在这个链接中(http://www.linuxidc.com/Linux/2014-10/107501.htm),我配置了cuda,有强大的GPU,自然不能暴殄天物,让资源白白空闲着,所以配置一下卷积神经网络跑一下程序喽。至于卷积神经网络的原理,容后再写。打算先写库的用法,再写原理,以行动带动对理论的追求。
话不多说,步入正题。
1. 预说明
关于cuda-convnet,起源于一篇经典论文①,论文中针对ILSVRC-2010的数据进行实验,然后公布了其实验使用的代码,链接为②。但是,事实往往跟论文是有差距的,链接②中的代码根本不能重现论文中的结果。在下不才,在使用这个链接的库很久之后才发现的,觉得很坑,希望后来者慎之。
之所以说它坑,首先,论文中提到特性中,multi-GPU和dropout就没有实现,而且也没有给出论文中8层卷积神经网络的配置文件。总之不能直接拿来用,需要自己探索。
虽然如此,但有总比没有好,毕竟这个库实现的卷积神经网络封装的很好,论文中的大神的贡献非我等小菜所能企及的。给大神点32个赞。
本文只对cuda-convnet和cuda-convnet2的配置进行说明,论文中的作者还公布了其他版本的库,尚未用到,故且按下不提。
Ubuntu 12.04 下 CUDA 编程 http://www.linuxidc.com/linux/2014-06/103056.htm
Ubuntu 12.04 安装 CUDA-5.5 http://www.linuxidc.com/Linux/2013-10/91101.htm
Ubuntu 11.10 上安装CUDA开发环境 http://www.linuxidc.com/Linux/2012-04/58913.htm
Fedora 15系统下配置CUDA环境 http://www.linuxidc.com/Linux/2011-12/49874.htm
Ubuntu 11.04 安装 NVIDIA CUDA 4.0 RC2 http://www.linuxidc.com/Linux/2011-10/46304.htm
Linux Mint 13/Ubuntu 12.04 配置CUDA 4.2 & OpenCV 2.4.2 方法 http://www.linuxidc.com/Linux/2013-10/91102.htm
CUDA入门教程 http://www.linuxidc.com/Linux/2014-07/104328.htm
2. Cuda-convnet配置
2.1. 源码下载
参考链接②,先将源码下载下来。
- svn checkout http://cuda-convnet.googlecode.com/svn/trunk/ cuda-convnet-read-only
svn checkout http://cuda-convnet.googlecode.com/svn/trunk/ cuda-convnet-read-only
取出的版本是562。
2.2. 安装必要的库
然后,安装必须的库,我使用的是ubuntu系统。所以命令为
- sudo apt-get install Python-dev python-numpy python-magic python-matplotlib libatlas-base-dev
sudo apt-get install python-dev python-numpy python-magic python-matplotlib libatlas-base-dev
当然,还要确认你安装了cuda,我安装的是cuda6.5,在/usr/local/目录下,如下所示:
- $ ls /usr/local
- bin cuda cuda-6.5 etc games include lib man sbin share src
$ ls /usr/local bin cuda cuda-6.5 etc games include lib man sbin share src
2.3. 更改build.sh
进入到刚才下载的cuda-convnet-read-only目录,更改build.sh文件中的配置路径。如下所示:
- # CUDA toolkit installation directory.
- export CUDA_INSTALL_PATH=/usr/local/cuda
- # CUDA SDK installation directory.
- export CUDA_SDK_PATH=/usr/local/cuda-6.5/samples/common/inc
- # Python include directory. This should contain the file Python.h, among others.
- export PYTHON_INCLUDE_PATH=/usr/include/python2.7
- # Numpy include directory. This should contain the file arrayobject.h, among others.
- export NUMPY_INCLUDE_PATH=/usr/lib/python2.7/dist-packages/numpy/core/include/numpy
- # ATLAS library directory. This should contain the file libcblas.so, among others.
- export ATLAS_LIB_PATH=/usr/lib/atlas-base
- make $*
# CUDA toolkit installation directory. export CUDA_INSTALL_PATH=/usr/local/cuda # CUDA SDK installation directory. export CUDA_SDK_PATH=/usr/local/cuda-6.5/samples/common/inc # Python include directory. This should contain the file Python.h, among others. export PYTHON_INCLUDE_PATH=/usr/include/python2.7 # Numpy include directory. This should contain the file arrayobject.h, among others. export NUMPY_INCLUDE_PATH=/usr/lib/python2.7/dist-packages/numpy/core/include/numpy # ATLAS library directory. This should contain the file libcblas.so, among others. export ATLAS_LIB_PATH=/usr/lib/atlas-base make $*
按照官网的教程,配置完build.sh后就可以进行编译了。但是会发生错误,还需要改如下几个地方才可以。
2.4. 头文件添加
直接编译会发生找不到cutil_inline.h头文件的错误。分析原因可能是原来有这个头文件,后来这个头文件的功能被实现到其他头文件中去了。
在include子文件夹下田间cutil_inline.h文件,并输入内容。
- #include “helper_cuda.h”
- #define cutilCheckMsg(a) getLastCudaError(a)
- #define cutGetMaxGflopsDeviceId() gpuGetMaxGflopsDeviceId()
- #define MIN(a,b) (a) < (b) ? (a) : (b)
#include "helper_cuda.h" #define cutilCheckMsg(a) getLastCudaError(a) #define cutGetMaxGflopsDeviceId() gpuGetMaxGflopsDeviceId() #define MIN(a,b) (a) < (b) ? (a) : (b)
2.5. MakeFile文件更改
MakeFile第3行,原文如下:
- INCLUDES := -I$(PYTHON_INCLUDE_PATH) -I$(NUMPY_INCLUDE_PATH) -I./include -I./include/common -I./include/cudaconv2 -I./include/nvmatrix
INCLUDES := -I$(PYTHON_INCLUDE_PATH) -I$(NUMPY_INCLUDE_PATH) -I./include -I./include/common -I./include/cudaconv2 -I./include/nvmatrix
添加cuda的路径后如下:
- INCLUDES := -I$(PYTHON_INCLUDE_PATH) -I$(NUMPY_INCLUDE_PATH) -I$(CUDA_SDK_PATH) -I./include -I./include/common -I./include/cudaconv2 -I./include/nvmatrix
INCLUDES := -I$(PYTHON_INCLUDE_PATH) -I$(NUMPY_INCLUDE_PATH) -I$(CUDA_SDK_PATH) -I./include -I./include/common -I./include/cudaconv2 -I./include/nvmatrix
保存之。
2.6. 最后的库链接错误
做完上述改动后,可以编译了,但到最后会发生一个库链接的错误,不用管,直接将那个库注释掉。
在common-gcc-cuda-4.0.mk文件的332行。直接用#号注释。
- # LIB += -lcutil_$(LIB_ARCH) $(LIBSUFFIX) -lshrutil_$(LIB_ARCH) $(LIBSUFFIX)
# LIB += -lcutil_$(LIB_ARCH) $(LIBSUFFIX) -lshrutil_$(LIB_ARCH) $(LIBSUFFIX)
至此,就可以完成cuda-convnet的编译了。
更多详情见请继续阅读下一页的精彩内容: http://www.linuxidc.com/Linux/2014-10/107500p2.htm
3. Cuda-convnet2配置
顾名思义,这是cuda-convnet的2.0版本,支持多GPU运行。
3.1. 源码下载
git clone https://code.google.com/p/cuda-convnet2/
3.2. 必要的库
- sudo apt-get install Python-dev python-numpy python-scipy python-magic python-matplotlib libatlas-base-dev libjpeg-dev libopencv-dev
sudo apt-get install python-dev python-numpy python-scipy python-magic python-matplotlib libatlas-base-dev libjpeg-dev libopencv-dev
3.3. 配置
我只能说,这个版本的比上个版本人性化多了,这个版本的build.sh直接如此。
- # CUDA toolkit installation directory.
- export CUDA_INSTALL_PATH=/usr/local/cuda
- # Python include directory. This should contain the file Python.h, among others.
- export PYTHON_INCLUDE_PATH=/usr/include/python2.7
- # Numpy include directory. This should contain the file arrayobject.h, among others.
- export NUMPY_INCLUDE_PATH=/usr/lib/python2.7/dist-packages/numpy/core/include/numpy/
- # ATLAS library directory. This should contain the file libcblas.so, among others.
- export ATLAS_LIB_PATH=/usr/lib/atlas-base
- # You don’t have to change these:
- export LD_LIBRARY_PATH=$CUDA_INSTALL_PATH/lib64:$LD_LIBRARY_PATH
- export CUDA_SDK_PATH=$CUDA_INSTALL_PATH/samples
- export PATH=$PATH:$CUDA_INSTALL_PATH/bin
# CUDA toolkit installation directory. export CUDA_INSTALL_PATH=/usr/local/cuda # Python include directory. This should contain the file Python.h, among others. export PYTHON_INCLUDE_PATH=/usr/include/python2.7 # Numpy include directory. This should contain the file arrayobject.h, among others. export NUMPY_INCLUDE_PATH=/usr/lib/python2.7/dist-packages/numpy/core/include/numpy/ # ATLAS library directory. This should contain the file libcblas.so, among others. export ATLAS_LIB_PATH=/usr/lib/atlas-base # You don't have to change these: export LD_LIBRARY_PATH=$CUDA_INSTALL_PATH/lib64:$LD_LIBRARY_PATH export CUDA_SDK_PATH=$CUDA_INSTALL_PATH/samples export PATH=$PATH:$CUDA_INSTALL_PATH/bin
如果是在Ubuntu下的话,这样就直接已经基本把路径都配置对了。
3.4. 链接错误
但是直接编译的话还是会遇到错误,如下所示:
- cd ./bin/ && g++ -O3 -DNUMPY_INTERFACE -shared -Wl,-no-undefined -o libutilpy.so src/matrix.o -L/usr/lib/atlas-base -latlas -lcblas -lpython2.7
- /usr/bin/ld: cannot find -latlas
- /usr/bin/ld: cannot find -lcblas
- collect2: error: ld returned 1 exit status
cd ./bin/ && g++ -O3 -DNUMPY_INTERFACE -shared -Wl,-no-undefined -o libutilpy.so src/matrix.o -L/usr/lib/atlas-base -latlas -lcblas -lpython2.7 /usr/bin/ld: cannot find -latlas /usr/bin/ld: cannot find -lcblas collect2: error: ld returned 1 exit status
主要是因为atlas库中没有libatlas.so和libctlas.so。查看atlas的目录发现结构如此:
- :/usr/lib/atlas-base$ ls -l
- 总用量 4292
- drwxr-xr-x 2 root root 4096 9月 22 11:41 atlas
- lrwxrwxrwx 1 root root 15 2月 4 2014 libatlas.so.3 -> libatlas.so.3.0
- -rw-r–r– 1 root root 3746968 2月 4 2014 libatlas.so.3.0
- lrwxrwxrwx 1 root root 15 2月 4 2014 libcblas.so.3 -> libcblas.so.3.0
- -rw-r–r– 1 root root 135376 2月 4 2014 libcblas.so.3.0
- lrwxrwxrwx 1 root root 17 2月 4 2014 libf77blas.so.3 -> libf77blas.so.3.0
- -rw-r–r– 1 root root 131000 2月 4 2014 libf77blas.so.3.0
- lrwxrwxrwx 1 root root 22 2月 4 2014 liblapack_atlas.so.3 -> liblapack_atlas.so.3.0
- -rw-r–r– 1 root root 369472 2月 4 2014 liblapack_atlas.so.3.0
:/usr/lib/atlas-base$ ls -l 总用量 4292 drwxr-xr-x 2 root root 4096 9月 22 11:41 atlas lrwxrwxrwx 1 root root 15 2月 4 2014 libatlas.so.3 -> libatlas.so.3.0 -rw-r--r-- 1 root root 3746968 2月 4 2014 libatlas.so.3.0 lrwxrwxrwx 1 root root 15 2月 4 2014 libcblas.so.3 -> libcblas.so.3.0 -rw-r--r-- 1 root root 135376 2月 4 2014 libcblas.so.3.0 lrwxrwxrwx 1 root root 17 2月 4 2014 libf77blas.so.3 -> libf77blas.so.3.0 -rw-r--r-- 1 root root 131000 2月 4 2014 libf77blas.so.3.0 lrwxrwxrwx 1 root root 22 2月 4 2014 liblapack_atlas.so.3 -> liblapack_atlas.so.3.0 -rw-r--r-- 1 root root 369472 2月 4 2014 liblapack_atlas.so.3.0
添加两个软链接,执行命令:
- sudo ln -s libatlas.so.3.0 libatlas.so
- sudo ln -s libcblas.so.3.0 libcblas.so
sudo ln -s libatlas.so.3.0 libatlas.so sudo ln -s libcblas.so.3.0 libcblas.so
目录结构变为如此:
- /usr/lib/atlas-base$ ls -l
- 总用量 4292
- drwxr-xr-x 2 root root 4096 9月 22 11:41 atlas
- lrwxrwxrwx 1 root root 15 10月 1 22:35 libatlas.so -> libatlas.so.3.0
- lrwxrwxrwx 1 root root 15 2月 4 2014 libatlas.so.3 -> libatlas.so.3.0
- -rw-r–r– 1 root root 3746968 2月 4 2014 libatlas.so.3.0
- lrwxrwxrwx 1 root root 15 10月 1 22:36 libcblas.so -> libcblas.so.3.0
- lrwxrwxrwx 1 root root 15 2月 4 2014 libcblas.so.3 -> libcblas.so.3.0
- -rw-r–r– 1 root root 135376 2月 4 2014 libcblas.so.3.0
- lrwxrwxrwx 1 root root 17 2月 4 2014 libf77blas.so.3 -> libf77blas.so.3.0
- -rw-r–r– 1 root root 131000 2月 4 2014 libf77blas.so.3.0
- lrwxrwxrwx 1 root root 22 2月 4 2014 liblapack_atlas.so.3 -> liblapack_atlas.so.3.0
- -rw-r–r– 1 root root 369472 2月 4 2014 liblapack_atlas.so.3.0
/usr/lib/atlas-base$ ls -l 总用量 4292 drwxr-xr-x 2 root root 4096 9月 22 11:41 atlas lrwxrwxrwx 1 root root 15 10月 1 22:35 libatlas.so -> libatlas.so.3.0 lrwxrwxrwx 1 root root 15 2月 4 2014 libatlas.so.3 -> libatlas.so.3.0 -rw-r--r-- 1 root root 3746968 2月 4 2014 libatlas.so.3.0 lrwxrwxrwx 1 root root 15 10月 1 22:36 libcblas.so -> libcblas.so.3.0 lrwxrwxrwx 1 root root 15 2月 4 2014 libcblas.so.3 -> libcblas.so.3.0 -rw-r--r-- 1 root root 135376 2月 4 2014 libcblas.so.3.0 lrwxrwxrwx 1 root root 17 2月 4 2014 libf77blas.so.3 -> libf77blas.so.3.0 -rw-r--r-- 1 root root 131000 2月 4 2014 libf77blas.so.3.0 lrwxrwxrwx 1 root root 22 2月 4 2014 liblapack_atlas.so.3 -> liblapack_atlas.so.3.0 -rw-r--r-- 1 root root 369472 2月 4 2014 liblapack_atlas.so.3.0
然后就可以正常编译了。
参考文献
① ImageNet Classification with Deep Convolutional Neural Networks
② https://code.google.com/p/cuda-convnet
③ https://code.google.com/p/cuda-convnet2
更多Ubuntu相关信息见Ubuntu 专题页面 http://www.linuxidc.com/topicnews.aspx?tid=2
本文永久更新链接地址:http://www.linuxidc.com/Linux/2014-10/107500.htm