cp.nan, dtype=mu.dtype)
+ nu = _full(mu.shape, cp.nan, dtype=mu.dtype)
  kernel(mu, order, scale, nu)
  return nu
 

python/cucim/src/cucim/skimage/metrics/_contingency_table.py

@@ -1,6 +1,8 @@
 import cupy as cp
 import cupyx.scipy.sparse as sparse
 
+from .._shared.compat import _full
+
 __all__ = ["contingency_table"]
 
 
@@ -36,7 +38,7 @@ def contingency_table(im_true, im_test, *, ignore_labels=None, normalize=False):
  data /= cp.count_nonzero(data)
  else:
  if normalize:
- data = cp.full((im_test_r.size,), 1 / im_test_r.size, dtype=float)
+ data = _full((im_test_r.size,), 1 / im_test_r.size, dtype=float)
  else:
  data = cp.ones((im_test_r.size,), dtype=float)
  cont = sparse.coo_matrix((data, (im_true_r, im_test_r))).tocsr()

python/cucim/src/cucim/skimage/metrics/tests/test_structural_similarity.py

@@ -16,7 +16,7 @@
 cam_noisy = cam_noisy.astype(cam.dtype)
 
 
-cp.random.seed(1234)
+cp.random.seed(cp.uint32(1234))
 
 assert_equal = cp.testing.assert_array_equal
 assert_almost_equal = cp.testing.assert_array_almost_equal
@@ -25,7 +25,7 @@
 
 def test_structural_similarity_patch_range():
  N = 51
- rstate = cp.random.RandomState(1234)
+ rstate = cp.random.RandomState(cp.uint32(1234))
  X = (rstate.rand(N, N) * 255).astype(cp.uint8)
  Y = (rstate.rand(N, N) * 255).astype(cp.uint8)
 
@@ -35,7 +35,7 @@ def test_structural_similarity_patch_range():
 
 def test_structural_similarity_image():
  N = 100
- rstate = cp.random.RandomState(1234)
+ rstate = cp.random.RandomState(cp.uint32(1234))
  X = (rstate.rand(N, N) * 255).astype(cp.uint8)
  Y = (rstate.rand(N, N) * 255).astype(cp.uint8)
 

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

@@ -159,6 +159,9 @@ def thin(image, max_num_iter=None):
 def _get_tiebreaker(n, seed):
  # CuPy generator doesn't currently have the permutation method, so
  # fall back to cp.random.permutation instead.
+ if np.isscalar(seed):
+ # TODO: remove this NumPy 2.0 compat. fix once CUPy 13.3 is released
+ seed = cp.uint32(seed)
  cp.random.seed(seed)
  if n < 2 << 31:
  dtype = np.int32

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

@@ -14,6 +14,8 @@
 import skimage
 from packaging.version import Version
 
+from .._shared.compat import _full
+
 old_reconstruction_pyx = Version(skimage.__version__) < Version("0.20.0")
 
 
@@ -190,7 +192,7 @@ def reconstruction(seed, mask, method="dilation", footprint=None, offset=None):
  # CuPy Backend: modified to allow images_dtype based on input dtype
  # instead of float64
  images_dtype = np.promote_types(seed.dtype, mask.dtype)
- images = cp.full(dims, pad_value, dtype=images_dtype)
+ images = _full(dims, pad_value, dtype=images_dtype)
  images[(0, *inside_slices)] = seed
  images[(1, *inside_slices)] = mask
  isize = images.size

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

@@ -5,6 +5,7 @@
 from skimage import data
 from skimage.morphology import thin as thin_cpu
 
+from cucim.skimage._shared.compat import _full
 from cucim.skimage.morphology import medial_axis, thin
 
 
@@ -99,7 +100,7 @@ def _test_vertical_line(self, dtype, **kwargs):
  img[:, 3] = 2
  img[:, 4] = 3
 
- expected = cp.full(img.shape, False)
+ expected = _full(img.shape, False)
  expected[:, 3] = True
 
  result = medial_axis(img, **kwargs)

python/cucim/src/cucim/skimage/registration/_optical_flow.py

@@ -10,6 +10,7 @@
 
 from .._shared._gradient import gradient
 from .._shared.utils import _supported_float_type
+from .._vendored.ndimage import uniform_filter # has numpy 2.0 compat fix
 from ..transform import warp
 from ._optical_flow_utils import _coarse_to_fine, _get_warp_points
 
@@ -303,7 +304,7 @@ def _ilk(
  filter_func = partial(gaussian_filter, sigma=sigma, mode="mirror")
  else:
  filter_func = partial(
- ndi.uniform_filter, size=ndim * (size,), mode="mirror"
+ uniform_filter, size=ndim * (size,), mode="mirror"
  )
 
  flow = flow0

python/cucim/src/cucim/skimage/registration/tests/test_tvl1.py

@@ -41,7 +41,7 @@ def _sin_flow_gen(image0, max_motion=4.5, npics=5):
 @pytest.mark.parametrize("dtype", [cp.float16, cp.float32, cp.float64])
 def test_2d_motion(dtype):
  # Generate synthetic data
- rnd = cp.random.RandomState(0)
+ rnd = cp.random.RandomState(cp.uint32(0))
  image0 = cp.array(rnd.normal(size=(256, 256)).astype(dtype))
  gt_flow, image1 = _sin_flow_gen(image0)
  image1 = image1.astype(dtype, copy=False)

python/cucim/src/cucim/skimage/restoration/deconvolution.py

@@ -4,6 +4,7 @@
 import cupy as cp
 import numpy as np
 
+from .._shared.compat import _full
 from .._shared.utils import (
  DEPRECATED,
  _supported_float_type,
@@ -456,7 +457,7 @@ def richardson_lucy(image, psf, num_iter=50, clip=True, filter_epsilon=None):
  float_type = _supported_float_type(image.dtype)
  image = image.astype(float_type, copy=False)
  psf = psf.astype(float_type, copy=False)
- im_deconv = cp.full(image.shape, 0.5, dtype=float_type)
+ im_deconv = _full(image.shape, 0.5, dtype=float_type)
  psf_mirror = cp.ascontiguousarray(psf[::-1, ::-1])
 
  # Small regularization parameter used to avoid 0 divisions

python/cucim/src/cucim/skimage/restoration/tests/test_denoise.py

@@ -10,7 +10,7 @@
 from cucim.skimage._shared.utils import _supported_float_type, slice_at_axis
 from cucim.skimage.metrics import structural_similarity
 
-cp.random.seed(1234)
+cp.random.seed(cp.uint32(1234))
 
 
 astro = img_as_float(data.astronaut()[:128, :128])
@@ -140,7 +140,7 @@ def test_denoise_tv_chambolle_4d():
 def test_denoise_tv_chambolle_weighting():
  # make sure a specified weight gives consistent results regardless of
  # the number of input image dimensions
- rstate = cp.random.RandomState(1234)
+ rstate = cp.random.RandomState(cp.uint32(1234))
  img2d = astro_gray.copy()
  img2d += 0.15 * rstate.standard_normal(img2d.shape)
  img2d = cp.clip(img2d, 0, 1)