Numpy 2.0.0 breaks ismember

[leuraph]

What has changed

From Numpy's 2.0.0 release notes, we know that

When multi-dimensional inputs are passed to np.unique with return_inverse=True, the unique_inverse output is now shaped such that the input can be reconstructed directly using np.take(unique, unique_inverse) when axis=None, and np.take_along_axis(unique, unique_inverse, axis=axis) otherwise.

Breaking Example

In practice, this means that the unit tests of ismember fail and, e.g. the following code breaks when switching from numpy==1.26.0 to numpy==2.0.0:

boundaries = np.array([
    [1,2],
    [3,4],
    [6,7],
    [10,11]
])
local_edges = np.array([
    [6,7],
    [3,4],
    [12,3]
])
is_local_boundary = ismember.is_row_in(
    boundaries, local_edges)
print(is_local_boundary.shape)
local_boundaries = boundaries[is_local_boundary]

IndexError: boolean index did not match indexed array along axis 1; size of axis is 2 but size of corresponding boolean axis is 1

Explanation

• numpy==2.0.0: is_local_boundary has shape (4,1)
• numpy==1.26.0: is_local_boundary has shape (4,)

[leuraph]

@erdogant Do you have any suggestions/ideas on how to handle the situation? In case you do not have time but would appreciate someone looking into it, just let me know and I'll try my best :) For now, I will force usage of numpy<2.0.0 by hand whenever I use ismember.

[erdogant]

thank you for posting this issue. For now, I restricted installation < 2.0.0. I will look into this at a later moment.