ENH: atleast_nd delegation

[Enderdead]

This merge request introduces the atleast_nd delegate function (#100).

Unlike other delegate functions, atleast_nd does not exist for all major backends. Instead, the functions atleast_1d, atleast_2d, and atleast_3d are available.

In this implementation, I use getattr to delegate when ndim is between 1 and 3. I’ve also added tests with ndim=3 to cover the full range.

Open to feedback!

[lucascolley]

looks great, I agree that this delegation strategy makes sense, thanks @Enderdead !

[mdhaber]

@lucascolley @Enderdead git bisect finds that something about this commit (708bce7) causes tests of scipy.spatial.transform.RigidTransform to fail. See e.g. first test run of scipy/scipy#23929.

[lucascolley]

Thanks for the bisection, I'll try to investigate today

[lucascolley]

Some of RigidTransform is relying on atleast_nd adding dimensions to the front of the shape, but np.atleast_3d adds dimensions to the back:

(Pdb) x.shape
(3, 3)
(Pdb) np.atleast_3d(x).shape
(3, 3, 1)
(Pdb) x = np.arange(4)
(Pdb) x.shape
(4,)
(Pdb) np.atleast_2d(x).shape
(1, 4)

[lucascolley]

@Enderdead would you be interested in opening a PR to document that we add dimensions to the front of the shape, and removing the delegation to np.atleast_3d?

[lucascolley]

Apparently according to the docs, np.atleast_3d applied to a 1-D array sandwiches the existing axis between new axes 🤔 https://numpy.org/doc/2.3/reference/generated/numpy.atleast_3d.html#numpy-atleast-3d

[amacati]

That's a weird decision by numpy, didn't know about that. I think the version assumed in RigidTransform still makes a lot of sense.

[Enderdead]

@lucascolley No problem, I’ll take care of fixing this issue. I haven’t looked into it deeply yet, but it’s concerning that the proposed unit test didn’t catch this behavior mismatch. I’ll try to push a fix by the end of the week (check unit test + remove delegation).