max_pooling.py

from PIL import Image
import numpy as np

def main():

    im = Image.open("image.jpg").convert('LA') #Can be many different formats.
    pix = im.load()

    #emulate CaffeNet implementation.
    im = im.resize((227, 227))

    kernel_size = 3
    stride = 2
    padding = 0
    pooling_type = max

    pooling_height = (im.size[1] + (2 * padding) - kernel_size) / (stride + 1);
    pooling_width = (im.size[0] + (2 * padding) - kernel_size) / (stride + 1);

    im = im.resize((pooling_width, pooling_height))

    val = []
    secondval = []

    for i in xrange(0, im.size[1] + padding, stride):
        for x in xrange(0, im.size[0] + padding, stride):
            for y in xrange(0, kernel_size**2, 2):
                #since we're assuming this is single channel, they will all ahve the same RGB value for R G and B.
                #This is why we are only looking at the first element in the tuple.
                if(i >= im.size[1] or x >= im.size[0]):
                    val.append(0)
                    continue


                val.append(pix[x, i][0])

                if x+1 >= im.size[0]:
                    secondval.append(0)
                    continue

                secondval.append(pix[x+1, i][0])

            pix[x,i] = (pooling_type(val), pooling_type(secondval))
            val = []
            secondval = []

    print im.size
    im.save('pooled_image.png')

if __name__ == '__main__':
    main()