Sorting a three-dimensional numpy array by a two-dimensional index

Question

I'm having real trouble with this. I have a three-dimensional numpy array, and I'd like to re-order it by a two-dimensional indexing array. In reality the arrays will be determined programatically and the three-dimensional array may be two or four-dimensioned, but to keep things simple, here's the desired outcome if both arrays were two-dimensional:

ph = np.array([[1,2,3], [3,2,1]])
ph_idx = np.array([[0,1,2], [2,1,0]])
for sub_dim_n, sub_dim_ph_idx in enumerate(ph_idx):
    ph[sub_dim_n] = ph[sub_dim_n][sub_dim_ph_idx]

This makes the ph array into:

array([[1, 2, 3],
       [1, 2, 3]])

Which is what I'd like. Could anyone please help if it's the same circumstance, but instead of ph, I have a three-dimensional array (psh), like:

psh = np.array(
    [[[1,2,3]], 
     [[3,2,1]]]
)

Hope that's clear and please ask if it's not. Thanks in advance!

Answer 1

If what you want is to end up with a ph.shape shaped array, you could simply np.squeeze ph_ixs so the shapes match, and use it to index ph :

print(ph)
[[[1 2 3]]
 [[3 2 1]]]

print(ph_idx)
[[0 1 2]
 [2 1 0]]

np.take_along_axis(np.squeeze(ph), ph_idx, axis=-1)

array([[1, 2, 3],
       [1, 2, 3]])

Answer 2

So, the clues are already in the helpful comments here, but for completeness, it's as simple as using np.take_along_axis and a broadcasted version of the 2d array:

psh = np.array(
    [[[1,2,3]], 
     [[3,2,1]]]
)
ph_idx = np.array(
    [[0,1,2], 
     [2,1,0]]
)
np.take_along_axis(psh, ph_idx[:, None, :], axis=2)

This has the advantage of also working if the 3d array has a dim1 of more than one element:

psh = np.array(
    [[[1,2,3],[4,5,6],[7,8,9]], 
     [[3,2,1],[6,5,4],[9,8,7]]]
)
ph_idx = np.array([[0,1,2], [2,1,0]])
np.take_along_axis(psh, ph_idx[:, None, :], axis=2)

which gives

array([[[1, 2, 3],
        [4, 5, 6],
        [7, 8, 9]],

       [[1, 2, 3],
        [4, 5, 6],
        [7, 8, 9]]])