Handle new warnings introduced in NumPy 1.24

[grlee77]

The pre-release test case on CI fails due to a few new warnings introduced in NumPy 1.24 (example of a recent test log). (Our pre-release test CI job treats warnings as errors via env var SKIMAGE_TEST_STRICT_WARNINGS_GLOBAL=1)

The changes here should resolve these failures.

[grlee77]

The fix for Cascade in f25ae85 does not seem like the correct solution as it will break again with an error once the deprecation cycle completes. I am not familiar with that code so would like to leave a proper fix to a separate PR. I will open an issue for it. The reported warning from the log is:

FAILED feature/tests/test_cascade.py::test_detector_astronaut - DeprecationWarning: NumPy will stop allowing conversion of out-of-bound Python integers to integer arrays.  The conversion of -67130709 to uint32 will fail in the future.
For the old behavior, usually:
    np.array(value).astype(dtype)`
will give the desired result (the cast overflows).

[grlee77]

Three failures remaining in test_ciede2000_dE that are a bit more concerning as at seems the tolerance to pass the test has changed.

FAILED color/tests/test_delta_e.py::test_ciede2000_dE[float64-0] - AssertionError: 
Not equal to tolerance rtol=0.0001, atol=0

Mismatched elements: 1 / 34 (2.94%)
Max absolute difference: 0.04027009
Max relative difference: 0.00560927
 x: array([ 2.04246 ,  2.86151 ,  3.441191,  0.999999,  1.000005,  1.000013,
        2.366859,  2.366859,  7.179172,  7.21947 ,  7.219472,  7.219474,
        4.804522,  4.804525,  4.746071,  4.306482, 27.149231, 22.897692,...
 y: array([ 2.0425,  2.8615,  3.4412,  1.    ,  1.    ,  1.    ,  2.3669,
        2.3669,  7.1792,  7.1792,  7.2195,  7.2195,  4.8045,  4.8045,
        4.7461,  4.3065, 27.1492, 22.8977, 31.903 , 19.4535,  1.    ,...
FAILED color/tests/test_delta_e.py::test_ciede2000_dE[float64-1] - AssertionError: 
Not equal to tolerance rtol=0.0001, atol=0

Mismatched elements: 1 / 34 (2.94%)
Max absolute difference: 0.04027009
Max relative difference: 0.00560927
 x: array([ 2.04246 ,  2.86151 ,  3.441191,  0.999999,  1.000005,  1.000013,
        2.366859,  2.366859,  7.179172,  7.21947 ,  7.219472,  7.219474,
        4.804522,  4.804525,  4.746071,  4.306482, 27.149231, 22.897692,...
 y: array([ 2.0425,  2.8615,  3.4412,  1.    ,  1.    ,  1.    ,  2.3669,
        2.3669,  7.1792,  7.1792,  7.2195,  7.2195,  4.8045,  4.8045,
        4.7461,  4.3065, 27.1492, 22.8977, 31.903 , 19.4535,  1.    ,...
FAILED color/tests/test_delta_e.py::test_ciede2000_dE[float64--1] - AssertionError: 
Not equal to tolerance rtol=0.0001, atol=0

Mismatched elements: 1 / 34 (2.94%)
Max absolute difference: 0.04027009
Max relative difference: 0.00560927
 x: array([ 2.04246 ,  2.86151 ,  3.441191,  0.999999,  1.000005,  1.000013,
        2.366859,  2.366859,  7.179172,  7.21947 ,  7.219472,  7.219474,
        4.804522,  4.804525,  4.746071,  4.306482, 27.149231, 22.897692,...
 y: array([ 2.0425,  2.8615,  3.4412,  1.    ,  1.    ,  1.    ,  2.3669,
        2.3669,  7.1792,  7.1792,  7.2195,  7.2195,  4.8045,  4.8045,
        4.7461,  4.3065, 27.1492, 22.8977, 31.903 , 19.4535,  1.    ,...
====== 3 failed, 8306 passed, 30 skipped, 5 warnings in 340.68s (0:05:40) ======
Error: Process completed with exit code 1.

[grlee77]

Regarding the relatively low rtolused in test_ciede2000_dE, I think it is likely there is substantial numerical error in the term:
np.sqrt(c7 / (c7 + 25 ** 7)) where c7 might have very small values while 25**7 is large, so there may be substantial floating point error. I don't know why the situation would have changed for the double precision test case only in NumPy 1.24rc1 vs. older releases, though.

updated the comment as I am not sure what term is the primary contributor to the numerical error and don't think it is the one I originally mentioned

[stefanv]

Should np.bool_ just be bool?

[grlee77]

I think they are all equivalent. They do have different entries here, though:

scikit-image/skimage/util/dtype.py

Lines 25 to 27 in f11c27f

	dtype_range = {bool: (False, True),
	np.bool_: (False, True),
	np.bool8: (False, True),

so I think that is why there was both bool and np.bool8 separately here previously. I just changed bool8 to bool_ since that one is not deprecated.

[grlee77]

@bmorris3, I pulled your Cascade fix from #6633 to here in 4b21720. Thanks!