ENH: Add strict parameter to assert_array_equal. Fixes #9542

[jontwo]

It's useful to have assert_array_equal broadcast a scalar to match every value in an array, but it would be nice to be able to turn this behaviour off. Sometimes you need to have an exact match.

[jontwo]

Not sure why those tests are failing - it's picked up the change to the tests but not to numpy.testing.

[seberg]

No, it is picking up things just fine the problem is that TestArrayEqual gets inherited below for TestEqual. (May look weird these days, but this code is older than pytest and other test fixtures.)

[jontwo]

Ok, could I add strict further up the chain and pass it though?

[jontwo]

Not sure what to do about these test fails. Can we rerun the pipeline when there is a network issue?

[charris]

Needs a release note in doc/release/upcoming_changes.

[WarrenWeckesser]

I've often wanted a strict version of the array equality test functions. I would like it to be even stricter than the change proposed here, and only pass if the shape and data types match exactly. So if x is np.array([2.0, 2.0, 2.0], dtype=np.float32), both the checks assert_array_equal(x, 2, strict=True) and assert_array_equal(x, [2, 2, 2], strict=True) would fail. The strict version of the test would be assert_array_equal(x, np.array([2, 2, 2], dtype=np.float32), strict=True).

[jontwo]

I've often wanted a strict version of the array equality test functions. I would like it to be even stricter than the change proposed here, and only pass if the shape and data types match exactly. So if x is np.array([2.0, 2.0, 2.0], dtype=np.float32), both the checks assert_array_equal(x, 2, strict=True) and assert_array_equal(x, [2, 2, 2], strict=True) would fail. The strict version of the test would be assert_array_equal(x, np.array([2, 2, 2], dtype=np.float32), strict=True).

This sounds like a good idea to me. It makes sense that a "strict" check would also enforce the types.

I was also wondering if we should add the same parameter to assert_array_almost_equal in case users are being tripped up by the same problem. What does everyone think?

[BvB93]

Can you also add the new function parameters to the utils.pyi stub file?

[jontwo]

Needs a release note in doc/release/upcoming_changes.

Does it count as a new_feature? The README says this is for new features "like kwargs". Or is it better as improvement or change?

[mattip]

The mailing list discussion is here. I think this is the best available solution.

[seberg]

I concur with Matti. I did not remember that it also makes the dtype strict.
This seems like a good idea and I suspect it makes using strict=True useful in practice.

[WarrenWeckesser]

I made a few nitpicky inline comments. The more significant (but still not major) issue that should be fixed is that when the shapes of two arrays match but the dtypes do not, the error message is not correct. E.g.

In [19]: assert_array_equal(np.eye(2), np.eye(2, dtype=np.float32), strict=True)
---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<...snip...> 
AssertionError: 
Arrays are not equal

(shapes (2, 2), (2, 2) mismatch)
 x: array([[1., 0.],
       [0., 1.]])
 y: array([[1., 0.],
       [0., 1.]], dtype=float32)

The error message says (shapes (2, 2), (2, 2) mismatch) but the shapes do match. In this case, the error message should report that the dtypes do not match.

[WarrenWeckesser]

Thanks @jontwo, the updates look good. I'll merge in a couple days if no one does it before then.

[WarrenWeckesser]

Merged. Thanks @jontwo and @BvB93.