Use np.testing when appropriate in unit tests about iodata HOT 5 CLOSED

@tovrstra It is safe to say that we use numpy.testing.assert_array_almost_equal for almost all cases and numpy.testing.assert_allclose in rare cases when precision is required?

from iodata.

The Numpy docs recommend using assert_allclose instead of assert_almost_equal. See https://docs.scipy.org/doc/numpy-1.15.0/reference/generated/numpy.testing.assert_almost_equal.html

from iodata.

Or maybe use the Python (not numpy) intrinsic:
math.isclose(a, b, *, rel_tol=1e-09, abs_tol=0.0)

Probably clunkier without the "assert" built in....

from iodata.

The numpy.testing.assert_allclose is asymmetric, which means that a=b might result b not equal a. For example (taken from https://stackoverflow.com/questions/48156460/isclose-function-in-numpy-is-different-from-math)

a = 0.142253
b = 0.142219
rel_tol = 1e-4
abs_tol = 1.9776e-05
# test failed
numpy.testing.assert_allclose(actual=a, desired=b, rtol=rel_tol, atol=abs_tol)
# test passed
np.testing.assert_allclose(actual=b, desired=a, rtol=rel_tol, atol=abs_tol)

But for math.isclose(a, b, *, rel_tol=1e-09, abs_tol=0.0) , it seems only support scaler instead of numpy array.

a = np.linspace(0., 1., 100).reshape(10, 10)
b = np.arcsin(np.sin(a))
# this will raise an TypeError
math.isclose(a, b, rel_tol=1e-09, abs_tol=0.0)

Once an agreement is reached, I can start working on this.

from iodata.

It sounds to me like you should use numpy.testing.assert_allclose

The nuance of directionality is unlikely to be important; the example is a rather evil one. To the extent that you can identify one value which is "reference" and one which is "approximate" you should make sure the "desired" is the reference.

from iodata.