caffe训练：shufflenet和mobilenet层定义和加速

爱被打了一巴掌 2022-01-28 08:39 275阅读 0赞

以下是最近做项目时用到的，亲测有效。

1、ShuffleNet

层定义：[https://github.com/farmingyard/ShuffleNet][https_github.com_farmingyard_ShuffleNet]

预训练模型：[https://pan.baidu.com/s/1eS8NOm2][https_pan.baidu.com_s_1eS8NOm2]，提取码：bcj6  
（1）将cpp 和cu文件复制到caffe/src/caffe/layers目录下，将hpp文件复制到caffe/include/caffe/layers/目录下，   
（2）修改在caffe/src/caffe/proto目录下的caffe.proto文件

message LayerParameter \{  
...  
optional ShuffleChannelParameter shuffle\_channel\_param = 164;  
...  
\}  
...  
message ShuffleChannelParameter \{  
  optional uint32 group = 1\[default = 1\]; // The number of group  
\}  
（3）编译caffe   
cd caffe\_ROOT   
make all -j8

shufflenetv1和v2都有的版本：

[https://github.com/miaow1988/ShuffleNet\_V2\_pytorch\_caffe/releases][https_github.com_miaow1988_ShuffleNet_V2_pytorch_caffe_releases]

2、MobileNet

层定义：[https://github.com/farmingyard/caffe-mobilenet][https_github.com_farmingyard_caffe-mobilenet]

（1）将cpp 和cu文件复制到caffe/src/caffe/layers目录下，将hpp文件复制到caffe/include/caffe/layers/目录下，   
（2）编译caffe   
cd caffe\_ROOT   
make all -j8   
（3）替换相应的层的名字

参考博文：[https://blog.csdn.net/fengzhongluoleidehua/article/details/80797660][https_blog.csdn.net_fengzhongluoleidehua_article_details_80797660]

[https_github.com_farmingyard_ShuffleNet]: https://github.com/farmingyard/ShuffleNet
[https_pan.baidu.com_s_1eS8NOm2]: https://pan.baidu.com/s/1eS8NOm2
[https_github.com_miaow1988_ShuffleNet_V2_pytorch_caffe_releases]: https://github.com/miaow1988/ShuffleNet_V2_pytorch_caffe/releases
[https_github.com_farmingyard_caffe-mobilenet]: https://github.com/farmingyard/caffe-mobilenet
[https_blog.csdn.net_fengzhongluoleidehua_article_details_80797660]: https://blog.csdn.net/fengzhongluoleidehua/article/details/80797660