Understanding Tensorflow maxpooling

Question

I am unable to understand why tensorflow maxpooling with my parameters.

When performed a maxpool with ksize=2 and strides=2 i get the following output with both padding SAME and padding VALID

input :  (?, 28, 28, 1)
conv2d_out :  (?, 28, 28, 32)
maxpool2d_out :  (?, 14, 14, 32)

But when I try to perfrom maxpool with ksize=3 and strides=1, I get the following output:

input :  (?, 28, 28, 1)
conv2d_out :  (?, 28, 28, 32)
maxpool2d_out :  (?, 28, 28, 32) PADDING SAME
maxpool2d_out :  (?, 26, 26, 32) PADDING VALID

maxpool with ksize=2 and strides=2 using padding SAME should have produced the output maxpool2d_out : (?, 28, 28, 32)

Is there something I missed out on how max pooling with padding works?

**CODE**==python_

Answer 1

I saw in your code you use padding=SAME. When using the SAME paddding and strides=1, the input and output size are the same. Why do you think the tensorflow implementation is wrong?

Update: According to tensorflow documentation:

Using SAME padding

out_height = ceil(float(in_height) / float(strides[1]))
out_width  = ceil(float(in_width) / float(strides[2]))

• It is 28/1=28 when k=3 and stride=1
• It is 28/2= 14 when k=2 and stride =2

Using VALID padding

out_height = ceil(float(in_height - filter_height + 1) / float(strides[1]))
out_width  = ceil(float(in_width - filter_width + 1) / float(strides[2]))

• It is celing((28-3+1)/1)= 26 when k=3, stride =1

• It is ceiling((28-2+1)/2)= 14 when k=2, stride=2

As you can see, because of the ceiling function, your result happens to be the same using different PADDING configuration

Answer 2

You're using padding='SAME', this means that your output will be padded with zeros in order to have the same size of the input.

If you change padding to VALID then the output is not padded with zeros and the pooling operation will work as you expect.