search

yonghenglh6 / DepthwiseConvolution

A personal depthwise convolution layer implementation on caffe by liuhao.(only GPU)
525 stars 184 forks source link

Why slower than tensorflow #16

Closed twmht closed 6 years ago

twmht commented 6 years ago

In Mobilenet-tensorflow (https://github.com/Zehaos/MobileNet), The speed in GTX-1080 is only 3ms, but caffe takes 8ms.

What makes the difference between caffe and tensorflow?

thanks

yonghenglh6 commented 6 years ago

This implement is a temporary patch in origin caffe, so the speed is not the fastest. My code is naive, which just make up the serial-process-order bug between groups in origin caffe and give us the ability to train the net. The difference is so big that I cannot estimate it.