cvjena / alpha_pooling Goto Github PK

Dear @cvjena,
I am trying to replace average pooing with alpha-pooling of ResNet. I plan to use this modified ResNet to train ImageNet-1000. I notice that your modified VGG-16 has some "SpatialTransformer" stuff, which is out of my best knowledge. Here is my question:

• Is it necessary to use "SpatialTransformer" and "alpha-pooling" at the same time?
• alpha-pooling = SignedPower + CompactBilinear ? or alpha-pooling = SignedPower + CompactBilinear + SignedPower?

I am really appreciated if you have other guide

Thanks

Hi, @MarcelSimon

Thanks for sharing this code!

I am trying to find the code that calculate the gradient w.r.t. alpha, and I check the file caffe_pp2/src/caffe/layers/signed_power_layer.cpp
but can not find it, would you please show me which lines compute this? I only find the gradient w.r.t. bottom data.

Thank you!

This work is so fancy. I wonder that how to get the intriguing visualization like Figure2? Could you share the plotting codes for it? Thank you so much.