kaiminghe / resnet-1k-layers Goto Github PK

@KaimingHe Referencing your paper (http://arxiv.org/abs/1603.05027) ,I reimplement the pre-activation unit of ResNet-110, also the baseline ResNet-110 in (http://arxiv.org/abs/1512.03385) with caffe. The baseline 's performance is pretty good, but the pre-activation unit is not. This is the network structure of pre-act ResNet-110: (http://ethereon.github.io/netscope/#/gist/41d6e48f15d1c0b154b499602fb4a529) Are there any problem?

"For the first Residual Unit (that follows a stand-alone convolutional layer, conv1), we adopt the first activation right after conv1 and before splitting into two paths."
This is a part of your paper "Identity Mappings in Deep Residual Networks". What make me confusing is that the "resnet-1k-layers" doesn't obey that. Why? I will be great thankful if you can answer the question.

I hope to train your network on Caffe or MXNet. Do you have your network spec file with Caffe or MXNet format?