安装TensorFlow(Ubuntu+CUDA+Cudnn)

0. 安装说明

由于之前安装caffe 的时候使用的cuda7.0,并不想破坏它，同时又不想使用过低版本的tensorflow，所以才用了Virtualenv installation 方式安装。

1. 本机配置：

系统：Ubuntu14.04 
显卡：GTX1080 
CPU：i7 6500K，其他不重要了

2. 安装的版本：

tensorflow r0.10
cuda 7.5
cudnn v5.1(我安装的5.0会出现问题)

3. 参考网址

官网（r0.10,使用Virtualenv installation 安装）：
https://www.tensorflow.org/versions/r0.10/get_started/os_setup.html#requirements
cuda: (需要选对版本)
https://developer.nvidia.com/cuda-downloads
cudnn:(需要注册)
https://developer.nvidia.com/rdp/cudnn-download

4. 安装参考网址

http://www.pyimagesearch.com/2016/07/04/how-to-install-cuda-toolkit-and-cudnn-for-deep-learning/

5. 安装流程

5.1 安装1080显卡驱动

1: Download the Nvidia driver in right version (NVIDIA GeForce GTX1080) 
2: Edit the blacklist.conf: 
sudo gedit /etc/modprobe.d/blacklist.conf 
add :

blacklist vga16fb
blacklist nouveau
blacklist rivafb
blacklist nvidiafb
blacklist rivatv
__block_line__
3: sudo apt-get remove –purge nvidia-* 
**sudo apt-get remove –purge xserver-xorg-video-nouveau** 
4: reboot and into computer using ‘Ctrl+ALT+F1’ using the ‘Ctrl+ALT+F7’ return xservers 
5: sudo /etc/init.d/gdm stop or sudo /etc/init.d/lightdm stop 
6: in the driver file,which downloaded in step 1 
sudo sh NVIDIA.run 
7: sudo /etc/init.d/gdm restart or sudo /etc/init.d/lightdm restart 
8: sudo reboot 

5.2 安装cuda

1: Download the cuda 7.5 from the cuda_driver 
   IMPORTANCE:download .runfile 
2: reboot and ctrl+alt+F1 
   sudo /etc/init.d/lightdm stop 
3: sudo sh cuda.run 
   IMPORTANCE:ALL SHOULD INSTALL EXCEPT:DRIVER(because we have installed the nvidia driver and the cuda.runfile don’t include the nvidia driver) 
4: sudo /etc/init.d/lightdm restart 
5: reboot 

5.3 安装cudnn

 tar -zxf cudnn-7.5-linux-x64-v5.1-ga.tgz
 cd cuda
 sudo cp lib64/* /usr/local/cuda/lib64/
 sudo cp include/* /usr/local/cuda/include/

5.4 添加环境变量

gedit .bashrc

export CUDA_HOME=/usr/local/cuda
export PATH="/usr/local/cuda/bin:/usr/local/lib/:${JAVA_HOME}/bin:${PATH}"
export LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64:${LD_LIBRARY_PATH}"

source .bashrc

5.5 Install tensorflow

sudo apt-get install python-pip python-dev python-virtualenv
virtualenv --system-site-packages ~/tensorflow
source ~/tensorflow/bin/activate
# Requires CUDA toolkit 7.5 and CuDNN v5. For other versions, see "Install from sources" below.
(tensorflow)$ export TF_BINARY_URL=https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow-0.10.0-cp27-none-linux_x86_64.whl
(tensorflow)$ pip install --upgrade $TF_BINARY_URL

6 Important Issue

1 https://github.com/tensorflow/tensorflow/issues/4251
2 如果测试的时候无法导入cudnn，注意LD_LIBRARY_PATH 路径下是否存在可用的cudnn
3 测试通过需要看看 python -m tensorflow.models.image.mnist.convolutional，不能仅仅看import tensorflow as tf 同时不能用任何问题才能认为安装成功
4 更新一个刚刚解决的问题，在安装tensorflow上经常会出现numpy 版本冲突问题
参考：http://stackoverflow.com/questions/16391335/using-old-version-of-numpy-even-with-a-newer-version-installed

先要找到numpy的存在位置
ipython
import numpy
print numpy.__version__
print numpy.__file__

在我自己的配置下，因为是Virtualenv installation 方式安装，所以numpy 会在两个地方:

/usr/local/lib/python2.7/dist-packages/numpy/__init__.pyc
/home/dex_hand/tensorflow/lib/python2.7/site-packages/numpy/__init__.pyc

所以我通过这样的方式把两个地方的numpy 文件夹 都删除，
rm -rf numpy
最后通过pip install numpy 重新安装回来，这个时候在tensorflow这个虚拟环境下也能使用到这个版本（1.11.0）的版本的numpy了