Do "C_Relu" really speed up the network?

sanghoon / pva-faster-rcnn

Demo code for PVANet

https://arxiv.org/abs/1611.08588

Other

651 stars 241 forks source link

Do "C_Relu" really speed up the network? #37

Open catsdogone opened 7 years ago

catsdogone commented 7 years ago

I test the C_Relu(with scale after concat) in a network. The original conv layer is 52877, I change this layer to C_Relu structure which means that now the conv layer is 26477. But the training time increased unexpected. I use CUDA8.0 and CUDNN5.1. I feel confused. Can you give me some advice? Thank you very much. @sanghoon

songjmcn commented 7 years ago

I think tthi is Python problem, not C_Relu

catsdogone commented 7 years ago

@songjmcn While there is not Python layer in may new net and I use the c++ interface of caffe.

songjmcn commented 7 years ago

The training use Python

sanghoon commented 7 years ago

Hi @catsdogone We've added C_ReLU mainly for test efficiency. I'm not 100% sure whether it can accelerate network training. And as @songjmcn mentioned, it'll depend on implementations