BUG: Fix np.unique with axis=0 and 1D input not collapsing NaNs with equal_nan=True #29336 #29372
Conversation
corrected white space
Review comment implementation
AnkitAhlawat7742
force-pushed
the
AnkitAhlawat_fix-unique-equal_nan-axis0-1d
branch
from
63104ca to
7f04ede
Compare
|
Hi @ngoldbaum , |
|
Haven't looked at this, but one note/question is how it ties back to: #27345 (and maybe there is even something relevant for this there). |
@seberg , |
|
@seberg are we going to merge this PR ? ,as it is a Quick fix for 1st case(For 1D arrays, np.nan values are not treated as equal even when equal_nan=True.) |
| @@ -290,7 +290,7 @@ def unique(ar, return_index=False, return_inverse=False, | |||
| """ | |||
| ar = np.asanyarray(ar) | |||
| if axis is None: | |||
| if axis is None or (axis == 0 and ar.ndim == 1): | |||
There was a problem hiding this comment.
are we going to merge this PR
I suppose it makes sense, yes. But this isn't quite right and you should add a test for all cases that make sense: that includes -1.
It also will allow axis == 0.0 here which we should reject, maybe just check ar.ndim == 1 first and then call normalize_axis_index to be definitely right.
(And since that is clearly easy to miss, maybe add a test for that too.)
EDIT: Actually, this just gets us back to what @ngoldbaum said: Use axis is None or ar.ndim == 1, then just add a normalize_axis_index() -- but you can ignore the result.
What was the Issue
Issue link :- #29336
Bug Description:- The equal_nan parameter in numpy.unique does not work correctly in some cases. Specifically:
For 1D arrays, np.nan values are not treated as equal even when equal_nan=True.
For 2D arrays with axis=0, the issue is similar — np.nan values are not treated as equal.
What This PR Fixes
This pull request fixes the issue for 1D arrays only.
Fixed: np.unique(..., equal_nan=True) now treats NaNs as equal in 1D arrays.
Notes
This PR does not address the 2D case (e.g., np.unique(array, axis=0, equal_nan=True)) described in the same issue.In coming days i will investigate and provide fix for this case as well