Numpy match indexing dimensions

Question

Problem

I have two numpy arrays, A and indices.

A has dimensions m x n x 10000. indices has dimensions m x n x 5 (output from argpartition(A, 5)[:,:,:5]). I would like to get a m x n x 5 array containing the elements of A corresponding to indices.

Attempts

indices = np.array([[[5,4,3,2,1],[1,1,1,1,1],[1,1,1,1,1]],
    [500,400,300,200,100],[100,100,100,100,100],[100,100,100,100,100]])
A = np.reshape(range(2 * 3 * 10000), (2,3,10000))

A[...,indices] # gives an array of size (2,3,2,3,5). I want a subset of these values
np.take(A, indices) # shape is right, but it flattens the array first
np.choose(indices, A) # fails because of shape mismatch. 

Motivation

I'm trying to get the 5 largest values of A[i,j] for each i<m, j<n in sorted order using np.argpartition because the arrays can get fairly large.

Answer 1

You can use advanced-indexing -

m,n = A.shape[:2]
out = A[np.arange(m)[:,None,None],np.arange(n)[:,None],indices]

Sample run -

In [330]: A
Out[330]: 
array([[[38, 21, 61, 74, 35, 29, 44, 46, 43, 38],
        [22, 44, 89, 48, 97, 75, 50, 16, 28, 78],
        [72, 90, 48, 88, 64, 30, 62, 89, 46, 20]],

       [[81, 57, 18, 71, 43, 40, 57, 14, 89, 15],
        [93, 47, 17, 24, 22, 87, 34, 29, 66, 20],
        [95, 27, 76, 85, 52, 89, 69, 92, 14, 13]]])

In [331]: indices
Out[331]: 
array([[[7, 8, 1],
        [7, 4, 7],
        [4, 8, 4]],

       [[0, 7, 4],
        [5, 3, 1],
        [1, 4, 0]]])

In [332]: m,n = A.shape[:2]

In [333]: A[np.arange(m)[:,None,None],np.arange(n)[:,None],indices]
Out[333]: 
array([[[46, 43, 21],
        [16, 97, 16],
        [64, 46, 64]],

       [[81, 14, 43],
        [87, 24, 47],
        [27, 52, 95]]])

For getting those indices corresponding to the max 5 elements along the last axis, we would use argpartition, like so -

indices = np.argpartition(-A,5,axis=-1)[...,:5]

To keep the order from highest to lowest, use range(5) instead of 5.

Answer 2

For posterity, the following uses Divakar's answer to accomplish the original goal, i.e. return the top 5 values for all i<m, j<n in sorted order:

m, n = np.shape(A)[:2]

# get the largest 5 indices for all m, n
top_unsorted_indices = np.argpartition(A, -5, axis=2)[...,-5:]

# get the values corresponding to top_unsorted_indices
top_values = A[np.arange(m)[:,None,None], np.arange(n)[:,None], top_unsorted_indices]

# sort the top 5 values
top_sorted_indices = top_unsorted_indices[np.arange(m)[:,None,None], np.arange(n)[:,None], np.argsort(-top_values)]