Issues about pooling output size

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

In the video “Inception Network Motivation” at 2:08, the orange volume which is the output of max pooling has the shape of (28, 28, 32) as shown below.

coursera问题
coursera问题878×421 172 KB

However, as taught in the previous videos, the output #channels should be the same as the input, thus it should be (28, 28, 128) with “same pooling”.
Is this an exception or a mistake?

1 Like
3

Hi NocturneJay,

What is missing in the Inception Network Motivation video is what is shown in the Inception Network video that follows. In that video, at 2:40 it becomes clear that the MAXPOOL layer is followed by a 1x1 convolution that shrinks the volume to 28x28x32 by using 32 filters.

4

Thank you @reinoudbosch ， I had found that truth after watching week 2’s videos.
But I think, this part may make some learners confused.

2 Likes
5

Yes NocturneJay, you are absolutely right.

It is good that you asked this question, so that an answer could be provided to those learners who may have become confused because of this. Thanks!

1 Like