Caffe安装过程记录(CentOS,无独立显卡,无GPU)
原文地址:http://www.aiuxian.com/article/p-2410195.html
参考资料:
http://www.tuicool.com/articles/uiuA3e Caffe 安装配置(CentOS + 无GPU)
http://blog.sina.com.cn/s/blog_990865340102vewt.html caffe 安装配置(CentOS 6.5 + 无GPU)
http://www.cnblogs.com/platero/p/3993877.html Caffe + Ubuntu 14.04 64bit + CUDA 6.5 配置说明
http://www.tuicool.com/articles/uiuA3e Caffe 安装配置(CentOS + 无GPU)
http://blog.sina.com.cn/s/blog_990865340102vewt.html caffe 安装配置(CentOS 6.5 + 无GPU)
http://www.cnblogs.com/platero/p/3993877.html Caffe + Ubuntu 14.04 64bit + CUDA 6.5 配置说明
http://blog.sciencenet.cn/blog-1583812-841855.html Ubuntu+CUDA6.5+Caffe安装配置汇总
Caffe( http://caffe.berkeleyvision.org/ )是一个清晰而高效的深度学习框架,其作者是博士毕业于UC Berkeley的贾扬清(http://daggerfs.com/ ),他目前在Google工作。Caffe是纯粹的C++/CUDA架构,支持命令行、Python和MATLAB接口;可以在CPU和GPU直接无缝切换.
3月11日补充:当时安装完,是可以执行caffe里面的相关示例,可以修改protobuf 和不同的网络配置,但是如果想用python编写程序,执行还是失败的。问题再最后补充。
1 首先先鉴定机器是什么显卡:
lspci |grep VGA
00:02.0 VGA compatible controller: Intel Corporation Xeon E3-1200 v2/3rd Gen Core processor Graphics Controller (rev 09)
机器悲剧的没有NVIDIA 的显卡。
不过没有关系,还是可以跑起来的。
基本软件要求
- CUDA library version 6.5 (recommended), 6.0, 5.5, or 5.0 and the latest driver version for CUDA 6 or 319.* for CUDA 5 (and NOT 331.*)
- BLAS (provided via ATLAS, MKL, or OpenBLAS).
- OpenCV (>= 2.4)
- Boost (>= 1.55, although only 1.55 and 1.56 are tested)
glog
,gflags
,protobuf
,leveldb
,snappy
,hdf5
,lmdb
- For the Python wrapper
Python 2.7
,numpy (>= 1.7)
, boost-providedboost.python
2 CUDA
从http://developer.download.nvidia.com/compute/cuda/6_5/rel/installers/cuda_6.5.14_linux_64.run 下载cuda
sudo ./cuda_6.5.14_linux_64.run
不要安装驱动,否则安装会出错(driver installation is unable to locate the kernel source)
第一个选项选No
Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 340.29? ((y)es/(n)o/(q)uit): n
Install the CUDA 6.5 Toolkit? ((y)es/(n)o/(q)uit): y
Enter Toolkit Location [ default is /usr/local/cuda-6.5 ]:
Do you want to install a symbolic link at /usr/local/cuda? ((y)es/(n)o/(q)uit): y
Install the CUDA 6.5 Samples? ((y)es/(n)o/(q)uit): y
Enter CUDA Samples Location [ default is /home/root1 ]:
Installing the CUDA Toolkit in /usr/local/cuda-6.5 ...
Missing recommended library: libGLU.so
Missing recommended library: libXmu.so
Installing the CUDA Samples in /home/root1 ...
Copying samples to /home/root1/NVIDIA_CUDA-6.5_Samples now...
Finished copying samples.
===========
= Summary =
===========
Driver: Not Selected
Toolkit: Installed in /usr/local/cuda-6.5
Samples: Installed in /home/root1, but missing recommended libraries
Please make sure that
- PATH includes /usr/local/cuda-6.5/bin
- LD_LIBRARY_PATH includes /usr/local/cuda-6.5/lib64, or, add /usr/local/cuda-6.5/lib64 to /etc/ld.so.conf and run ldconfig as root
然后修改$PATH ,$LD_LIBRARY_PATH
vi /etc/profile
#在文档最后,添加:
export PATH="/usr/local/cuda-6.5/bin:$PATH"
#保存,退出,然后运行:
#source /etc/profile
vi ~/.bash_profile ~/目录下打开.bash_profile
$ export LD_LIBRARY_PATH=/usr/local/cuda-6.5/lib64:$LD_LIBRARY_PATH
#source ~/.bash_profile
3 安装一系列依赖项
sudo yum install boost-devel
sudo yum install opencv-devel
yum -y install freeglut
yum install atlas-devel
yum install snappy-devel
glog, gflags, protobuf, leveldb, snappy, hdf5, lmdb
glog: 记录日志信息的c++库,编程使用方式与c++的stream操作类似
wget https://google-glog.googlecode.com/files/glog-0.3.3.tar.gz
tar zxf glog-0.3.3.tar.gz
cd glog-0.3.3
./configure
make
sudo make install
其他人有使用安装相对应的依赖库
sudo yum install gflags-devel protobuf-devel leveldb-devel snappy-devel hdf5-devel
但是我安装的时候 还是很多失败了。所以列一下 各个的安装方法
gflags: 一套命令行参数解析工具
tar zxvf gflags-2.1.1.tar.gz && cd gflags-2.1.1
mkdir build && cd build
export CXXFLAGS="-fPIC" && cmake .. && make VERBOSE=1
make
sudo make install
protobuf:序列化协议, Google 公司内部的混合语言数据标准
第一次错误的方法:
tar zxvf protobuf-cpp-3.0.0-alpha-1.tar.gz && cd protobuf-cpp-3.0.0-alpha-1&& ./configure && make
make install
第二次正确的方法:
wget https://github.com/google/protobuf/releases/download/v3.0.0-alpha-2/protobuf-python-3.0.0-alpha-2.tar.gz
不过当时没有下载下来,所以我从windows机器上下载了,然后用rz -b上传的。 之前不带-b参数,导致一直解压失败。
然后进去按照readme里面的步骤
#./configure
#make
#make check
#make install
记得要去python文件夹内 安装python所需要的模块
#cd ./python
#python setup.py build
#python setup.py test
#python setup.py install
安装完成,验证Linux命令
#protoc –version
验证Python模块是否被正确安装
#python
>>>import google.protobuf
如果没有报错,说明安装正常。
leveldb:google开源的一个key-value存储引擎库
git clone https://github.com/google/leveldb.git
make
sudo cp -r include/leveldb /usr/local/include
sudo cp libleveldb.so /usr/local/lib
lmdb: mmap'd database for OpenLDAP aka Lightning MDB
git clone git://gitorious.org/mdb/mdb.git
cd mdb/libraries/liblmdb
make && make install
hdf5 : Hierarchical Data Format 5 rpm -i hdf5-1.8.10-2.1.1.src.rpm
tar -xjf hdf5-1.8.8.tar.bz2
./configure
make && make install
ccmakewget http://www.cmake.org/files/v2.8/cmake-2.8.10.2.tar.gz
tar -zxvf cmake-2.8.10.2.tar.gz
cd cmake-2.8.10.2
./bootstrap
4 OPENCV
Caffe作者默认你已经配置好了OpenCV环境,文档里没有说这一步。好在有人已经写好了配置OpenCV的脚本, https://github.com/jayrambhia/Install-OpenCV ,直接拿来用。
(这一步我没有用到)如果脚本运行失败,则详细阅读 opencv_install.sh 的代码,然后手工敲入命令进行安装。
git clone https://github.com/jayrambhia/Install-OpenCV
cd Install-OpenCV/RedHat
sudo ./opencv_latest.sh
(这一步我没有用到)如果脚本运行失败,则详细阅读 opencv_install.sh 的代码,然后手工敲入命令进行安装。
cd opencv-2.4.7
sed -i '/string(MD5/d' cmake/cl2cpp.cmake
mkdir build
cd build
cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local ..
make -j 4
sudo make install
sudo sh -c 'echo "/usr/local/lib" > /etc/ld.so.conf.d/opencv.conf'
sudo ldconfig
/usr/local/cuda-6.5/samples/bin/x86_64/linux/release/Install-OpenCV/OpenCV/opencv-2.4.10/build/lib
5 安装caffe git clone https://github.com/BVLC/caffe.git
make all
看上去是不是应该就可以看到美好结局了。。。。。实际上不是的。。。
1 error while loading shared libraries: libprotobuf.so.10: cannot open shared object file: No such file or directory
解决方法: 将protobuf 库函数所在路径加入LD_LIBRARY_PATH
2 /usr/bin/ld: cannot find -lhdf5_hl
解决方法:尝试了修改各种路径无效后,采取建立软链接
解决方法:尝试了修改各种路径无效后,采取建立软链接
cd /etc/ld.so.conf.d
sudo ln -sv /usr/local/weiwei/hdf5-1.8.8/hdf5/lib/libhdf5_hl.so /usr/lib/libhdf5_hl.so
sudo ln -sv /usr/local/weiwei/hdf5-1.8.8/hdf5/lib/libhdf5_hl.so /usr/lib/libhdf5_hl.so
并修改# vi Makefile.config 加入所要使用到的lib路径
3/usr/bin/ld: cannot find -lcblas
BLAS CBLAS
vi Makefile.in 修改库路径
/usr/local/BLAS
/usr/local/CBLAS/lib
sudo ln -sv /usr/lib64/atlas/libatlas.a /usr/lib/libatlas.a
sudo ln -sv /usr/lib64/atlas/libcblas.a /usr/lib/libcblas.a
vi Makefile.in 修改库路径
/usr/local/BLAS
/usr/local/CBLAS/lib
sudo ln -sv /usr/lib64/atlas/libatlas.a /usr/lib/libatlas.a
sudo ln -sv /usr/lib64/atlas/libcblas.a /usr/lib/libcblas.a
4 boost库里面 boost_thread调用一直有问题
首先重装了boost库
cd cd boost_1_57_0
./bootstrap.sh
./b2
./b2 install
echo "/usr/local/lib" >> /etc/ld.so.conf
ldconfig
看到成功提示
The Boost C++ Libraries were successfully built!
The Boost C++ Libraries were successfully built!
The following directory should be added to compiler include paths:
/usr/local/boost_1_57_0
The following directory should be added to linker library paths:
/usr/local/boost_1_57_0/stage/lib
但是还是一直有问题。一直报错 undefined reference to `boost::thread::join()'
但是在boost库文件夹内建立的test文件编译运行都没问题。
最后定位到之前第一次编译caffe 使用的boost版本有问题,刚刚重装boost库,于是caffe定位到的库和最后编译的版本不一致
在caffe目录下make clean后 重新编译,ok
但是在boost库文件夹内建立的test文件编译运行都没问题。
最后定位到之前第一次编译caffe 使用的boost版本有问题,刚刚重装boost库,于是caffe定位到的库和最后编译的版本不一致
在caffe目录下make clean后 重新编译,ok
PS: BLAS库的选择
Basic Linear Algebra Subprograms (BLAS)用来优化矢量内积、矩阵矢量乘积、矩阵乘积等运算。在Intel处理器上运行的的BLAS库有两种:MKL(The Intel Math Kernel Library)和Automatically Tuned Linear Algebra Software(ATLAS)。
ATLAS* (Automatically Tuned Linear Algebra Software) 是比较流行的线性代数的计算函数库。它实现了绝大多数的BLAS 与LAPACK 的函数。ATLAS 的网站是: http://math-atlas.sourceforge.net/.*
Intel MKL库中有些函数使用了多线程并行优化, 用户可用环境变量OMP_NUM_THREADS来控制MKL运行时的线程数。在双核和多核处理器上,它能够为DGEMM函数带来2-5倍的性能提升。(通用矩阵计算(DGEMM )是一个广为使用的数学函数。) 1)相比于ATLAS 函数库, Intel MKL 能够有明显性能提 2)Intel MKL BLAS 函数提供性能优越的多线程的支持
2) ScaLAPACK 是集群上求解线性方程的标准程序包。它是线性代数程序包(LAPACK)函数的一个子集。在64 位英特尔至强处理器的 32 节点集群上, 1)Intel MKL ScaLAPACK 性能明显超过 NETLIB ScaLAPACK。 2). 与使用 ATLAS* BLAS 与 NETLIB ScaLAPACK 代码相比,英特尔 MKL 的性能更为突出。
3). Intel MKL傅立叶变换函数已针对中等和大型问题进行高度优化。将Intel MKL 的性能与广为使用的FFT计算函数FFTW的性能进行了比较,表明:1)Intel MKL提供优异的傅立叶变换的性能。2) Intel MKL 为多处理器系统提供出色的可伸缩性。 3)在中等和大型转换方面,Intel MKL 比 FFTW 更出色。
MKL库中除了快速BLAS库外还包括LAPACK(Linear Algebra Package), FFT等。ATLAS有单线程C BLAS库、单线程Fortran BLAS库、多线程C BLAS库、多线程Fortran BLAS库等几种;单线程与多线程库的区别在于使用单线程库时, ATLAS函数内部是串行的(相当于MKL中OMP_NUM_THREADS=1), 而使用多线程库时, ATLAS的部分函数中使用了多线程SMP并行(相当于MKL中OMP_NUM_THREADS=2)。
综合比较,MKL库比ATLAS库性能稍好一点
ATLAS* (Automatically Tuned Linear Algebra Software) 是比较流行的线性代数的计算函数库。它实现了绝大多数的BLAS 与LAPACK 的函数。ATLAS 的网站是: http://math-atlas.sourceforge.net/.*
Intel MKL库中有些函数使用了多线程并行优化, 用户可用环境变量OMP_NUM_THREADS来控制MKL运行时的线程数。在双核和多核处理器上,它能够为DGEMM函数带来2-5倍的性能提升。(通用矩阵计算(DGEMM )是一个广为使用的数学函数。) 1)相比于ATLAS 函数库, Intel MKL 能够有明显性能提 2)Intel MKL BLAS 函数提供性能优越的多线程的支持
2) ScaLAPACK 是集群上求解线性方程的标准程序包。它是线性代数程序包(LAPACK)函数的一个子集。在64 位英特尔至强处理器的 32 节点集群上, 1)Intel MKL ScaLAPACK 性能明显超过 NETLIB ScaLAPACK。 2). 与使用 ATLAS* BLAS 与 NETLIB ScaLAPACK 代码相比,英特尔 MKL 的性能更为突出。
3). Intel MKL傅立叶变换函数已针对中等和大型问题进行高度优化。将Intel MKL 的性能与广为使用的FFT计算函数FFTW的性能进行了比较,表明:1)Intel MKL提供优异的傅立叶变换的性能。2) Intel MKL 为多处理器系统提供出色的可伸缩性。 3)在中等和大型转换方面,Intel MKL 比 FFTW 更出色。
MKL库中除了快速BLAS库外还包括LAPACK(Linear Algebra Package), FFT等。ATLAS有单线程C BLAS库、单线程Fortran BLAS库、多线程C BLAS库、多线程Fortran BLAS库等几种;单线程与多线程库的区别在于使用单线程库时, ATLAS函数内部是串行的(相当于MKL中OMP_NUM_THREADS=1), 而使用多线程库时, ATLAS的部分函数中使用了多线程SMP并行(相当于MKL中OMP_NUM_THREADS=2)。
综合比较,MKL库比ATLAS库性能稍好一点
MKL的申请路径(非商业版)
https://registrationcenter.intel.com/RegCenter/NComForm.aspx?ProductID=1461&pass=yes
caffe网络python接口编写程序失败及解决方法记录。
caffe 提供了python 接口和matlab接口,要使用python接口,需要在caffe网络执行完之后,再执行 make pycaffe
编译python文件时:
错误1 ImportError: No module named yaml
需要安装 PyYAML 3.11 http://pyyaml.org/download/pyyaml/PyYAML-3.11.tar.gz
wget http://pyyaml.org/download/pyyaml/PyYAML-3.11.tar.gz
tar -zxvf PyYAML-3.11.tar.gz
python setup.py install
错误2 ImportError: No module named argparse
easy_install argparse
wget http://pyyaml.org/download/pyyaml/PyYAML-3.11.tar.gz
tar -zxvf PyYAML-3.11.tar.gz
python setup.py install
错误2 ImportError: No module named argparse
easy_install argparse
输出为
Searching for argparse
Reading http://pypi.python.org/simple/argparse/
Best match: argparse 1.3.0
Downloading https://pypi.python.org/packages/source/a/argparse/argparse-1.3.0.tar.gz#md5=9bcf7f612190885c8c85e30ba41db3c7
Processing argparse-1.3.0.tar.gz
Running argparse-1.3.0/setup.py -q bdist_egg --dist-dir /tmp/easy_install-AVn40q/argparse-1.3.0/egg-dist-tmp-O6YdRA
warning: no previously-included files matching '*.pyc' found anywhere in distribution
warning: no previously-included files matching '*.pyo' found anywhere in distribution
warning: no previously-included files matching '*.orig' found anywhere in distribution
warning: no previously-included files matching '*.rej' found anywhere in distribution
no previously-included directories found matching 'doc/_build'
no previously-included directories found matching 'env24'
no previously-included directories found matching 'env25'
no previously-included directories found matching 'env26'
no previously-included directories found matching 'env27'
zip_safe flag not set; analyzing archive contents...
Adding argparse 1.3.0 to easy-install.pth file
Installed /usr/lib/python2.6/site-packages/argparse-1.3.0-py2.6.egg
Processing dependencies for argparse
Finished processing dependencies for argparse
Searching for argparse
Reading http://pypi.python.org/simple/argparse/
Best match: argparse 1.3.0
Downloading https://pypi.python.org/packages/source/a/argparse/argparse-1.3.0.tar.gz#md5=9bcf7f612190885c8c85e30ba41db3c7
Processing argparse-1.3.0.tar.gz
Running argparse-1.3.0/setup.py -q bdist_egg --dist-dir /tmp/easy_install-AVn40q/argparse-1.3.0/egg-dist-tmp-O6YdRA
warning: no previously-included files matching '*.pyc' found anywhere in distribution
warning: no previously-included files matching '*.pyo' found anywhere in distribution
warning: no previously-included files matching '*.orig' found anywhere in distribution
warning: no previously-included files matching '*.rej' found anywhere in distribution
no previously-included directories found matching 'doc/_build'
no previously-included directories found matching 'env24'
no previously-included directories found matching 'env25'
no previously-included directories found matching 'env26'
no previously-included directories found matching 'env27'
zip_safe flag not set; analyzing archive contents...
Adding argparse 1.3.0 to easy-install.pth file
Installed /usr/lib/python2.6/site-packages/argparse-1.3.0-py2.6.egg
Processing dependencies for argparse
Finished processing dependencies for argparse
3 Python-2.7.8/Include/pyport.h:886:2: error: #error "LONG_BIT definition appears wrong for platform (bad gcc/glibc config?)."
In file included from /usr/include/numpy/ndarrayobject.h:68,
from /usr/include/numpy/arrayobject.h:14,
from python/caffe/_caffe.hpp:8,
from python/caffe/_caffe.cpp:12:
/usr/include/numpy/npy_common.h:79:2: error: #error Must use Python with unicode enabled.
make: *** [python/caffe/_caffe.so] Error 1
In file included from /usr/include/numpy/ndarrayobject.h:68,
from /usr/include/numpy/arrayobject.h:14,
from python/caffe/_caffe.hpp:8,
from python/caffe/_caffe.cpp:12:
/usr/include/numpy/npy_common.h:79:2: error: #error Must use Python with unicode enabled.
make: *** [python/caffe/_caffe.so] Error 1
发现有Py_UNICODE_SIZE 的问题,提示 undefined symbol: PyUnicodeUCS4_AsEnco 。还有 error while loading shared libraries: libpython2.7.so.1.0:
原因是交叉编译python: pyconfig.h 中定义的 #define Py_UNICODE_SIZE 2
而交叉编译其他 python 库,比如 dbus-python, 默认的头文件路径包含: /usr/include/python2.6/pyconfig.h 定义的是#define Py_UNICODE_SIZE 4
所以会调用 UCS4 的函数,但是交叉版本的python里定义的是 UCS2 的函数
所以解决方法是 重新编译一次python
./configure --prefix=/usr/local/weiwei/python --enable-shared -enable-unicode=ucs4
make
make install
所以会调用 UCS4 的函数,但是交叉版本的python里定义的是 UCS2 的函数
所以解决方法是 重新编译一次python
./configure --prefix=/usr/local/weiwei/python --enable-shared -enable-unicode=ucs4
make
make install
编译完之后,要对所有用到python的地方进行修改,还要做对应的软链接
ln -sf /usr/local/weiwei/python/bin/python /usr/bin/python
查到他的记录:
1. 安装pycaffe必须的一些依赖项:
$ sudo apt-get install python-numpy python-scipy python-matplotlib python-sklearn python-skimage python-h5py python-protobuf python-leveldb python-networkx python-nose python-pandas
python-gflags Cython ipython
$ sudo apt-get install protobuf-c-compiler protobuf-compiler
$ sudo apt-get install python-numpy python-scipy python-matplotlib python-sklearn python-skimage python-h5py python-protobuf python-leveldb python-networkx python-nose python-pandas
python-gflags Cython ipython
$ sudo apt-get install protobuf-c-compiler protobuf-compiler
但是centos用不上这些指令,用yum也有很多是无法装。首先先给python装easy_istall
然后根据需要装
easy_install numpy
easy_install networkx
easy_install networkx
4 出现import caffe 失败, 直接没有输出,采用killall -9 python 来结束进程。
采用 gdb python 来进行调试。看到如下错误提示:
GNU gdb (GDB) Red Hat Enterprise Linux (7.2-60.el6)
Copyright (C) 2010 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law. Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-redhat-linux-gnu".
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>...
Reading symbols from /usr/local/weiwei/python/bin/python...done.
(gdb) set args a.py
(gdb) r
Copyright (C) 2010 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law. Type "show copying"
and "show warranty" for details.
This GDB was configured as "x86_64-redhat-linux-gnu".
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>...
Reading symbols from /usr/local/weiwei/python/bin/python...done.
(gdb) set args a.py
(gdb) r
[Thread debugging using libthread_db enabled]
Traceback (most recent call last):
File "a.py", line 17, in <module>
import caffe
File "../python/caffe/__init__.py", line 1, in <module>
from .pycaffe import Net, SGDSolver
File "../python/caffe/pycaffe.py", line 13, in <module>
import caffe.io
File "../python/caffe/io.py", line 2, in <module>
import skimage.io
ImportError: No module named skimage.io
skimage没有安装,查到官网上的指令如下
make pycaffe
easy_install -U scikit-image
但是依赖PIL,于是先装PIL easy_install PIL
发现PIL装完之后,显示安装成功,但是编译程序时还是一直报错,还是找不到。可能是因为机器上的python版本比较多,也不知道为啥一安装就说已经安装,但是程序编译又找不到,所以只好再重新下源码尝试安装。
wget http://effbot.org/downloads/Imaging-1.1.7.tar.gz
python setup.py build_ext -i
然后重新安装skimage,下载skimage 后, pip install .
然后安装 easy_install Cython
4 安装完之后出现protobuf找不到的问题。
这个之前是安装过的。不过之前安装的是cpp版本,而且安装后没有装python里面对应python-protobuf 。
5 ImportError: No module named caffe
make pycaffe