Positive children usually take the CPU to run the eight Tensorflow it does not matter, until recently I ran a RNN model after a epoch s + 300 of the CPU so I really can not stand it, so I chose the bitter experience of the GPU to run the algorithm . But it is embarrassing, I use the agricultural enterprises of graphics, will not escape the mainstream CUDA, after two or three days after the configuration, and finally successfully configured the open OpenCL support Tensorflow, ran LSTM still very pleasant.

Install Ubuntu 16.04 LTS

Because the current agricultural enterprises has essentially fglrx does not support it, and for my graphics card Advanced Micro Devices, Inc. [AMD/ATI] Mars [Radeon HD 8670A/8670M/8750M]is, fglrx driver is always a bunch of questions, so I decisively chose this version 16.04. In this version, the open-source drivers are supported, such as Radeonwell AMD GPU Pro.

Ready to work

First you need to install cmakeas well as other tools to compile frequently required.

      
      
       
       # install build essentials
      
      
      
      
       
       sudo apt-get install cmake
      
      
      
      
       
       sudo apt-get update && sudo apt-get install build-essential

Next, you need to install the famous boostandlibssl

      
      
       
       sudo apt-get install libssl0.9.8:i386
      
      
      
      
       
       sudo apt-get install libboost-all-dev
      
      
      
      
      
      
       
       sudo apt-get install libgtest-dev
      
      
      
      
       
       cd /usr/src/gtest
      
      
      
      
       
       sudo cmake .
      
      
      
      
       
       sudo make
      
      
      
      
       
       sudo mv libg* /usr/lib/

安装 Python

建议使用anaconda来安装Python 3.5。

现在主流的包和库都支持Python 3并且有些库（例如Django）已经明确宣称不支持Python 2，同时，Python 3 对Unicode良好的兼容性对于开发有中文的数据来说，非常省心。

其实这里推荐anaconda的主要原因还是因为Theano。因为如果你想使用带有OpenCL支持的Theano，你就必须使用其后端libgpuarray。而这个包在issue里面明确对不使用conda安装该库的用户提供非常有限的支持。还有一点就是anaconda自带了众多的科学计算库，一步到位确实非常省心。

对不使用anaconda的少年的说明

对于不想安装anaconda的少年（conda速度真的不多说，推荐使用清华tuna的源），自然我提出一些小小的参考。

你需要安装numpy等一众科学计算的库，你可能还需要安装BLAS、MKL等一系列库以求支持TF。建议知道这些东西的人手工配齐库。

如果遇到问题，百度或者google吧。

安装OpenCL

就我来说，AMD网上的驱动几乎不能成功。所以我选择了一种非常优雅的方式，apt安装。

      
      

       
       $ sudo add-apt-repository ppa:paulo-miguel-dias/mesa 
      
      
      
      

       
       $ sudo apt-get update
      
      
      
      

       
       $ sudo apt-get install libclc-amdgcn mesa-opencl-icd

移除这个PPA

1 2	$ sudo apt-get install ppa-purge $ sudo ppa-purge ppa:oibaf/graphics-drivers

更多的信息你可以参考这篇博文：https://laanwj.github.io/2016/05/06/opencl-ubuntu1604.html

为Theano安装的准备

如果你不需要安装Theano，那么你可以跳过这一项，下面这项目会自动引用系统自带的Python，与下面所说的安装Python有所出入。

Theano sucks

为theano的支持安装clblas库

      
      

       
       ## clBlas
      
      
      
      
       
       sudo apt-get install git
      
      
      
      
       
       git clone https://github.com/clMathLibraries/clBLAS.git
      
      
      
      
       
       cd clBLAS/
      
      
      
      
       
       mkdir build
      
      
      
      
       
       cd build/
      
      
      
      
       
       sudo apt-cache search openblas
      
      
      
      
       
       sudo apt-get install libopenblas-base libopenblas-dev
      
      
      
      
       
       sudo apt-get install liblapack3gf liblapack-doc liblapack-dev
      
      
      
      
       
       cmake ../src
      
      
      
      
       
       make
      
      
      
      
       
       sudo make install

接下来安装libgpuarray

      
      
       
       git clone https://github.com/Theano/libgpuarray.git
      
      
      
      
       
       cd libgpuarray
      
      
      
      
       
       mkdir Build
      
      
      
      
       
       cd Build
      
      
      
      
       
       cmake .. -DCMAKE_BUILD_TYPE=Release -DOPENCL_INCLUDE_DIRS=/opt/AMDAPPSDK-3.0-0-Beta/include
      
      
      
      
       
       make
      
      
      
      
       
       sudo make install
      
      
      
      
       
       cd ..
      
      
      
      
       
       sudo apt-get install cython
      
      
      
      
       
       sudo apt-get install python-numpy python-scipy python-dev python-pip python-nose g++ libopenblas-dev git
      
      
      
      
       
       python setup.py build
      
      
      
      
       
       sudo python setup.py install

安装 Theano

      
      

       
       ## Theano
      
      
      
      
       
       pip install Theano
      
      
      
      
       
       sudo pip install Theano

安装Tensorflow

安装computeCPP

这个库真的挺奇怪的，但是安装带有OpenCL支持的Tensorflow却不得不需要这个奇怪的库。首先你需要在其官网上注册一个账号来下载ComputeCPP，和牙膏厂一个尿性，然后解压后放到/usr/lib或者其他你能看得到的位置。

Tensorflow对于OpenCL的支持至今都是很有限的，所以你需要的是使用源码来安装Tensorflow而不是优雅的使用pip3 install tensorflow-gpu（都怪农企喜欢造挖矿卡）

直接参考Tensorflow的如何编译源码来安装就可以了。需要注意的是在enable Tensorflow support那块需要选择是，而CUDA那一块则要选择否。假设出现了要你选择OpenCL路径地方的时候，他显示的那个路径一般都是没有错的。所以这个时候你需要检查那个路径下是否生成了libOpencl.so这类型的文档。

这里就是英文的说明了：https://www.tensorflow.org/install/install_sources

安装Keras

一句话 pip3 install keras

如何知道我是用了OpenCL

和使用CPU一般出现的TF让你编译TF到那些奇奇怪怪的（AXS是什么）格式，我的就会出现这种信息：

      
      
       
       2017-07-07 20:35:24.377623: I tensorflow/compiler/xla/service/platform_util.cc:58] platform Host present with 8 visible devices
      
      
      
      
       
       2017-07-07 20:35:24.378838: I tensorflow/compiler/xla/service/service.cc:198] XLA service 0xaa19620 executing computations on platform Host. Devices:
      
      
      
      
       
       2017-07-07 20:35:24.378870: I tensorflow/compiler/xla/service/service.cc:206]   StreamExecutor device (0): <undefined>, <undefined>

postscript

In order to pip install quickly, you can use some image library, such as source USTC, Taobao and Tsinghua tuna and the like. Here I would rather recommend students (such as juvenile side Village Polytechnic) using Qinghua source to avoid sharp McNair kill. The following method is used:

https://mirrors.tuna.tsinghua.edu.cn/help/pypi/

Then Theano how I did not installed this, is libgpuarray limited graphics support on my show, and I can not read anyway error Orz.

Finally, I wish computers do not explode.

Original link large column https://www.dazhuanlan.com/2019/08/18/5d585c96d8131/

Ubuntu 16.04 + OpenCL + Tensorflow + Keras Configuration Reference