物体检测框架RON的shufflenet实现

一。修改caffe源码

1.shuffle_channel_layer.cpp

2.shuffle_channel_layer.cu

3.shuffle_channel_layer.hpp

4.conv_dw_layer.cpp

5.conv_dw_layer.cu

6.conv_dw_layer.hpp

文件下载：百度网盘

将cpp和cu文件放在caffe-path/src/caffe/layers中

将hpp头文件放在caffe-path/include/caffe/layers中

修改caffe-path/src/caffe/proto/caffe.proto

添加

optional ShuffleChannelParameter shuffle_channel_param = 164;

message ShuffleChannelParameter {
optional uint32 group = 1[default = 1]; // The number of group
}

重新编译caffe

mkdir build
cd build
cmake ..
make -j8
make pycaffe

二。修改RON网络，用shufflenet前16层替换VGG16，具体网络见： 百度网盘

train_shufflenet_1x_g3_ron_cudnn320.prototxt

三。截取shufflenet前16层网络作为预训练网络模型

python cutNetwork.py

四。training

#!/usr/bin/env bash
python tools/train_net.py --gpu 0 \
                          --solver ./ShuffleNet/solver.prototxt \
                          --imdb voc_2007_trainval+voc_2012_trainval \
                          --weights ./yesmodel_shufflenet/shufflenet_1x_g3_16layer.caffemodel --batchsize 16 \
                          --iters 200000