Ubuntu 16.04下1.5版TensorFlow-gpu升级记录

发表于 更新于 分类于 ，

使用pip升级TensorFlow时，发现TF已经升级至1.5版本。以下记录更新步骤 系统信息： 

1
2
Ubuntu 16.04 LTS x86_64
Python 3.5.4 :: Anaconda custom (64-bit)
# 1、更新TensorFlow 如果直接使用pip install -U升级，会出现futures的错误： 
1
2
3
Collecting futures>=3.1.1 (from tensorflow-tensorboard<1.6.0,>=1.5.0->tensorflow)
  Downloading https://pypi.tuna.tsinghua.edu.cn/packages/1f/9e/7b2ff7e965fc654592269f2906ade1c7d705f1bf25b7d469fa153f7d19eb/futures-3.2.0.tar.gz
Unknown requires Python '>=2.6, <3' but the running Python is 3.5.4
解决办法先安装futures的3.1.1版本，然后安装TensorFlow 1.5。注意不能使用'-U'参数更新： 
1
2
pip install futures==3.1.1
pip install tensorflow-gpu==1.5.0
安装完成执行以下测试语句，报错说找不到libcublas.so.9.0： 
1
2
import tensorflow as tf
print(tf.__version__)
错误： 
1
2
ImportError: libcublas.so.9.0: cannot open shared object file: 
             No such file or directory

2、更新cuda 9和cudnn 7

(1) 下载以下两个文件至本地：

cuda-repo-ubuntu1604_9.0.176-1_amd64.deb 7fa2af80.pub

(2) 执行以下两个命令：

1
2
sudo dpkg -i cuda-repo-ubuntu1604_9.0.176-1_amd64.deb
sudo apt-key add ./7fa2af80.pub

nvidia文档说明如下： 

1
2
sudo dpkg -i cuda-repo-<distro>_<version>_<architecture>.deb
sudo apt-key add /var/cuda-repo-<version>/7fa2af80.pub

(3) 给apt-get设置代理：

网址http://developer.download.nvidia.com/无法通过ipv6访问，设置代理： 

1
sudo vi /etc/apt/apt.conf
添加代理服务器配置后保存： 
1
2
Acquire::http::Proxy "http://127.0.0.1:8122";
Acquire::https::Proxy "http://127.0.0.1:8122";

(4) 执行更新动作

1
sudo apt-get update

(5) 列出需要更新的软件版本

1
sudo apt-cache policy cuda

命令格式： 

1
sudo apt-cache policy <package name>
输出： 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
cuda:
  已安装：8.0.61-1
  候选： 9.1.85-1
  版本列表：
     9.1.85-1 500
        500 http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64  Packages
     9.0.176-1 500
        500 http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64  Packages
        100 /var/lib/dpkg/status
 *** 8.0.61-1 500
        500 http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64  Packages
     8.0.44-1 500
        500 http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64  Packages

(6) 选取指定的版本

1
sudo apt-get update cuda=9.0.176-1

(7) 创建软链接并验证安装

创建软链接： 

1
2
cd /usr/local/
sudo ln -s cuda-9.0 cuda
安装完成之后，使用以下命令验证安装情况，注意，需要设定环境变量(可在~/.bashrc中添加)： 
1
2
3
4
export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64"
export CUDA_HOME=/usr/local/cuda

显示驱动基本情况，执行nvidia-smi，输出： 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
Sun Feb  4 11:36:36 2018
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 390.12                 Driver Version: 390.12                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 108...  Off  | 00000000:03:00.0 Off |                  N/A |
|  0%   19C    P5    26W / 250W |      0MiB / 11176MiB |      2%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-----------------------------------------------------------------------------+
执行nvcc -V，输出： 
1
2
3
4
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Sep__1_21:08:03_CDT_2017
Cuda compilation tools, release 9.0, V9.0.176
## (8) 安装cudnn 然后进入选定目录，执行以下命令： 
1
2
3
4
5
6
wget -c http://developer.download.nvidia.com/compute/machine-learning/cudnn/secure/v7.0.5/prod/9.0_20171129/cudnn-9.0-linux-x64-v7.tgz
tar -zxvf cudnn-9.0-linux-x64-v7.tgz
sudo cp cuda/include/cudnn.h /usr/local/cuda/include/
sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64/ -d
sudo chmod a+r /usr/local/cuda/include/cudnn.h
sudo chmod a+r /usr/local/cuda/lib64/libcudnn*

(9) 更新完成后，恢复原有的配置

去除apt-get代理: 

1
sudo mv /etc/apt/apt.conf /etc/apt/apt.conf.with_proxy
去除nvidia更新源，将cuda.list的内容注释掉： 
1
sudo vi /etc/apt/sources.list.d/cuda.list

(10) 如果以后需要更新，则将第(9)步的内容恢复即可。

参考文档

[1] Failed install on Windows
[2] nvidia文档
[3] Configure proxy for APT?
[4] How to install specific version of some package
[5] 深度学习服务器环境配置