BUG: Fix np.testing utils failing for masked scalar vs. scalar (#29317) #29318

danra · 2025-07-04T05:19:34Z

Resolves #29317

Fixes np.testing.assert_XXX utilities failing when comparing a masked scalar against another scalar, with a "nan location mismatch" error. The consistent behavior is to ignore any masked elements, in this case the only one, and therefore succeed.

See also #11121

ngoldbaum · 2025-07-07T15:24:49Z

What is currently broken? What does this PR change?

I think the reason this PR hasn't been reviewed yet is because it has a blank description and does nontrivial things that aren't obvious just from the PR title.

danra · 2025-07-07T18:48:01Z

What is currently broken?

Described in the linked bug: np.testing.assert_XXX utilities fail when comparing a masked scalar against another scalar.

it has a blank description

The title says that it fixes the bug though... I moved a later comment "Resolves #29317" into the description, if that's helpful.

ngoldbaum · 2025-07-07T19:27:09Z

Described in the linked bug: np.testing.assert_XXX utilities fail when comparing a masked scalar against another scalar.

I don't think the np.testing functions are designed to work with masked arrays at all, see e.g.: #14624.

I'm not sure it makes sense to do one-off fixes like this, especially this one that does a cast based on the result of an isinstance call, which feels pretty brittle to me.

Much better IMO to add e.g. a new np.ma.testing module that intentionally accepts masked arrays and handles all these cases, just like we have for all the other masked operations.

Sadly masked arrays are somewhat of an unloved feature no one as far as I'm aware it working on any systematic fixes for them like this.

danra · 2025-07-07T20:33:02Z

Thanks for the reference to #14624.

Personally, I agree it makes more sense to have a separate np.ma.testing module. However, it looks like there have been numerous commits to the current testing module to support masked arrays; since those had been accepted and merged, I took it as a given that that support does need to be kept in case I make any changes. See for example #11121. But, if there's consensus on that direction, then I'd be happy to make such a split to simplify np.testing even though that's more work than I'd planned :)

I've made these one-off changes not for their own sake, but as incremental fixes towards #28827. I'd described that in some more detail under the Context section of #29317:

Ran into this while working on #28827
- while amending assert_array_compare's nan/inf handling to be able to completely remove the additional inf handling in assert_array_almost_equal
- which is needed to avoid assert_array_almost_equal.compare potentially producing an output shape different from its input
- which in turn would avoid additional logic to produce the correct indices of the max abs/rel errors in the original array.

ngoldbaum · 2025-07-07T20:37:08Z

ping @mhvk since you merged a fix a while back along these lines - what do you think?

ngoldbaum · 2025-07-07T20:40:32Z

numpy/testing/_private/utils.py

        # support them if possible.
-        if np.bool(x_id == y_id).all() != True:
+        result = x_id == y_id
+        if isinstance(result, bool):


This feels pretty brittle to me. Is there another way to do this without isinstance?

There are other ways, but have you seen the rationale in the commit message for d1963d9 ?

There are a few ways to resolve this; I went with testing the
comparison result with isinstance(bool) to check if a conversion to
array is necessary, which is the same approach already taken in
assert_array_compare after evaluating comparison(x, y).

The difference here is that logic has been there a long time, so people have written code to handle it. In this case, you're introducing a possibility for NumPy to suddently generate AttributeErrors in downstream tests when NumPy didn't do that before.

Pushed a more robust approach

numpy/testing/_private/utils.py

The two added test cases fail with: E AssertionError: E Arrays are not equal E E nan location mismatch: E ACTUAL: MaskedArray(3.) E DESIRED: array(3.) and E AssertionError: E Arrays are not equal E E nan location mismatch: E ACTUAL: MaskedArray(3.) E DESIRED: array(nan)

…#29317) TestArrayEqual.test_masked_scalar now passes. This case regressed since 7315145 (merged in numpy#12119) due to: - `<masked scalar> == <scalar>` returning np.ma.masked (not a 0-dim masked bool array), followed by - `np.bool(np.ma.masked)` unintentionally converting it to np._False Note on the modified comment: Confusingly, "isinstance(..., bool) checks" in the previous wording actually incorrectly referred to the ones towards the end of the function, which are not actually related to __eq__'s behavior but to the possibility of `func` returning a bool.

- Use same language as elsewhere below to explain `!= True` used to handle np.ma.masked - Clarify committed to support standard MaskedArrays - Restore note lost in 7315145 comment changes about how the np.bool casts towards the end of the function handle np.ma.masked, and expand further.

ngoldbaum

LGTM. Since this is load-bearing code in testing utilities that are used across the ecosystem I'd appreciate it if someone else could also look this over and think about possible issues before merging.

tylerjereddy

I'd appreciate it if someone else could also look this over and think about possible issues before merging.

I didn't really think, but I did build this NumPy feature branch from source locally, then built SciPy main from scratch locally against it, and the full testsuite was fine on ARM Mac. I guess astropy may be a nice one to check as well, I believe you already pinged Marten.

mhvk

The fix looks good - I like the extra care (and it reminds me inventing all that trickery in the first place...).

Just some small comments on the added tests.

Aside: in principle, one could use pytest.mark.parametrize for those, but I think this is fine (and probably faster).

numpy/testing/tests/test_utils.py

mhvk · 2025-07-12T11:09:57Z

p.s. I checked this branch on astropy -- no problems. Also, pretty confident this change is all OK.

mhvk

Approving, since the two comments I had made in my previous review were based on a misunderstanding of what the tests were trying to do. Thanks, @danra!

ngoldbaum · 2025-07-13T18:35:43Z

Thanks for working on this @danra and for being patient with our caution around changing this code. I hope you continue on your efforts to improve the testing utilities.

github-actions bot added the 00 - Bug label Jul 4, 2025

danra mentioned this pull request Jul 4, 2025

MNT: Cleanup infs handling in np.testing assertion utilities #29321

Open

ngoldbaum reviewed Jul 7, 2025

View reviewed changes

numpy/testing/_private/utils.py Show resolved Hide resolved

danra force-pushed the compare_masked_scalar branch from 11cff5c to b8d4da1 Compare July 10, 2025 03:12

danra requested a review from ngoldbaum July 10, 2025 03:14

danra added 3 commits July 9, 2025 20:15

TST: Expand TestArrayEqual.test_masked_scalar

faa1ed8

danra force-pushed the compare_masked_scalar branch from b8d4da1 to faa1ed8 Compare July 10, 2025 03:17

ngoldbaum approved these changes Jul 10, 2025

View reviewed changes

tylerjereddy reviewed Jul 10, 2025

View reviewed changes

mhvk reviewed Jul 12, 2025

View reviewed changes

numpy/testing/tests/test_utils.py Show resolved Hide resolved

numpy/testing/tests/test_utils.py Show resolved Hide resolved

mhvk approved these changes Jul 13, 2025

View reviewed changes

ngoldbaum merged commit 08916dc into numpy:main Jul 13, 2025
76 checks passed

danra deleted the compare_masked_scalar branch July 13, 2025 18:40

Uh oh!

BUG: Fix np.testing utils failing for masked scalar vs. scalar (#29317) #29318

BUG: Fix np.testing utils failing for masked scalar vs. scalar (#29317) #29318

Conversation

danra commented Jul 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ngoldbaum commented Jul 7, 2025

Uh oh!

danra commented Jul 7, 2025

Uh oh!

ngoldbaum commented Jul 7, 2025

Uh oh!

danra commented Jul 7, 2025

Uh oh!

ngoldbaum commented Jul 7, 2025

Uh oh!

ngoldbaum Jul 7, 2025

Choose a reason for hiding this comment

Uh oh!

danra Jul 7, 2025

Choose a reason for hiding this comment

Uh oh!

ngoldbaum Jul 8, 2025

Choose a reason for hiding this comment

Uh oh!

danra Jul 10, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

ngoldbaum left a comment

Choose a reason for hiding this comment

Uh oh!

tylerjereddy left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

mhvk left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

mhvk commented Jul 12, 2025

Uh oh!

mhvk left a comment

Choose a reason for hiding this comment

Uh oh!

ngoldbaum commented Jul 13, 2025

Uh oh!

Uh oh!

Uh oh!

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

danra commented Jul 4, 2025 •

edited

Loading

tylerjereddy left a comment •

edited

Loading