assert_allclose(..., equal_nan=False) doesn't raise an assertion error

[akaihola]

np.testing.assert_allclose is supposed to raise an AssertionError when comparing two NaN values with equal_nan=False. However, it succeeds without exceptions for both scalars and vectors.

np.allclose(..., equal_nan=False) does work as documented as seen below, but for assert_allclose the behavior is inconsistent with both the documentation and allclose.

>>> import numpy as np
>>> a = np.array([42.0, np.nan])
>>> b = np.array([42.0, np.nan])
>>> np.allclose(np.nan, np.nan, equal_nan=True)
True
>>> np.allclose(a, b, equal_nan=True)
True
>>> np.testing.assert_allclose(np.nan, np.nan, equal_nan=True)
>>> np.testing.assert_allclose(a, b, equal_nan=True)
>>> np.allclose(np.nan, np.nan, equal_nan=False)
False
>>> np.allclose(a, b, equal_nan=False)
False
>>> # These two should raise an AssertionError:
>>> np.testing.assert_allclose(np.nan, np.nan, equal_nan=False)
>>> np.testing.assert_allclose(a, b, equal_nan=False)

I tested this with master branch commit b0a6e55 from today (2016-10-12). I'm running Python 3.5.1 on Fedora 23.

[akaihola]

Both allclose() and assert_allclose() use numpy.core.numeric.isclose(), but there's a subtle difference:

• allclose() aggregates the results with Python's all()
• assert_allclose() uses assert_array_compare() and passes a curried isclose() as the comparison function

assert_array_compare() has a hard-coded behavior of always comparing NaNs and infinite values separately, and only applying the comparison function to numeric values. As a result, the equal_nan= argument doesn't affect NaN handling at all in assert_allclose().

My proposed fix keeps the behavior of assert_array_compare() intact, unless the newly introduced equal_nan= argument is set to False. In that case it runs both NaNs and numeric values through the comparison function.

Update: Corrected the empty_nan typo to equal_nan. I also typed that in one of the commit messages.

[charris]

Where is empty_nan?

[akaihola]

@charris, sorry empty_nan was just a typo in the commit message and the comment above. Should be equal_nan.

[charris]

Closed by #8165.