Numpy advanced indexing in multidimensional arrays; not just taking simple rows or columns

Question

I have a three dimensional numpy array. What is the fastest way to get a 3D array that has the largest item of each of final axis of the array without writing a loop.(I will later use CuPy with the same syntax, and loops would take away the GPU parallelism and the speed which is the most important factor here.)

Getting the indexes for largest items is easy:

>>> arr = np.array(
        [[[ 6, -2, -6, -5],
        [ 1, 12,  3,  9],
        [21,  7,  9,  8]],

       [[15, 12, 20, 12],
        [17, 15, 17, 23],
        [22, 18, 27, 32]]])
>>> indexes = arr.argmax(axis=2, keepdims=True)
>>> indexes
array([[[0],
        [1],
        [0]],

       [[2],
        [3],
        [3]]])

but how to use those indexes to get the chosen values from arr? All the ways I have tried either produce errors (such as arr[indexes]), or wrong results. What I would like to get in this example is

array([[[6],
        [12],
        [21]],

       [[20],
        [23],
        [32]]])

argmax docs suggest you see take_along_axis numpy.org/doc/stable/reference/generated/… — hpaulj
– hpaulj, Commented Aug 9, 2024 at 4:04
Thank you! I assumed there would be a simple answer but couldn't find it among so many numpy functions — Mikael
– Mikael, Commented Aug 9, 2024 at 7:54
Do you really need the intermediate step of getting the indices of the largest elements? If you are only interested in the largest elements themselves, you could use np.amax(), as proposed in the answer of "tax evader" below. Also you might want to give credit to the ones that helped you by (1) upvoting answers that you find helpful and (2) accepting the answer that solves your problem. — simon
– simon, Commented Aug 9, 2024 at 8:53
I will definitely upvote those who helped, when I have first acquired enough points to vote (=15 reputation points), currently I am at 11 points.... In my specific case I need the information of the indeces, yet without such need the np.amax() would indeed be better. I was about to make a comment on that after "tax evader"' answer, but then I read the guide that supposedly "Comments are used to ask for clarification or to point out problems in the post." This being my second post ever, I played safe and didn't make such other type of comment yet. — Mikael
– Mikael, Commented Aug 9, 2024 at 9:49
@Mikael Thanks for the clarification. Sorry I didn't remember that there was a minimum number of points to vote. In any case: welcome to Stack Overflow! — simon
– simon, Commented Aug 9, 2024 at 10:07

MSS · Accepted Answer · 2024-08-09 09:51:16Z

2

As pointed by hpaulj, you can use take_along_axis

>>>arr = np.array(
        [[[ 6, -2, -6, -5],
        [ 1, 12,  3,  9],
        [21,  7,  9,  8]],

       [[15, 12, 20, 12],
        [17, 15, 17, 23],
        [22, 18, 27, 32]]])
>>>indexes = arr.argmax(axis=2, keepdims=True)
>>>np.take_along_axis(arr , indexes , axis=2)

array([[[6],
        [12],
        [21]],

       [[20],
        [23],
        [32]]])

edited Aug 9, 2024 at 9:51

answered Aug 9, 2024 at 4:59

MSS

3,6932 gold badges28 silver badges57 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

vht981230 · Accepted Answer · 2024-08-09 04:04:48Z

2

I think you can use np.amax for that

import numpy as np

arr = np.array(
        [[[ 6, -2, -6, -5],
        [ 1, 12,  3,  9],
        [21,  7,  9,  8]],

       [[15, 12, 20, 12],
        [17, 15, 17, 23],
        [22, 18, 27, 32]]])

max_arr = np.amax(arr, axis=2, keepdims=True)
print(max_arr)

Output

[[[ 6]
  [12]
  [21]]

 [[20]
  [23]
  [32]]]

answered Aug 9, 2024 at 4:04

vht981230

5,3252 gold badges11 silver badges18 bronze badges

1 Comment

Mikael Over a year ago

Thank you! If the information of what indeces were the largest is not needed, this answer would be ideal.

Collectives™ on Stack Overflow

Numpy advanced indexing in multidimensional arrays; not just taking simple rows or columns

2 Answers 2

Comments

1 Comment

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

1 Comment

Your Answer

Sign up or log in

Post as a guest

Related