BUG/COMPAT: fix assert_* functions for nested arrays with latest numpy #50396
Conversation
jorisvandenbossche
added
Testing
| # reached in groupby aggregations, make sure we use np.any when checking | ||
| # if the comparison is truthy | ||
| left = np.array([np.array([50, 70, 90]), np.array([20, 30, 40])], dtype=object) |
There was a problem hiding this comment.
should we have both same-sized and mismatched-size cases? i expect these will be non-equivalent through different paths
There was a problem hiding this comment.
I intentionally changed that here, because I assume the intent of the test was to test a numpy array of arrays. But if the arrays are the same length, the np.array(..) constructor actually converts this into a 2D array.
Below in another test, I added a case for same length (and then constructed the array with a workaround first creating an empty and then filling), see # same-length lists.
There was a problem hiding this comment.
Added a same-length subarrays case in this test as well.
…assert-nested-data-numpy-125
|
Any other comments here? |
| assert not array_equivalent(left, right[::-1], strict_nan=strict_nan) | ||
|
|
||
|
|
||
| @pytest.mark.xfail(reason="failing") |
There was a problem hiding this comment.
Just to clarify my understanding, these xfails are not dependent on a future numpy version correct?
There was a problem hiding this comment.
I think not, but not 100% sure by heart. In any case, there were a bunch of cases that also with current numpy already failed with a direct array_equivalent while passing with assert_almost_equal because those two take slightly different code paths.
|
Thanks @jbrockmendel |
…nested arrays with latest numpy
This adds tests for
assert_almost_equal(used byassert_series_equalet al in case of object dtype arrays) that all pass on pandas main with released numpy (except for two cases with dict of array), and fixes the implementation ofarray_equivalent(used byassert_almost_equal) to return False instead of raising an error if numpy cannot compare the arrays (numpy 1.25.dev starts to raise for this instead of returning False).I also added a bunch of equivalent tests for
array_equivalentitself, but those don't all pass (and added xfails). Trying to fix those as well might need more extensive changes, and I would prefer to do that separate from this PR (to keep this one possible to backport).See #50360 (comment) for an illustration of the behaviour that changed in numpy 1.25.dev that causes this.
doc/source/whatsnew/vX.X.X.rstfile if fixing a bug or adding a new feature.