Inconsistency between caffe and pytorch for max-pooling

I think this is what you want :slight_smile: