0
$\begingroup$

Question RE this research paper if anyone has experience with CNN's, pooling & skip connections: https://arxiv.org/pdf/1606.01781.pdf

  • In figure 1, the input to the first convolutional block has shape (batch_size, 64, s)
  • The output from block 1 must be (batch_size, 64, s)
  • The output from block 2 must be (batch_size, 64, s)

However the output from the pooling step has shape (batch_size, 128, s/2). How can a pooling step increase the number of parameters on axis 1 from 64 to 128?!

My guess is that the input to the pooling layer is actually a concatenation of 2 of the previous layer's outputs. In this case it would have input shape (batch_size, 128, s). However, the paper does not appear to clearly specify at what outputs are concatenated...

Can anyone clarify how this is the case?

$\endgroup$
0
$\begingroup$

As you mentioned, the paper doesn't clarify. However, my guess is that this is not due to concatenating 2 previous layers (I don't really see an specific reason to do this here) but because of concatenating the ResNet shorcut.

Generally, all conv layers have a number of filters, thus determining the output size (num_filters, size) regardless the inputs. On the other hand, MaxPooling does keep the input num_filters (though in this case reducing the size). In the paper, note that the num_filters is doubled at the output of all convlayer except for the one that does not keep the ResNet shorcut (last 512 conv layer). So my guess is that they are concatenating the output of the conv layer and the shorcut which would explain the output size.

Hope this helps!

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.