cross maps pooling

[Ran Manor]

Hi,

Is it possible with Caffe to do cross maps pooling?

For example, to average the outputs of each 8 maps ?

Thanks.

[Yangqing Jia]

You can decompose that to a group convolution with 1/8 being the
weights (initialized as this constant value), and then if you want to
do spatial pooling, add a normal pooling layer on the top.
Yangqing

> --
> You received this message because you are subscribed to the Google Groups
> "Caffe Users" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to caffe-users...@googlegroups.com.
> To post to this group, send email to caffe...@googlegroups.com.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/caffe-users/efd88d27-79ad-4aba-ab65-a24cbe975e38%40googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Ran Manor]

Thanks. And if I want to do max pooling between the maps?

[Yangqing Jia]

Then a custom pooling needs to be written :) Would not be too hard
though, computation wise it will look like the local response
normalization layer albeit much simpler.
Yangqing