Fix canny and butterworth (recent CuPy and NumPy compatibility) (#574)

It is possible in some cases for array scalar `low_threshold` to cause a failure to launch the non-maximum suppression kernel. This MR casts the array scalar to a float to avoid the issue.

I encountered this as a test failure locally with the existing `test_use_quantiles` test case for `canny` . I am unsure what underlying change caused this to suddenly start failing (possibly an underlying change in CuPy's kernel launching behavior).

A second issue with NumPy >= 1.25 compatibility for `butterworth` is also fixed in this MR. The `cupyx.scipy.fft` call raises an error in comparing an array fft_shape to a tuple shape internally. Casting `fft_shape` to a tuple avoids this.

Authors:
  - Gregory Lee (https://github.com/grlee77)

Approvers:
  - https://github.com/jakirkham

URL: #574

python/cucim/src/cucim/skimage/feature/_canny.py

@@ -235,6 +235,13 @@ def _nonmaximum_suppression_bilinear(
     kernel = _get_nonmax_kernel(large_int=large_int)
 
     out = cp.empty_like(magnitude)
+
+    if isinstance(low_threshold, cp.ndarray):
+        # if array scalar was provided, make sure dtype matches other arrays
+        if low_threshold.ndim > 0:
+            raise ValueError("expected scalar low_treshold")
+        low_threshold = float(low_threshold)
+
     kernel(isobel, jsobel, magnitude, eroded_mask, low_threshold, out)
     return out
 

python/cucim/src/cucim/skimage/filters/_fft_based.py

@@ -159,8 +159,8 @@ def butterworth(
     elif npad > 0:
         center_slice = tuple(slice(npad, s + npad) for s in image.shape)
         image = pad(image, npad, mode='edge')
-    fft_shape = (image.shape if channel_axis is None
-                 else np.delete(image.shape, channel_axis))
+    fft_shape = tuple(image.shape if channel_axis is None
+                      else np.delete(image.shape, channel_axis))
     is_real = cp.isrealobj(image)
     float_dtype = _supported_float_type(image.dtype, allow_complex=True)
     image = image.astype(float_dtype, copy=False)