安装TensorFlow

anaconda

下载: python3.6版本

安装: sudo sh Anaconda3-5.2.0-Linux-x86_64.sh

tensorflow

从pypi下载已经编译好的tensorflow_gpu

https://pypi.org/project/tensorflow-gpu/#files

tensorflow_gpu依赖的python package, 一部分可以在python uci package下载, 一部分仍在在pypi下载.

tensorflow_gpu-1.10.1-cp36-cp36m-manylinux1_x86_64.whl
absl_py-0.4.0-py2.py3-none-any.whl
astor-0.7.1-py2.py3-none-any.whl
gast-0.2.0-py2.py3-none-any.whl
grpcio-1.14.1-cp36-cp36m-manylinux1_x86_64.whl
msgpack-0.5.6-cp36-cp36m-manylinux1_x86_64.whl
numpy-1.14.5-cp36-cp36m-manylinux1_x86_64.whl(版本限制!)
protobuf-3.6.1-cp36-cp36m-manylinux1_x86_64.whl
termcolor-1.1.0-py2.py3-none-any.whl
tensorboard-1.10.0-py3-none-any.whl
Markdown-2.6.11-py2.py3-none-any.whl

安装: pip install XXX.whl

安装Cuda

下载: cuda9.0
cuda 下载
安装: sudo sh cuda_9.0.176_384.81_linux.run
- 当提示安装openGL, 若为双显卡且主显为非NVIDIA GPU, 则选择no
- 安装cuda出现的问题. 因为cuda需要修改显卡驱动, 因此需要关闭图形界面.

# error:
It appears that an X server is running. 
Please exit X before installation. 
If you're sure that X is not running, 
but are getting this error, 
please delete any X lock files in /tmp.

# solution:
/etc/init.d/lightdm stop
# 然后重启
reboot
# refs: https://www.cnblogs.com/liyuanhong/articles/4919755.html
# 重启图形界面
sudo service lightdm start
或sudo lightdm restart

验证
- 输入ls /dev/nvidia*
  - 若显示/dev/nvidia0 /dev/nvidiactl /dev/nvidia-uvm, 则安装成功.
  - 若显示ls: cannot access/dev/nvidia*: No such file or directory, 或/dev/nvidia0 /dev/nvidiactl(安装成功中显示中的一个或两个), 则需修改启动文件
    - sudo vi /etc/rc.local
    - 把第一行的#!/bin/sh -e中的-e去掉
    - 然后在exit 0前面插入下面运行脚本
    - 输入脚本后, 运行source /etc/rc.local更新
    - 在运行ls /dev/nvidia*将会看到三个文件, 则安装成功

/sbin/modprobe nvidia

if [ "$?" -eq 0 ]; then
    # Count the number of NVIDIA controllers found.
    NVDEVS=`lspci | grep -i NVIDIA`
    N3D=`echo "$NVDEVS" | grep "3D controller" | wc -l`
    NVGA=`echo "$NVDEVS" | grep "VGA compatible controller" | wc -l`
    N=`expr $N3D + $NVGA - 1`
    for i in `seq 0 $N`; do
        mknod -m 666 /dev/nvidia$i c 195 $i
    done
    mknod -m 666 /dev/nvidiactl c 195 255
else
    exit 1
fi

/sbin/modprobe nvidia-uvm

if [ "$?" -eq 0 ]; then
    # Find out the major device number used by the nvidia-uvm driver
    D=`grep nvidia-uvm /proc/devices | awk '{print $1}'`
    mknod -m 666 /dev/nvidia-uvm c $D 0
else
    exit 1
fi

配置环境变量
- 运行sudo gedit /etc/profile , 在末尾添加下面文本, 然后运行source /etc/profile更新

export CUDA_HOME=/usr/local/cuda-9.0
export PATH=${PATH}:${CUDA_HOME}/bin
export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${CUDA_HOME}/lib64

安装完成后, 运行nvidia-smi进行测试

安装cudnn. 这里需要注册一个NVIDIA账户

下载: cuDNN7.2.1
cudnn 下载
- 安装(第一个就可以了): sudo dpkg -i libcudnn7_7.2.1.38-1+cuda9.0_amd64.deb
安装出现问题

# error:
/sbin/ldconfig.real: /usr/lib/nvidia-375/libEGL.so.1 is not a symbolic link

# solution:
https://askubuntu.com/questions/900285/libegl-so-1-is-not-a-symbolic-link
第二个脚本可用

安装完成后进行测试

运行python

import tensorflow
print(tensorflow.__version__)

运行问题

2018-08-31 16:11:56.214798: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-08-31 16:11:56.217297: E tensorflow/stream_executor/cuda/cuda_driver.cc:397] failed call to cuInit: CUDA_ERROR_NO_DEVICE
2018-08-31 16:11:56.217342: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:157] no NVIDIA GPU device is present: /dev/nvidia0 does not exist

运行nvidia-smi

Failed to initialize NVML: Driver/library version mismatch

原因分析:
- 安装cuda失败, 重新安装cuda, cuda的安装可以见cuda 安装.
- 安装的cuda版本不对, 则安装对应的cuda, 或者是cuda版本自动更新, 重启电脑解决.

参考

TensorFlow官方安装教程

https://www.tensorflow.org/install/install_linux?hl=zh-cn

cuda 安装

https://blog.csdn.net/QLULIBIN/article/details/78714596

cudnn 安装

http://www.itread01.com/content/1501166415.html

其他

代理上网

windows安装CCproxy
ubuntu设置环境变量

# proxy envir
MY_PROXY_URL=http://XXX.XX.XX.XX:808/
export ftp_proxy=${MY_PROXY_URL}
export http_proxy=${MY_PROXY_URL}
export https_proxy=${MY_PROXY_URL}

# proxy envir or
sudo apt-get -o Acquire::http::proxy="http://XXX.XX.XX.XX:808/" update