some tests fail on aarch64 dawin

[natsukium]

Description:

It may be due to the initial values in the array, but when I run the test on aarch64-darwin, the following two tests fail.

measure/tests/test_fit.py

_____________________________________________________________________________ test_ellipse_parameter_stability _____________________________________________________________________________

    def test_ellipse_parameter_stability():
        """The fit should be modified so that a > b
        """
    
        for angle in np.arange(0, 180 + 1, 1):
            # generate rotation matrix
            theta = np.deg2rad(angle)
            c = np.cos(theta)
            s = np.sin(theta)
            R = np.array([
                [c, -s],
                [s, c]]
            )
    
            # generate points on ellipse
            t = np.linspace(0, 2 * np.pi, 20)
            a = 100
            b = 50
            points = np.array([a * np.cos(t), b * np.sin(t)])
            points = R @ points
    
            # fit model to points
            ellipse_model = EllipseModel()
            ellipse_model.estimate(points.T)
            _, _, a_prime, b_prime, theta_prime = ellipse_model.params
    
>           assert_almost_equal(theta_prime, theta)

.venv/lib/python3.10/site-packages/skimage/measure/tests/test_fit.py:247: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

args = (2.356194490192345, 0.7853981633974483), kwds = {}

    @wraps(func)
    def inner(*args, **kwds):
        with self._recreate_cm():
>           return func(*args, **kwds)
E           AssertionError: 
E           Arrays are not almost equal to 7 decimals
E            ACTUAL: 2.356194490192345
E            DESIRED: 0.7853981633974483

/nix/store/6xmv46hawz0iqgvzwllmagzqc9pws00d-python3-3.10.12/lib/python3.10/contextlib.py:79: AssertionError

measure/tests/test_moments.py

_____________________________________________________________________ test_analytical_moments_calculation[3-1-float32] _____________________________________________________________________

dtype = <class 'numpy.float32'>, order = 1, ndim = 3

    @pytest.mark.parametrize('dtype', [np.uint8, np.int32, np.float32, np.float64])
    @pytest.mark.parametrize('order', [1, 2, 3, 4])
    @pytest.mark.parametrize('ndim', [2, 3, 4])
    def test_analytical_moments_calculation(dtype, order, ndim):
        if ndim == 2:
            shape = (256, 256)
        elif ndim == 3:
            shape = (64, 64, 64)
        else:
            shape = (16, ) * ndim
        rng = np.random.default_rng(1234)
        if np.dtype(dtype).kind in 'iu':
            x = rng.integers(0, 256, shape, dtype=dtype)
        else:
            x = rng.standard_normal(shape, dtype=dtype)
        # setting center=None will use the analytical expressions
        m1 = moments_central(x, center=None, order=order)
        # providing explicit centroid will bypass the analytical code path
        m2 = moments_central(x, center=centroid(x), order=order)
        # ensure numeric and analytical central moments are close
        thresh = 1e-4 if x.dtype == np.float32 else 1e-9
>       compare_moments(m1, m2, thresh=thresh)

.venv/lib/python3.10/site-packages/skimage/measure/tests/test_moments.py:237: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

m1 = array([[[140.60745,   0.     ],
        [  0.     ,   0.     ]],

       [[  0.     ,   0.     ],
        [  0.     ,   0.     ]]], dtype=float32)
m2 = array([[[ 1.4060745e+02, -9.5214844e-03],
        [ 1.5533447e-02, -1.1226189e+06]],

       [[ 5.2490234e-03, -1.1670678e+06],
        [-8.2869125e+05,  1.7116478e+08]]], dtype=float32)
thresh = 0.0001

    def compare_moments(m1, m2, thresh=1e-8):
        """Compare two moments arrays.
    
        Compares only values in the upper-left triangle of m1, m2 since
        values below the diagonal exceed the specified order and are not computed
        when the analytical computation is used.
    
        Also, there the first-order central moments will be exactly zero with the
        analytical calculation, but will not be zero due to limited floating point
        precision when using a numerical computation. Here we just specify the
        tolerance as a fraction of the maximum absolute value in the moments array.
        """
        m1 = m1.copy()
        m2 = m2.copy()
    
        # make sure location of any NaN values match and then ignore the NaN values
        # in the subsequent comparisons
        nan_idx1 = np.where(np.isnan(m1.ravel()))[0]
        nan_idx2 = np.where(np.isnan(m2.ravel()))[0]
        assert len(nan_idx1) == len(nan_idx2)
        assert np.all(nan_idx1 == nan_idx2)
        m1[np.isnan(m1)] = 0
        m2[np.isnan(m2)] = 0
    
        max_val = np.abs(m1[m1 != 0]).max()
        for orders in itertools.product(*((range(m1.shape[0]),) * m1.ndim)):
            if sum(orders) > m1.shape[0] - 1:
                m1[orders] = 0
                m2[orders] = 0
                continue
            abs_diff = abs(m1[orders] - m2[orders])
            rel_diff = abs_diff / max_val
>           assert rel_diff < thresh
E           assert 0.000110473855 < 0.0001

.venv/lib/python3.10/site-packages/skimage/measure/tests/test_moments.py:50: AssertionError

Way to reproduce:

on Aarch64 darwin

pytest skimage/measure/tests/test_moments.py skimage/measure/tests/test_fit.py

Version information:

>>> import sys; print(sys.version)
3.10.12 (main, Jul 29 2023, 02:25:35) [Clang 11.1.0 ]
>>> import platform; print(platform.platform())
macOS-13.3.1-arm64-arm-64bit
>>> import skimage; print(f'scikit-image version: {skimage.__version__}')
scikit-image version: 0.21.0
>>> import numpy; print(f'numpy version: {numpy.__version__}')

[hmaarrfk]

I think it is worth mentioning that for test_fit

np.pi - 0.7853981633974483 = 2.356194490192345

For test_moments it seems like the precision might have to be relaxed.

[lagru]

So is it an orientation issue? Seems #2646 related but might also conflate a few different problems.

[risicle]

On darwin x86_64, I see this failure too, though as of 0.22.0 we're only seeing it for test_ellipse_parameter_stability.

[lagru]

This test seems to be problematic / unstable to me judging from other similar issues. I hope that we can maybe address the test or code in question and make it more stable. See #7348.