Maxpooling 2x2 array only using numpy

Question

I want help in maxpooling using numpy. I am learning Python for data science, here I have to do maxpooling and average pooling for 2x2 matrix, the input can be 8x8 or more but I have to do maxpool for every 2x2 matrix. I have created an matrix by using

k = np.random.randint(1,64,64).reshape(8,8)

So hereby I will be getting 8x8 matrix as a random output. Form the result I want to do 2x2 max pooling. Thanks in advance

Answer 1

You don't have to compute the necessary strides yourself, you can just inject two auxiliary dimensions to create a 4d array that's a 2d collection of 2x2 block matrices, then take the elementwise maximum over the blocks:

import numpy as np

# use 2-by-3 size to prevent some subtle indexing errors
arr = np.random.randint(1, 64, 6*4).reshape(6, 4)

m, n = arr.shape
pooled = arr.reshape(m//2, 2, n//2, 2).max((1, 3))

An example instance of the above:

>>> arr
array([[40, 24, 61, 60],
       [ 8, 11, 27,  5],
       [17, 41,  7, 41],
       [44,  5, 47, 13],
       [31, 53, 40, 36],
       [31, 23, 39, 26]])

>>> pooled
array([[40, 61],
       [44, 47],
       [53, 40]])

---

For a completely general block pooling that doesn't assume 2-by-2 blocks:

import numpy as np

# again use coprime dimensions for debugging safety
block_size = (2, 3)
num_blocks = (7, 5)
arr_shape = np.array(block_size) * np.array(num_blocks)
numel = arr_shape.prod()
arr = np.random.randint(1, numel, numel).reshape(arr_shape)

m, n = arr.shape  # pretend we only have this
pooled = arr.reshape(m//block_size[0], block_size[0],
                     n//block_size[1], block_size[1]).max((1, 3))