# TensorFlow 安装配置

---

## 现有的机器环境

> `CentOS 7 (Linux 3.10.0-327.el7.x86_64)`

> `gcc 4.8.5`

> `glibc 2.17`

> `cuda 7.5`

> 没有 `g++`、`libstdc++`、`cudnn`、`Java 1.8` 和 `bzip2`、`unzip`、`swig`、`zlib-devel`

> 因为使用 `Anaconda` 提供 `python` 环境，所以并不依赖系统的 `python 2.7.5`

## 先使用管理员权限做一些必要的准备工作

### 1. 设置新用户 danliu

```

[root ~]# useradd danliu

[root ~]# passwd danliu

```

### 2. 安装必备软件

#### g++ 4.8.5

```

[root ~]# yum install gcc-c++

```

libstdc++ 会作为依赖项一并安装

### bzip2, unzip, swig 和 Java 1.8

```

[root ~]# yum install bzip2 unzip swig zlib-devel java-1.8.0-openjdk-devel

```

### package.zip

[root ~]# cd /home/

[root ~]# unzip packages.zip

[root ~]# cd /packages/

[root ~]# mv * /home/danliu

[root ~]# cd ..

[root ~]# rm -rf packages

[root ~]# cd /home/danliu/

### bazel

bazel 作为构建工具，安装在系统里更方便。

```

[root ~]# chmod +x bazel-0.3.1-installer-linux-x86_64.sh

[root ~]# ./bazel-0.3.1-installer-linux-x86_64.sh

```

## 使用普通账户 danliu 登陆并继续接下来的安装

### 安装 Anaconda

```

[root ~]# chown -R danliu:danliu /home/danliu

[root ~]# su danliu

[danliu ~]$ chmod +x Anaconda2-4.1.1-Linux-x86_64.sh

[danliu ~]$ ./Anaconda2-4.1.1-Linux-x86_64.sh

Do you wish the installer to prepend the Anaconda2 install location

to PATH in your /home/danliu/.bashrc ? [yes|no]

[no] >>> yes

```

默认的安装目录是 `~/anaconda2/` (`/root/anaconda2`)

安装完成后

```

source ~/.bashrc

```

### 添加已有的 cuda 7.5 安装

和下面的安装 cudnn v5 一起做

### 安装 cudnn v5

```

[danliu ~]$ tar -xzvf cudnn-7.5-linux-x64-v5.0-ga.tar.gz

[danliu ~]$ mv cuda/ cudnnv5

```

把下面的内容添加到 `~/.bashrc` 里

```

# CUDA

vim ~/.bashrc

export PATH=/usr/local/cuda-7.5/bin:$PATH

export LD_LIBRARY_PATH=/usr/local/cuda-7.5/lib64:$LD_LIBRARY_PATH

export LD_LIBRARY_PATH=$HOME/cudnnv5/lib64:$LD_LIBRARY_PATH

```

### 修改 TensorFlow 的一处 bug

TensorFlow 下的 `third_party/gpus/crosstool/CROSSTOOL` 存在一个可能导致构建出错的问题，如果使用 Ubuntu 14.04 可以无视，使用 CentOS 7 则必须要修改

```

[danliu ~]$ unzip tensorflow-src.zip

[danliu ~]$ mv tensorflow-master tensorflow-cudnnv5

[danliu ~]$ cd tensorflow-cudnnv5

[danliu ~]$ vi third_party/gpus/crosstool/CROSSTOOL

```

在第 66 行和第 67 行之间添加一行`cxx_builtin_include_directory: "/usr/local/cuda-7.5/include"`

```

@@ -64,6 +64,7 @@

cxx_builtin_include_directory: "/usr/lib/gcc/"

cxx_builtin_include_directory: "/usr/local/include"

cxx_builtin_include_directory: "/usr/include"

+  cxx_builtin_include_directory: "/usr/local/cuda-7.5/include"

tool_path { name: "gcov" path: "/usr/bin/gcov" }

# C(++) compiles invoke the compiler (as that is the one knowing where

```

### 配置 TensorFlow

```

[danliu *]$ ./configure

```

注意 cudnn 的路径是之前设置的 `/home/danliu/cudnnv5`

> ERROR: Problem getting numpy include path.  Is numpy installed?

### 编译和测试

```

# 下面的 // 不是注释是一行

[danliu *]$ bazel build -c opt --config=cuda //tensorflow/cc:tutorials_example_trainer

[danliu *]$ bazel-bin/tensorflow/cc/tutorials_example_trainer --use_gpu

```

### 构建可以迁移的 tensorflow wheel 包

```

[danliu *]$ bazel build -c opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

[danliu *]$ bazel-bin/tensorflow/tools/pip_package/build_pip_package ./tensorflow_pkg

```

> bazel-bin/tensorflow/tools/pip_package/build_pip_package:行65: rsync: 未找到命令