DeePMD-kit是一个流行的多体（many-body）深度学习势（deep learning potential）框架，需要从头安装tensorflow，比较繁琐，官网的安装教程也有一些暗坑，本文主要讲解如何在Ubuntu下顺利安装DeePMD-kit，主要参考资料[安装deepmd]和[源码安装TensorFlow]

https://blog.csdn.net/xszyqbr/article/details/82961934

https://github.com/deepmodeling/deepmd-kit/blob/master/doc/install-tf.1.12.md

在线安装必须的工具

#automake 工具
sudo apt-get install autoconf automake libtool
#根据是否已经安装make,git和cmake决定是否安装
sudo apt-get install make cmake git

安装JAVA环境

• 安装OpenJDK 8 JDK sudo apt install openjdk-8-jdk

• 设置JAVA_HOME变量

  #查看java可执行文件的实际目录，一般指向/etc/alternatives/java
  ls -lrt /usr/bin/java
  #查看/etc/alternatives/java的实际目录
  ls -lrt /etc/alternatives/java
  #一般指向 [/usr/lib/jvm/java-1.8.0-openjdk.x86_64](接下来用的到的)/bin/java
  export JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk.x86_64
  export JRE_HOME=/usr/lib/jvm/java-1.8.0-openjdk.x86_64/jre
  export CLASSPATH=.:$JAVA_HOME/lib:$JRE_HOME/lib:$CLASSPATH
  export PATH=$JAVA_HOME/bin:$JRE_HOME/bin:$JAVA_HOME:$PATH
  

安装bazel，必须安装0.15.0及以上版本

1. 参考安装deepmd教程安装，但版本不一样，如果安装失败，按照方法2安装

2. 使用预编译好的0.15.0版本的bazel二进制版本（https://blog.csdn.net/darkrabbit/article/details/81262556）

3. 安装完成后，将bazel位置放在环境变量里,方法2默认会自动把可执行文件放到 /home/usr/bin目录下，如果默认这个目录已经在PATH里无需主动添加。

安装Anaconda3.5，并用conda安装blas

1. 按照[安装deepmd]安装Anaconda，并加入到环境变量

2. 使用conda安装blas sudo conda install --channel https://conda.anaconda.org/anaconda blas

3. 按照[源码安装TensorFlow]教程使用pip安装tensorflow1.12.0

下载tensorflow源码，安装tensorflow的C++接口

1. 进入想放置tensorflow安装包的目录，比如cd ~

2. git clone https://github.com/tensorflow/tensorflow tensorflow -b v1.12.0 --depth=1

3. cd tensorflow

4. ./configure,有较多的选项，参考博客（https://blog.csdn.net/yhily2008/article/details/79967118）选择配置。

5. 使用bazel 编译tensorflow，可以控制编译时的所用资源 等待非常久的时间bazel build -c opt --verbose_failures //tensorflow:libtensorflow_cc.so

6. 设置tensorflow的主安装目录,我装在用户目录~/tensorflow_root下

   export tensorflow_root=~/tensorflow_root
   mkdir -p ${tensorflow_root}
   

在当前目录下下载tensorflow的依赖文件:Protobuf, Eigen, nsync and absl

• Protobuf

  #设置在${tensorflow_root}目录下新建tmps文件夹用于放置中间文件
  mkdir -p ${tensorflow_root}/tmps
  mkdir -p ${tensorflow_root}/tmps/proto
  sed -i 's;PROTOBUF_URL=.*;PROTOBUF_URL=\"https://mirror.bazel.build/github.com/google/protobuf/archive/v3.6.0.tar.gz\";g' tensorflow/contrib/makefile/download_dependencies.sh
  tensorflow/contrib/makefile/download_dependencies.sh
  cd tensorflow/contrib/makefile/downloads/protobuf/
  ./autogen.sh
  ./configure --prefix=${tensorflow_root}/tmps/proto
  make
  make install
  

• Eigen

  mkdir -p ${tensorflow_root}/tmps/eigen
  cd ../eigen
  mkdir build_dir
  cd build_dir
  cmake -DCMAKE_INSTALL_PREFIX=${tensorflow_root}/tmps/eigen/ ../
  make install
  

• nsync

  mkdir -p ${tensorflow_root}/tmps/nsync
  cd ../../nsync
  mkdir build_dir
  cd build_dir
  cmake -DCMAKE_INSTALL_PREFIX=${tensorflow_root}/tmps/nsync/ ../
  make
  make install
  

• absl

  cd ../../absl
  bazel build
  mkdir -p ${tensorflow_root}/include/
  sudo rsync -avzh --include '*/' --include '*.h' --exclude '*' absl ${tensorflow_root}/include/
  cd ../../../../..
  

• 将以上库文件拷贝到tensorflow的安装目录

  mkdir -p ${tensorflow_root}/lib
  cp bazel-bin/tensorflow/libtensorflow_cc.so ${tensorflow_root}/lib/
  cp bazel-bin/tensorflow/libtensorflow_framework.so ${tensorflow_root}/lib/
  cp ${tensorflow_root}/tmps/proto/lib/libprotobuf.a ${tensorflow_root}/lib/
  cp ${tensorflow_root}/tmps/nsync/lib/libnsync.a ${tensorflow_root}/lib/
  #可能是nsync/lib64/libnsync.a
  

• 然后拷贝头文件

  mkdir -p ${tensorflow_root}/include/tensorflow
  cp -r bazel-genfiles/* ${tensorflow_root}/include/
  cp -r tensorflow/cc ${tensorflow_root}/include/tensorflow
  cp -r tensorflow/core ${tensorflow_root}/include/tensorflow
  cp -r third_party ${tensorflow_root}/include
  cp -r ${tensorflow_root}/tmps/proto/include/* ${tensorflow_root}/include
  cp -r ${tensorflow_root}/tmps/eigen/include/eigen3/* ${tensorflow_root}/include
  cp -r ${tensorflow_root}/tmps/nsync/include/*h ${tensorflow_root}/include
  

• 清理头文件目录下的源文件

  cd ${tensorflow_root}/include
  find . -name "*.cc" -type f -delete
  

• 可以删除所有依赖的缓存文件

  rm -fr ${tensorflow_root}/tmps/eigen ${tensorflow_root}/tmps/nsync ${tensorflow_root}/tmps/proto
  

安装DeePMD-kit

• 先在${tensorflow_root}的上级目录克隆DeePMD-kit源码

  cd ${tensorflow_root} && cd ../
  git clone https://github.com/deepmodeling/deepmd-kit.git deepmd-kit
  

• 通过deepmd_source_dir变量记录源代码的位置

  cd deepmd-kit
  export deepmd_source_dir=`pwd`
  cd ${deepmd_source_dir}/source
  mkdir build
  cd build
  

• 设置deepmd-kit的主安装目录,我装在用户目录~/deepmd_root下,执行cmake，

  export deepmd_root=~/deepmd_root
  mkdir -p ${deepmd_root}
  #如果用gcc>5.0 去掉选项-DTF_GOOGLE_BIN=true，如果运行时出现deepmd-kit/lib/deepmd/libop_abi.so: undefined symbol:请打开选项-DTF_GOOGLE_BIN=true
  cmake -DTF_GOOGLE_BIN=true -DTENSORFLOW_ROOT=${tensorflow_root} \
  -DCMAKE_INSTALL_PREFIX=${deepmd_root} ..
  

• 如果cmake执行成功，make以及make install

  make
  make install
  

• 如果一切顺利地话，你将在$deepmd_root/bin目录下找到以下可执行文件

  $ ls ${deepmd_root}/bin
  dp_frz dp_ipi dp_test dp_train
  #可以加入系统PATH
  echo 'PATH='${deepmd_root}'/bin:$PATH' >> ~/.bashrc
  source ~/.bashrc
  

Install LAMMPS’s DeePMD-kit module

参考官方安装教程，首先得学会lammps的安装。

Build DeePMD-kit with GPU support

参考官方安装教程。