"Down-resolution" a numpy array

Question

I have a numpy array of shape [5sx,5sy,5sz] and I’d like to create a tensor with shape [sx,sy,sz] where the value of each element is:

new_array[x,y,z] = mean(array[5x:5x+5,5y:5y+5,5z:5z+5])

I could do this in a loop, but is there a faster one-line approach?

(I'd really ideally like to do it where the size of the origin array is not an integer multiple of the new one, but that seems like a much harder question, and I thought this would be a good first step at least.)

Answer 1

You could use scikit-image block_reduce:

import numpy as np
import skimage.measure

arr5 = np.random.rand(20, 20, 20)
arr  = skimage.measure.block_reduce(arr5, (5, 5, 5), np.mean)
print(f'arr.shape = {arr.shape}')
print(f'arr5[:5, :5, :5].mean() = {arr5[:5, :5, :5].mean()}')
print(f'arr[0, 0, 0] = {arr[0, 0, 0]}')

output:

arr.shape = (4, 4, 4)
arr5[:5, :5, :5].mean() = 0.47200241666948467
arr[0, 0, 0] = 0.4720024166694848

If you're dealing with large arrays and you have a GPU available I would advise you to look into pytorch's average pooling.