This is an efficiency implementation(GPU Version) of Depthwise Convolution in caffe. It is even better and faster than other implementations. Just copy the files to merge your original caffe,and compile it directly without adding any new parameter to the caffe.proto Have fun!