Skip to content

This repository implements a ShuffleNet prototxt generator and a more efficient channel shuffle implementation

Notifications You must be signed in to change notification settings

twangnh/ShuffleNet-prototxt-generator

Repository files navigation

more efficient shuffle operation and prototxt generator

This repository implements a ShuffleNet prototxt generator and a more efficient channel shuffle operation, gives a benchmark of forward backward time of shufflenet 1x_g3 on Pascal Titan X.

farmingyard's implementation(https://github.com/farmingyard/ShuffleNet) of shufflechannel layer is not efficient enough as it use each thread to do a full channel shuffle operation. And if input image size(i.e., height and width) increase, the overhead of shufflechannel layer would be large. Here a permute layer of weiliu's implementation Permute_layer is used for more efficient channel shuffle operation.

Benchmark forward backward time

Test on TiTan X Pascal

To generate shufflenet prototxt

base net settings are the same as the original paper: more details can be found: ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile

first put shufflenet_generator.py in the same directory as caffe directory, or put it somewhere else and set the CAFFE_ROOT to point to caffe directory

then run something like python /pathto/shufflenet_generator.py --scale_factor 0.5 --group 1

scale_factor can be 0.25 0.5 1 1.5 2 group can be 1 2 3 4 8

To use the generated shufflenet prototxt

if you don't have permute layer yet, then put permute_layer.hpp into CAFFE_ROOT/include/caffe/layers/, put permute_layer.cpp and permute_layer.cu into CAFFE_ROOT/src/caffe/layers/

if you don't have depthWise convolution layer yet, then do the same as permute layer for source files from depthWise convolution layer

then you can safely change the line in caffe sourece to CHECK_LE(num_axes(), 5)

and then just do make under CAFFE_ROOT/

About

This repository implements a ShuffleNet prototxt generator and a more efficient channel shuffle implementation

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published