[关闭]
@rg070836rg 2017-10-20T10:15:52.000000Z 字数 4627 阅读 2646

安装cuda8+cudnn5.1+tensorflowgpu+keras

毕业设计 GPU并行计算课程实验与报告

0.由于对centos不熟悉，以及超算中心的centos版本比较低，实在没法装有些库，于是先换成了ubuntu16.04server。

一 创建用户

1.1 创建用户
adduser dluser01
passwd xxxxxxxx
dluser01~10

1.2 增加root权限
vim /etc/sudoers
## Allow root to run any commands anywhere
root ALL=(ALL) ALL
ubuntu ALL=(ALL) ALL
dluser01 ALL=(ALL) ALL
dluser02 ALL=(ALL) ALL

二 修改源

参见 http://mirrors.ustc.edu.cn/help/ubuntu.html

三 安装python2.7

换成ubuntu16.04后自带

四 安装pip

4.0 网速够，ubuntu16.04下 直接sudo apt-get install python-pip python-dev

4.1 安装easyinstall

  1. wget -q http://peak.telecommunity.com/dist/ez_setup.py
  2. python ez_setup.py

4.2 编译安装python
下载 https://github.com/pypa/pip/releases

  1. tar zvxf pip-9.0.1.tar.gz    #解压文件
  2. cd pip-9.0.1/
  3. python setup.py install

4.3 修改pip源（阿里源）

  1. cd ~
  2. mkdir .pip
  3. vim ~/.pip/pip.conf
  5. [global] 
  6. index-url = http://mirrors.aliyun.com/pypi/simple/
  7. [install] 
  8. trusted-host=mirrors.aliyun.com

五 安装NVIDIA驱动

5.1 查找对应驱动
下载并传至服务器，进入root

  1. sudo init 3
  2. sudo sh NVIDIA-Linux-x86_64-375.39.run
  3. sudo reboot

装好了用nvidia-smi，检查一下：
1490615001(1).jpg-20.6kB

六 安装cuda

6.1 下载
下载地址
image_1bc7pgehh1u1kv99iidh5518mel.png-65.7kB
下载runfile。。

6.2安装

  1. sudo sh xxxxx.run
  2. 刷屏漫长的EULA条文，接下来这么选：
  3. accept/decline/quit: accept
  5. Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 367.48?
  6. (y)es/(n)o/(q)uit: y
  8. Do you want to install the OpenGL libraries?
  9. (y)es/(n)o/(q)uit [ default is yes ]: y
  11. Do you want to run nvidia-xconfig?
  12. This will update the system X configuration file so that the NVIDIA X driver
  13. is used. The pre-existing X configuration file will be backed up.
  14. This option should not be used on systems that require a custom
  15. X configuration, such as systems with multiple GPU vendors.
  16. (y)es/(n)o/(q)uit [ default is no ]: n
  18. Install the CUDA 8.0 Toolkit?
  19. (y)es/(n)o/(q)uit: y
  21. Enter Toolkit Location
  22.  [ default is /usr/local/cuda-8.0 ]: 
  24. Do you want to install a symbolic link at /usr/local/cuda?
  25. (y)es/(n)o/(q)uit: y
  27. Install the CUDA 8.0 Samples?
  28. (y)es/(n)o/(q)uit: y
  30. Enter CUDA Samples Location
  31.  [ default is /home/ubuntu ]: 
  33. Installing the NVIDIA display driver...
  34. Installing the CUDA Toolkit in /usr/local/cuda-8.0 ...
  35. Missing recommended library: libGLU.so
  36. Missing recommended library: libX11.so
  37. Missing recommended library: libXi.so
  38. Missing recommended library: libXmu.so
  40. Installing the CUDA Samples in /home/ubuntu ...
  41. Copying samples to /home/ubuntu/NVIDIA_CUDA-8.0_Samples now...
  42. Finished copying samples.
  44. ===========
  45. = Summary =
  46. ===========
  48. Driver:   Installed
  49. Toolkit:  Installed in /usr/local/cuda-8.0
  50. Samples:  Installed in /home/ubuntu, but missing recommended libraries
  52. Please make sure that
  53.  -   PATH includes /usr/local/cuda-8.0/bin
  54.  -   LD_LIBRARY_PATH includes /usr/local/cuda-8.0/lib64, or, add /usr/local/cuda-8.0/lib64 to /etc/ld.so.conf and run ldconfig as root
  56. To uninstall the CUDA Toolkit, run the uninstall script in /usr/local/cuda-8.0/bin
  57. To uninstall the NVIDIA Driver, run nvidia-uninstall
  59. Please see CUDA_Installation_Guide_Linux.pdf in /usr/local/cuda-8.0/doc/pdf for detailed information on setting up CUDA.

6.3 配置环境变量（当前用户）

  1. sudo vim ~/.bashrc
  3. 最后加入
  4. export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/cuda-8.0/lib64:/usr/local/cuda-8.0/extras/CUPTI/lib64"
  5. export CUDA_HOME=/usr/local/cuda-8.0
  7. source ~/.bashrc   刷新文件

七 安装cudnn

与8匹配的是cudnn5.1，下载地址
首先需要注册,填一个问卷。
然后下载这个cuDNN v5.1 Runtime Library for Ubuntu14.04 (Deb)
16.06的那个不是amd64平台的。。下载14.04的

  1. sudo dpkg -i libcudnn5_5.1.10-1+cuda8.0_amd64.deb

八 安装tensorflow gpu

为了保证稳定，不在root配置tensorflow，转而在各个用户下配置，所以需要每个用户配置下pip源(参照上文)，配置好之后，执行

  1. pip install tensorflow-gpu

注意，环境变量也是随着用户的，所以每增加一个用户，需要重新配一下这个用户的环境变量，打开python测试一下：

  1. dluser02@ubuntu:~$ python
  2. Python 2.7.12 (default, Nov 19 2016, 06:48:10) 
  3. [GCC 5.4.0 20160609] on linux2
  4. Type "help", "copyright", "credits" or "license" for more information.
  5. >>> import tensorflow
  6. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcublas.so.8.0 locally
  7. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcudnn.so.5 locally
  8. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcufft.so.8.0 locally
  9. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcuda.so.1 locally
  10. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcurand.so.8.0 locally
  11. >>>

注意，安装版本过低，建议按照官网推荐的方法，找到gpu字样

  1. sudo pip install https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow_gpu-1.0.1-cp27-none-linux_x86_64.whl

九 安装keras

装好前面的前提下，直接pip install keras,等待安装好即可,测试如下：

  1. dluser02@ubuntu:~$ python
  2. Python 2.7.12 (default, Nov 19 2016, 06:48:10) 
  3. [GCC 5.4.0 20160609] on linux2
  4. Type "help", "copyright", "credits" or "license" for more information.
  5. >>> import keras
  6. Using TensorFlow backend.
  7. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcublas.so.8.0 locally
  8. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcudnn.so.5 locally
  9. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcufft.so.8.0 locally
  10. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcuda.so.1 locally
  11. I tensorflow/stream_executor/dso_loader.cc:135] successfully opened CUDA library libcurand.so.8.0 locally
  12. >>>

注意如果版本过低，去github上面下载源码安装

添加新批注
在作者公开此批注前，只有你和作者可见。
回复批注