centos6.5 下安装caffe-睿象云平台

centos6.5 下安装caffe

本站部分文章、图片属于网络上可搜索到的公开信息，均用于学习和交流用途，不能代表睿象云的观点、立场或意见。我们接受网民的监督，如发现任何违法内容或侵犯了您的权益，请第一时间联系小编邮箱jiasou666@gmail.com 处理。

centos6.5 下安装caffe

注：系统安装好后，先确认kernel kernel-headers kernel-devel kernel-firmware四个包的版本要相同

#rpm -qa |grep kernel

1	#rpm -qa \|grep kernel

注：先修改yum配置文件 /etc/yum.conf 修改 keepcache=1

1. 安装库

2、JDK安装

编译安装python2.7（必须先安装zlib与openssl的包再执行编译）

先安装gcc zlib openssl 等包

tar -xvf Python-2.7.9.tgz -C /usr/srccd /usr/src/Python-2.7.9./configure --enable-sharedmake -j12make altinstall

tar-xvf Python-2.7.9.tgz-C/usr/src

cd/usr/src/Python-2.7.9

./configure--enable-shared

make-j12

make altinstall

（altinstall在安装时会区分已存在的版本）(解决libpython2.7.so.1.0办法：vi /etc/ld.so.conf 添加/usr/local/lib，然后ldconfig)

替换系统中的python

ls -l `which python python2 python2.6`rm /usr/bin/pythonln -s -f /usr/local/bin/python2.7 /usr/bin/python

ls-l`which python python2 python2.6`

rm/usr/bin/python

ln-s-f/usr/local/bin/python2.7/usr/bin/python

保持yum可用性

vim /usr/bin/yum#!/usr/bin/python 改为 #!/usr/bin/python2.6

1 2	vim/usr/bin/yum #!/usr/bin/python 改为 #!/usr/bin/python2.6

4、安装pip

（必须先安装openssl-devel与zlib的包，再执行python编译，若执行该命令的时候出现红色cann't remove easy-install.pth的提醒，但目录下又无此文件，可新建后再次执行一遍命令，安装系统的时候最好把开发工具的选项给勾上，出现“Successfully installed pip-6.0.8 setuptools-14.3.1为安装成功”）

5、安装cuda-6.5及驱动

GTX 660显卡装cuda后会导致Xorg狂奔，直至系统死机，需要将/etc/inittab中的启动级别改为3注，驱动包文件结构不对，导致nvidia_uvm.ko模块无法编译，需手动解决

cd /var/lib/dkms/nvidia/346.46cp -rv /usr/src/nvidia-346.46 build

1 2	cd/var/lib/dkms/nvidia/346.46 cp-rv/usr/src/nvidia-346.46build

如果使用yum 方式安装的使用下载下的驱动包升级下

chmod +x NVIDIA-Linux-x86_64-346.72.run./NVIDIA-Linux-x86_64-346.72.run

1 2	chmod+xNVIDIA-Linux-x86_64-346.72.run ./NVIDIA-Linux-x86_64-346.72.run

重启后，dkms会在开机时完成nvidia_uvm.ko的编译

/lib/modules/版本号/extra/下有两个包：nvidia.ko nvidia_uvm.ko

lsmod|grep nvidiavi /etc/rc.local #编辑该文件modprobe nvidia_uvm #添加该条

lsmod|grep nvidia

vi/etc/rc.local#编辑该文件

modprobe nvidia_uvm#添加该条

5.1 run包安装方式

chmod +x cuda_6.5.19_linux_64.run./ cuda_6.5.19_linux_64.run

1 2	chmod+xcuda_6.5.19_linux_64.run ./cuda_6.5.19_linux_64.run

6、安装blas

yum -y install blas.x86_64 blas-devel.x86_64 \atlas.x86_64 atlas-devel.x86_64 atlas-sse3.x86_64 atlas-sse3-devel.x86_64

7、安装opencv

yum -y install ant.x86_64 gcc.x86_64 gcc-c++.x86_64 cmake.x86_64 git.x86_64 pkgconfig.x86_64 gtk2.x86_64 gtk2-devel.x86_64 libdc1394.x86_64 libdc1394-devel.x86_64 libjpeg-turbo.x86_64 libjpeg-turbo-devel.x86_64 libpng.x86_64 libpng-devel.x86_64 libtiff.x86_64 libtiff-devel.x86_64 jasper.x86_64 jasper-libs.x86_64 jasper-devel.x86_64 yasm.x86_64 yasm-devel.x86_64pip install numpy

yum-yinstall ant.x86_64 gcc.x86_64 gcc-c++.x86_64 cmake.x86_64 git.x86_64 pkgconfig.x86_64 gtk2.x86_64 gtk2-devel.x86_64 libdc1394.x86_64 libdc1394-devel.x86_64 libjpeg-turbo.x86_64 libjpeg-turbo-devel.x86_64 libpng.x86_64 libpng-devel.x86_64 libtiff.x86_64 libtiff-devel.x86_64 jasper.x86_64 jasper-libs.x86_64 jasper-devel.x86_64 yasm.x86_64 yasm-devel.x86_64

pip install numpy

安装ffmpeg： #此包不需要通过yum安装，yum安装版本不对

tar -xf ffmpeg-2.6.1.tar.bz2 -C /usr/srccd /usr/src/ffmpeg-2.6.1/./configure --enable-shared #要以共享库方式配置，否则opencv编译时链接静态库会出错make -j12 && make installunzip opencv-2.4.9cd opencv-2.4.9mkdir release && cd release

tar-xf ffmpeg-2.6.1.tar.bz2-C/usr/src

cd/usr/src/ffmpeg-2.6.1/

./configure--enable-shared#要以共享库方式配置，否则opencv编译时链接静态库会出错

make-j12&&make install

unzip opencv-2.4.9

cd opencv-2.4.9

mkdir release&&cd release

修改源文件NCVPixelOperations.hpp，文件替换到opencv路径下的modules/gpu/src/nvidia/core/NCVPixelOperations.hpp配置环境变量：

vim /etc/profile.d/custom.sh 配置完成source /etc/profile.d/custom.sh#!/bin/bashexport PATH=/usr/local/MATLAB/R2014a/bin:/usr/local/cuda-6.5/bin:$PATHexport LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/boost-1.55.0/lib:/usr/local/cuda-6.5/lib64:/opt/caffe-master/build/lib:/usr/lib64/atlasexport LIBRARY_PATH=$LIBRARY_PATH:/usr/local/boost-1.55.0/lib:/usr/local/cuda-6.5/lib64:/opt/caffe-master/build/lib:/usr/lib64/atlasexport C_INCLUDE_PATH=$C_INCLUDE_PATH:/usr/local/boost-1.55.0/include:/usr/local/cuda-6.5/include:/opt/caffe-master/build/src:/opt/caffe-master/includeexport CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/usr/local/boost-1.55.0/include:/usr/local/cuda-6.5/include:/opt/caffe-master/build/src:/opt/caffe-master/includeexport PYTHONPATH=$PYTHONPATH:/opt/caffe-master/pythonexport HISTTIMEFORMAT="%F %T "cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local ..make –j12make install

vim/etc/profile.d/custom.sh配置完成source/etc/profile.d/custom.sh

#!/bin/bash

export PATH=/usr/local/MATLAB/R2014a/bin:/usr/local/cuda-6.5/bin:$PATH

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/boost-1.55.0/lib:/usr/local/cuda-6.5/lib64:/opt/caffe-master/build/lib:/usr/lib64/atlas

export LIBRARY_PATH=$LIBRARY_PATH:/usr/local/boost-1.55.0/lib:/usr/local/cuda-6.5/lib64:/opt/caffe-master/build/lib:/usr/lib64/atlas

export C_INCLUDE_PATH=$C_INCLUDE_PATH:/usr/local/boost-1.55.0/include:/usr/local/cuda-6.5/include:/opt/caffe-master/build/src:/opt/caffe-master/include

export CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/usr/local/boost-1.55.0/include:/usr/local/cuda-6.5/include:/opt/caffe-master/build/src:/opt/caffe-master/include

export PYTHONPATH=$PYTHONPATH:/opt/caffe-master/python

export HISTTIMEFORMAT="%F %T "

cmake-DCMAKE_BUILD_TYPE=RELEASE-DCMAKE_INSTALL_PREFIX=/usr/local..

make–j12

make install

8、安装boost-1.55（1.56不兼容）

yum -y install libicu.x86_64 libicu-devel.x86_64 bzip2-libs.x86_64 bzip2-devel.x86_64tar –xf boost_1_55_0.tar.gz && cd boost_1_55_0./bootstrap.sh./b2./b2 install

yum-yinstall libicu.x86_64 libicu-devel.x86_64 bzip2-libs.x86_64 bzip2-devel.x86_64

tar–xf boost_1_55_0.tar.gz&&cd boost_1_55_0

./bootstrap.sh

./b2

./b2 install

运行./b2 install命令，默认安装在/usr/local/lib目录下，头文件在/usr/local/include/boost目录下

9、安装caffe其他依赖：

yum -y install snappy.x86_64 snappy-devel.x86_64 hdf5.x86_64 hdf5-devel.x86_64 epel-release leveldb.x86_64 leveldb-devel.x86_64 libgfortran.x86_64------ 编译安装protobuf-2.5.0 protobuf-2.5.0------ tar -xvf protobuf-2.5.0.tar.gz------ cd /usr/src/protobuf-2.5.0./configuremakemake checkmake install------ export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

yum-yinstall snappy.x86_64 snappy-devel.x86_64 hdf5.x86_64 hdf5-devel.x86_64 epel-release leveldb.x86_64 leveldb-devel.x86_64 libgfortran.x86_64

------编译安装protobuf-2.5.0protobuf-2.5.0

------tar-xvf protobuf-2.5.0.tar.gz

------cd/usr/src/protobuf-2.5.0

./configure

make

make check

make install

------export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

10、编译安装caffe其他依赖包

11、将matlab上传至服务器，通过图形方式安装

安装秘钥12345-67890-12345-67890安装完成后导入lic文件，然后替换libmwservices.so到/usr/local/MATLAB/R2014a/bin/glnxa64/进行覆盖，结束安装。

12、解决python依赖

pip install 'six>=1.3'easy_install -U distributepip2.7 install PIL --allow-external PIL --allow-unverified PIL

pip install'six>=1.3'

easy_install-Udistribute

pip2.7install PIL--allow-external PIL--allow-unverified PIL

解包caffe-master.zip，并将该包移至opt目录

cd /opt/caffe-master/pythonfor i in $(cat requirements.txt); do pip install $i; done #需要多执行几遍

1 2	cd/opt/caffe-master/python foriin$(cat requirements.txt);dopip install$i;done#需要多执行几遍

注：会出现一个报错，关于PIL.Image >= 1.1.7，则可使用命令pip install 'PIL' 进行安装后再次执行以上的for循环语句，需要将python升级至2.7以上版本（安装及注意事项下：）

13、安装caffe

修改caffe-master/Makefile.config文件，增加如下几句

cp /opt/caffe-master/Makefile.config.example Makefile.configvim Makefile.configMATLAB_DIR := /usr/local/MATLAB/R2014a/BLAS := atlasBLAS_LIB := /usr/lib64/atlasPYTHON_INCLUDE := /usr/include/python2.7 \/usr/local/lib/python2.7/site-packages/numpy/core/include \/usr/local/include/python2.7执行 ldconfigmake all -j12make –j12 pycaffemake –j12 matcaffemake test –j12make runtest –j12

cp/opt/caffe-master/Makefile.config.example Makefile.config

vim Makefile.config

MATLAB_DIR:=/usr/local/MATLAB/R2014a/

BLAS:=atlas

BLAS_LIB:=/usr/lib64/atlas

PYTHON_INCLUDE:=/usr/include/python2.7\

/usr/local/lib/python2.7/site-packages/numpy/core/include\

/usr/local/include/python2.7

执行ldconfig

make all-j12

make–j12 pycaffe

make–j12 matcaffe

make test–j12

make runtest–j12

如果matlab要使用静态编译libprotobuf.a的话，修改Makefile在MATLAB_CXXFLAGS项上添加-static参数即可但使用动态库的matlab模型可能不可用

若一切没有问题，至此caffe环境安装结束，待测试。

以下为可选部分

编译安装protobuf-2.5.0 protobuf-2.5.0

tar -xvf protobuf-2.5.0.tar.gzcd /usr/src/protobuf-2.5.0./configuremakemake checkmake installexport LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

tar-xvf protobuf-2.5.0.tar.gz

cd/usr/src/protobuf-2.5.0

./configure

make

make check

make install

export LD_LIBRARY_PATH=/usr/local/lib:$LD_LIBRARY_PATH

继续安装protobuf的python模块(如果不用python，可跳过这一步)

#cd ./python #python setup.py build #python setup.py test #python setup.py install

#cd ./python

#python setup.py build

#python setup.py test

#python setup.py install

安装cudnnLINUX

a. 编辑确保Makefile.config,启用GPU “# CPU_ONLY := 1”,并设置 “USE_CUDNN := 1”。

b. 安装cuDNN

tar -xzvf cudnn-6.5-linux-R1.tgzcd cudnn-6.5-linux-R1cp lib* /usr/local/cuda-6.5/lib64/cp cudnn.h /usr/local/cuda-6.5/include/cd /usr/local/cuda-6.5/lib64/rm -rf libcudnn.so libcudnn.so.6.5chmod u=rwx,g=rx,o=rx libcudnn.so.6.5.18ln -s libcudnn.so.6.5.18 libcudnn.so.6.5ln -s libcudnn.so.6.5 libcudnn.soldconfig

tar-xzvf cudnn-6.5-linux-R1.tgz

cd cudnn-6.5-linux-R1

cp lib*/usr/local/cuda-6.5/lib64/

cp cudnn.h/usr/local/cuda-6.5/include/

cd/usr/local/cuda-6.5/lib64/

rm-rf libcudnn.so libcudnn.so.6.5

chmodu=rwx,g=rx,o=rx libcudnn.so.6.5.18

ln-slibcudnn.so.6.5.18libcudnn.so.6.5

ln-slibcudnn.so.6.5libcudnn.so

ldconfig

注1：将相关的头文件，库文件放到profile中定义的系统路径里即可，matlab的mex运行时需要加载对应库

caffe编译时也可在Makefile.config中修改，添加cuDNN的路径/cache/INSTALL_cuDNN/cudnn-6.5-linux-R1

INCLUDE_DIRS := $(PYTHON_INCLUDE) /usr/local/include /cache/INSTALL_cuDNN/cudnn-6.5-linux-R1

LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib /cache/INSTALL_cuDNN/cudnn-6.5-linux-R1

注2：在使用tesla-c2050显卡时，需要在Makefile.config里改如下几个地方：

PYTHON_LIB := /usr/lib64 #原为PYTHON_LIB := /usr/libLIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib64 /usr/lib64 #原为如下：LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib

注3：protobuf手动安装，不需要通过yum，yum安装版本不对，make runtest会报错，使用protobuf2.5的版本，安装方式见上，在编译caffe前安装好后再进行编译。

包：咖啡环境需要上传的包：gflags-master.zip、opencv-2.4.9.zip、boost_1_55_0.tar.gz、caffe-master.zip、glog-0.3.3.tar.gz、protobuf-2.5.0.tar.gz、cuda-repo-rhel6-6.5-14.x86_64.rpm、jdk-7u25-linux-x64.tar.gz、lmdb.tar、Python-2.7.9.tgz、 ffmpeg-2.6.1.tar.bz2、rpmforge-release-0.5.3-1.el6.rf.x86_64.rpm、NVIDIA-Linux-x86_64-346.72.run、NCVPixelOperations.hpp、matlab文件夹

作者：峥叔、小葱（③群），感谢两位投稿！