np.testing.assert_equal does not work with quantity lists

[edmundsj]

When creating a list of quantities from a numpy array, numpy's built-in test methods for array comparison no longer work. This is breaking a lot of my test suite now that I am trying to incorporate pint's unit capabilities into my packages. Is there an easy way around this, or am I just going to have to write a wrapper function and swap it out for numpy's?

import pint
import numpy as np


ureg = pint.UnitRegistry()
data = ureg.mV * np.array([1, 2, 3])
np.testing.assert_equal(data, data) # This fails

[andrewgsavage]

There's some helper functions in the testsuite

from pint.testsuite import helpers
helpers.assert_quantity_equal(x, y)

[hgrecco]

I was also starting to look into this. Any idea what triggered this?

[edmundsj]

It looks like numpy fails to recognize the object as array-like, and so doesn't invoke all() appropriately. I see there has been discussion of this for a while (#1128). The error I get is below:

if not (desired == actual):
ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

[hgrecco]

Yes, but you mentioned that is working before. That is what surprised me.

[keewis]

looking at the code of assert_equal it seems that there is a special case for ndarray (using isinstance), for which assert_array_equal is used. If we directly use numpy.testing.assert_array_equal that casts to numpy using asarray.

I think for testing it would be good to expose pint.testsuite.helpers.assert_quantity_equal as pint.testing.assert_equal.

[edmundsj]

Yes, but you mentioned that is working before. That is what surprised me.

Ah! Apologies, not working with pint, but working with regular arrays of dimensionless numbers.

[hgrecco]

Ahh, ok. Then as mentioned by @andrewgsavage and @keewis mentions we should expose and document assert_quantity_equal (and assert_quantity_almost_equal) which work like the corresponding numpy counterparts but deal with unit conversions.

[keewis]

following numpy.testing, assert_quantity_almost_equal should be exposed as pint.testing.assert_allclose

[hgrecco]

True (see https://github.com/numpy/numpy/blob/v1.20.0/numpy/testing/_private/utils.py#L936 )

My recomendation is tha we double check the implementation of assert_quantity_almost_equal so tha it follows the semantics of assert_allclose.
If yes, then we expose assert_quantity_almost_equal as assert_allclose
If no, then we define assert_allclose to match the semantics of numpy.