setting warnings to errors changes the behaviour of np.asarray on quantities

[keewis]

Using warnings.filterwarnings to promote warnings to errors changes the behaviour of np.asarray (or rather np.array as np.asarray delegates to that):

>>> array = np.arange(10) * ureg.m
>>> np.asarray(array)  # warning emitted because units were stripped
[0 1 2 3 4 5 6 7 8 9]
>>> warnings.filterwarnings("error")
>>> np.asarray(array)  # warning should have become an error, but completely disappeared
array([<Quantity(0, 'meter')>, <Quantity(1, 'meter')>,
       <Quantity(2, 'meter')>, <Quantity(3, 'meter')>,
       <Quantity(4, 'meter')>, <Quantity(5, 'meter')>,
       <Quantity(6, 'meter')>, <Quantity(7, 'meter')>,
       <Quantity(8, 'meter')>, <Quantity(9, 'meter')>], dtype=object)

This happens with pint 0.9, but also with the version from #764 (because __array_function__ does not allow overloading np.asarray)

Edit: I do think this is an issue within numpy, but I think it might help to also track this here. Possibly related: numpy/numpy#9761

[eric-wieser]

In some sense this is a feature - numpy was told it was not allowed to convert to an int in the second case, so it picked an object array instead.

What behavior would you like here instead?

[keewis]

The use case here is catching warnings in unit tests. Obviously, if the warning -- that was converted to an error -- gets swallowed, the test won't work.

[keewis]

what I would like is that warnings get passed through. So something like:

try:
    ...
except Warning:
    raise
except Exception:
    ...

or catching using AttributeError

[jthielen]

Based on discussion in numpy/numpy#14735, this seems to be a pint issue with using __getattr__ to handle __array_* attributes/methods, in particular, __array__. With this in mind, I think that pint should explicitly set the __array_* attributes/methods it relies on, or at the very least, explicitly define __array__ (see also #878 (comment))?

[keewis]

with the merge of numpy/numpy#14745 this won't happen on numpy master and from 1.18 on. We still need to decide if we want to change pint, though, before we can close this

[jthielen]

This relates to some other issues with Pint's handling of __array_* attributes, so I've created #924 to hopefully discuss this and the other issues in general.