RFC: zero-dimensional `indices` array for `take`

[mdhaber]

The standard specifies that the indices array of take must be one-dimensional.

NumPy, PyTorch, JAX, and CuPy all support zero-dimensional indices arrays, too. Of course, all libraries support zero-dimensional (and int) indices for __getitem__.

Should the standard allow zero-dimensional arrays for indices? How about ints?

[kgryte]

This was discussed previously. See #855.

[ev-br]

One sticky bit is that the spec requires that

The output array must have the same number of axes as x

While major array libraries do not conform for 0D indices:

In [1]: import jax.numpy as jnp

In [2]: import cupy as cp
im
In [3]: import numpy as np

In [4]: for xp in [np, cp, jnp]:
   ...:     x = xp.arange(6).reshape(2, 3)
   ...:     taken = xp.take(x, xp.asarray(0), axis=0)
   ...:     print(xp.__name__, x.ndim, taken.ndim)
   ...: 
numpy 2 1
cupy 2 1
jax.numpy 2 1

[mdhaber]

Oops, sorry I missed that issue.
This is orthogonal: that is only about a 0-d first argument (x), and this is about a 0-d second (indices) argument.
And @ev-br's point is valid, too.