Caffe移植到windows

一、前言

Caffe是一个很棒的深度学习框架，值得花时间学习和分享，最近在做移植的工作，有必要做一个总结，俗话说的好，好记性不如C博。

我的开发环境是：windows7 64位、gpu 1080ti、opencv2.4.11、boost1.61.0

二、依赖库

1、安装OpenCV

可以参考博客OpenCV在VS中的配置

2、安装boost

可以参考博客编译并使用boost库(win7+boost1.60+vs2013)

3、安装cuda

可以参考博客对应部分win7+1080ti+cuda8.0+cudnn5.0+caffe安装

4、下载其他库

比较常用的是Neil Z. SHAO提供的已编译好的库，风翼冰舟提供的百度云下载地址，链接：http://pan.baidu.com/s/1eSku9aE 密码：qrj7。

解压之后有两个文件夹，新建一个文件夹这里命名为caffe_in_windows，将3rdparty文件夹拖进caffe_in_windows中。

三、Caffe在VS中的配置、编译

接下来分两部分说，第一部分配置编译libcaffe，我用到的相关文件下载地址：链接：http://pan.baidu.com/s/1pLnho0N 密码：e580

1、下载官方最新的caffe源文件，地址https://github.com/BVLC/caffe。

2、将caffe-master内部所有文件放进caffe_in_windows文件夹。

3、新建caffe_in_windows/windows文件夹，在windows文件夹下新建一个VS控制台空白工程，命名为Caffe，然后添加一个lib工程，命名为libcaffe。

4、打开VIEW->Other Windows->Property Manager属性管理窗口。

5、右键libcaffe工程打开属性窗口，将平台改为x64

6、在属性窗口中找到General->Output Directory，在Debug和Release均将其修改为

$(SolutionDir)..\$(Platform)\$(Configuration)\

7、找到C/C++->General->Additional Include Directories，在Debug和Release均添加以下路径

$(Boost_path)$(CUDA_PATH)\include$(SolutionDir)..\..\src$(SolutionDir)..\..\include$(OpenCV_build_path)\include$(OpenCV_build_path)\include\opencv$(OpenCV_build_path)\include\opencv2$(SolutionDir)..\..\3rdparty\include$(SolutionDir)..\..\3rdparty\include\lmdb$(SolutionDir)..\..\3rdparty\include\hdf5$(SolutionDir)..\..\3rdparty\include\openblas

其中$(SolutionDir)代表解决方案.sln所在的路径，其余的$(XXX)都代表电脑的相关路径，可以在系统变量中新建对应变量，将路径添加进去，然后在path变量末尾添加系统变量，程序运行时就能找到对应的dll。

8、找到Librarian->General->Additional Library Directories，在Debug和Release均添加以下路径

$(CUDA_LIB_PATH)$(Boost_lib_path)$(SolutionDir)..\..\3rdparty\lib$(OpenCV_build_path)\x64\vc12\lib

9、找到Librarian->General->Additional Dependencies，在Debug模式下添加以下依赖

opencv_core2411d.lib;opencv_calib3d2411d.lib;opencv_contrib2411d.lib;opencv_flann2411d.lib;opencv_highgui2411d.lib;opencv_imgproc2411d.lib;opencv_legacy2411d.lib;opencv_ml2411d.lib;opencv_gpu2411d.lib;opencv_objdetect2411d.lib;opencv_photo2411d.lib;opencv_features2d2411d.lib;opencv_nonfree2411d.lib;opencv_stitching2411d.lib;opencv_video2411d.lib;opencv_videostab2411d.lib;cudart.lib;cuda.lib;nppi.lib;cufft.lib;cublas.lib;curand.lib;

在Release下添加以下依赖

opencv_core2411.lib;opencv_calib3d2411.lib;opencv_contrib2411.lib;opencv_flann2411.lib;opencv_highgui2411.lib;opencv_imgproc2411.lib;opencv_legacy2411.lib;opencv_ml2411.lib;opencv_gpu2411.lib;opencv_objdetect2411.lib;opencv_photo2411.lib;opencv_features2d2411.lib;opencv_nonfree2411.lib;opencv_stitching2411.lib;opencv_video2411.lib;opencv_videostab2411.lib;cudart.lib;cuda.lib;nppi.lib;cufft.lib;cublas.lib;curand.lib;

10、新建windows\scripts文件夹，将ProtoCompile.cmd放进windows\scripts中，在属性窗口中找到Build Events -> Pre-Build Event -> Command Line，在debug和Release模型下均添加

"$(SolutionDir)..\scripts\ProtoCompile.cmd" "$(SolutionDir)" "$(SolutionDir)..\..\3rdparty\bin\"

编译libcaffe工程，可以看到输出窗口已经生成了caffe.pb.h、caffe.pb.cc

11、将src/caffe下的所有文件添加进libcaffe

12、打开common.cpp，在开头添加

#if defined(_MSC_VER)#include <process.h>#define getpid() _getpid()#endif

找到::google::InstallFailureSignalHandler();所在的行，改为

#if !defined(_MSC_VER)  ::google::InstallFailureSignalHandler();#endif

在属性窗口找到C/C++->Precompiled Header->Precompiled Header，改为Not Using Precompiled Headers

找到C/C++->Advanced->Disable Specific Warnings，添加4996/4703
找到C/C++->Preprocessor->Preprocessor Definitions，添加

_CRT_SECURE_NO_WARNINGSUSE_OPENCVUSE_CUDNNUSE_LEVELDBUSE_LMDB

13、打开layers/bnll_layer.cu，在开头找到const float kBNLL_THRESHOLD = 50.;所在的行，改为

//const float kBNLL_THRESHOLD = 50.;#define kBNLL_THRESHOLD 50.0

14、打开util/io.cpp，在开头添加

#if defined(_MSC_VER)#include <io.h>#endif

15、打开util/signal_handler.cpp，在第16行左右的地方，修改为

#ifdef _MSC_VERcase SIGBREAK:  // there is no SIGHUP in windows, take SIGBREAK instead.got_sighup = true;break;#elsecase SIGHUP:got_sighup = true;break;#endif

在第36行左右的地方，修改为

#ifdef _MSC_VERif (signal(SIGBREAK, handle_signal) == SIG_ERR) {LOG(FATAL) << "Cannot install SIGBREAK handler.";}if (signal(SIGINT, handle_signal) == SIG_ERR) {LOG(FATAL) << "Cannot install SIGINT handler.";}#elsestruct sigaction sa;// Setup the handlersa.sa_handler = &handle_signal;// Restart the system call, if at all possiblesa.sa_flags = SA_RESTART;// Block every signal during the handlersigfillset(&sa.sa_mask);// Intercept SIGHUP and SIGINTif (sigaction(SIGHUP, &sa, NULL) == -1) {LOG(FATAL) << "Cannot install SIGHUP handler.";}if (sigaction(SIGINT, &sa, NULL) == -1) {LOG(FATAL) << "Cannot install SIGINT handler.";}#endif

在第64行左右的地方，修改为

#ifdef _MSC_VERif (signal(SIGBREAK, SIG_DFL) == SIG_ERR) {LOG(FATAL) << "Cannot uninstall SIGBREAK handler.";}if (signal(SIGINT, SIG_DFL) == SIG_ERR) {LOG(FATAL) << "Cannot uninstall SIGINT handler.";}#elsestruct sigaction sa;// Setup the sighub handlersa.sa_handler = SIG_DFL;// Restart the system call, if at all possiblesa.sa_flags = SA_RESTART;// Block every signal during the handlersigfillset(&sa.sa_mask);// Intercept SIGHUP and SIGINTif (sigaction(SIGHUP, &sa, NULL) == -1) {LOG(FATAL) << "Cannot uninstall SIGHUP handler.";}if (sigaction(SIGINT, &sa, NULL) == -1) {LOG(FATAL) << "Cannot uninstall SIGINT handler.";}#endif

16、打开util/db_lmdb.cpp，在开头添加

#if defined(_MSC_VER)#include <direct.h>#define mkdir(X, Y) _mkdir(X)#endif

17、打开proto/caffe.pb.cc，将

#include "caffe.pb.h"

改为

#include "caffe/proto/caffe.pb.h"

至此，libcaffe工程应该没有什么错误了，可以先一个一个.cpp、.cu文件编译看有没有错误，.cu文件需要右键属性，在General->Item Type中将其修改为CUDA C/C++，并且所有的.cu文件都需要修改，可以按住Ctrl选中所有的cu文件，然后一次性修改Type。没有错误的话，可以编译libcaffe工程了，应该没有太大问题。

现在开始配置编译Caffe工程

1、右键Caffe工程打开属性窗口，将平台改为x64

2、在属性窗口中找到General->Output Directory，在Debug和Release均将其修改为

$(SolutionDir)..\$(Platform)\$(Configuration)\

3、找到General->Character Set，将其改为使用多字节字符

4、找到C/C++->SDL checks，更改为no

5、找到VC++ Directories->Include Directories，在Debug和Release均添加以下路径

$(Boost_path)$(CUDA_PATH)\include$(SolutionDir)..\..\include$(OpenCV_build_path)\include$(OpenCV_build_path)\include\opencv$(OpenCV_build_path)\include\opencv2$(SolutionDir)..\..\3rdparty\include$(SolutionDir)..\..\3rdparty\include\lmdb$(SolutionDir)..\..\3rdparty\include\hdf5$(SolutionDir)..\..\3rdparty\include\openblas

6、找到VC++ Directories->Library Directories，在Debug和Release均添加以下路径

$(OutDir)$(CUDA_LIB_PATH)$(Boost_lib_path)$(SolutionDir)..\..\3rdparty\lib$(OpenCV_build_path)\x64\vc12\lib

7、找到Linker->Input->Additional Dependencies，在Debug模型下添加以下依赖

libcaffe.libopencv_core2411d.libopencv_calib3d2411d.libopencv_contrib2411d.libopencv_flann2411d.libopencv_highgui2411d.libopencv_imgproc2411d.libopencv_legacy2411d.libopencv_ml2411d.libopencv_gpu2411d.libopencv_objdetect2411d.libopencv_photo2411d.libopencv_features2d2411d.libopencv_nonfree2411d.libopencv_stitching2411d.libopencv_video2411d.libopencv_videostab2411d.libcudart.libcuda.libnppi.libcufft.libcublas.libcurand.libgflagsd.libleveldbd.liblibglog.liblibhdf5_D.liblibhdf5_hl_D.liblibopenblas.dll.alibprotobufd.liblmdbd.libboost_system-vc120-mt-gd-1_61.libboost_chrono-vc120-mt-gd-1_61.libboost_date_time-vc120-mt-gd-1_61.libboost_filesystem-vc120-mt-gd-1_61.libboost_thread-vc120-mt-gd-1_61.libShlwapi.lib

在Release模式下添加

libcaffe.libopencv_core2411.libopencv_calib3d2411.libopencv_contrib2411.libopencv_flann2411.libopencv_highgui2411.libopencv_imgproc2411.libopencv_legacy2411.libopencv_ml2411.libopencv_gpu2411.libopencv_objdetect2411.libopencv_photo2411.libopencv_features2d2411.libopencv_nonfree2411.libopencv_stitching2411.libopencv_video2411.libopencv_videostab2411.libcudart.libcuda.libnppi.libcufft.libcublas.libcurand.libgflags.libleveldb.liblibglog.liblibhdf5.liblibhdf5_hl.liblibopenblas.dll.alibprotobuf.liblmdb.libboost_system-vc120-mt-1_61.libboost_chrono-vc120-mt-1_61.libboost_date_time-vc120-mt-1_61.libboost_filesystem-vc120-mt-1_61.libboost_thread-vc120-mt-1_61.libShlwapi.lib

编译Caffe工程，应该没有太大问题，如果没有配置对的话，可能会出现应用程序无法正常启动的错误。

这是release模式下的运行结果

这是debug模式下运行结果

四、参考文献

我主要用到了两篇博客，在这里表示感谢

【caffe-windows】Linux至Windows平台的caffe移植

Build Caffe in Windows with Visual Studio 2013 + CUDA 6.5 + OpenCV 2.4.9