Replace usage of np.allclose in testing with assert_allclose

[davidorme]

I was expecting these two commands to be equivalent:

In [3]: import numpy as np
   ...: from numpy.testing import assert_allclose
   ...: a = np.array([[2,2]])
   ...: b = np.array([[2,2], [2,2]])
   ...: a.shape == b.shape
   ...: False
   ...: assert np.allclose(a, b)  # DOES NOT RAISE!
   ...: assert_allclose(a, b)
   ...: 
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
...

(shapes (1, 2), (2, 2) mismatch)
 ACTUAL: array([[2, 2]])
 DESIRED: array([[2, 2],
       [2, 2]])

Turns out np.allclose broadcasts the inputs before testing, so this does not guarantee the inputs are the correct shape. This is a good reason to prefer np.testing.assert_allclose(a, b) over assert np.allclose(a, b). I can sort of see why it does it - the broadcasting is such a fundamental part of numpy that it happens first? - but it is counterintuitive.

I've used np.allclose a lot in the tests. I doubt we've got any issues, but it's tempting fate.

[davidorme]

I doubt we've got any issues, but it's tempting fate.

It was tempting fate.