原作者:曾晋哲
原链接: https://mp.weixin.qq.com/s/DeQZKwoyxXK7rHhhmdgTmw https://mp.weixin.qq.com/s/MsEtgnN\_mi-auiFyXYUquA
假定已经安装了Anaconda(建议使用最新版2019.07),已连接互联网,则
1.安装tensorflow(如仅需CPU版本的TensorFlow,则将tensorflow-gpu改为tensorflow):
1 | python -m pip install tensorflow Successfully installed google-pasta-0.1.7 keras-applications-1.0.8 opt-einsum-3.1.0 tensorboard-2.0.0 tensorflow-estimator-2.0.0 tensorflow-gpu-2.0.0 |
2.安装deepmd-kit v1.0:
1 | python -m pip install git+https://github.com/deepmodeling/deepmd-kit |
Building wheels for collected packages: deepmd-kit
Building wheel for deepmd-kit (PEP 517) … done
Created wheel for deepmd-kit: filename=deepmd_kit-1.0.0-cp37-cp37m-linux_x86_64.whl size=268836 sha256=1f5b1149bbf35c0c96c713cc8b607e0626ad0df6451a7 171d4f6b46acc2d4290
Stored in directory: /tmp/pip-ephem-wheel-cache-zlksq4dl/wheels/a2/80/6c/a26fba79e43199eb4cdba7a3686c5370d3620916f5a0ea23ac
Successfully built deepmd-kit
Installing collected packages: deepmd-kit
Successfully built deepmd-kit
Installing collected packages: deepmd-kit
Successfully installed deepmd-kit-1.0.0
大功告成!现在看一看是否成功安装:
1 | dp -h |
usage: dp [-h] {train,freeze,test} …
DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics
optional arguments: -h, –help
show this help message and exit Valid subcommands: {train,freeze,test}
train train a model
freeze freeze the model
test test the model
现在,DeePMD-kit v1.0.0已成功安装。下一期将介绍如何用DP编译LAMMPS
TensorFlow 安装
最新版本的 TensorFlow 要求 GLIBC 2.17 以上,尽管推荐做法是找一台最新系统的机子,但是有时候系统的类型不是由自己决定的,通常又没有root权限,又想在所有机子上都能运行 TensorFlow 。 刚好手里有一个超算账号,系统是 Red Hat 4.4.7 ,GLIBC 版本是 2.12 ,就以此为例,安装CPU版本的TensorFlow(反正没有权限也安装不了GPU版本需要的驱动)。
一、用 Anaconda 3 安装 TensorFlow 1.8
1.安装 Anaconda 3 见Linux软件安装②Anaconda3 2.创建 TensorFlow 环境
1 | conda create -n tensorflow pip python=3.6 |
二、安装 gcc
这时候打开 Python ,执行 import tensorflow ,提示:
ImportError: /usr/lib64/libstdc++.so.6: version `CXXABI_1.3.7’ not found
1 | conda install -c psi4 gcc-5 |
再此运行 Python,不再提示这个问题。
三、安装 GLIBC 2.21
但是提示:
ImportError: /lib64/libc.so.6: version `GLIBC_2.16’ not found
本来应该安装GLIBC 2.17,但是我发现从2.16到2.19都有个bug,不能运行Python 3.6。于是我们安装GLIBC 2.21。 1.下载GLIBC 2.21并编译GLIBC 2.21
1 | wget http://mirror.rit.edu/gnu/libc/glibc-2.21.tar.gz |
然后就报错了:
checking version of as… 2.20.51.0.2, bad checking version of ld… 2.20.51.0.2, bad These critical programs are missing or too old: as ld
仔细看看INSTALL文件,要求GNU ‘binutils’ 2.22 or later,但系统只装了2.20。 2.下载并编译binutils 2.30
1 | wget ftp://ftp.gnu.org/gnu/binutils/binutils-2.30.tar.gz |
3.重新编译glibc 2.21
1 | cd glibc-2.21-build |
Warning: ignoring configuration file that cannot be opened: … /software/glibc-2.21-install/etc/ld.so.conf: No such file or directory
将/etc 目录的ld.so.conf复制到指定目录后重新安装:
1 | cp /etc/ld.so.conf ../glibc-2.21-install/etc/ |
安装成功。
四、运行TensorFlow
1 | source activate tensorflow |
在Python内输入:
1 | # Python |
b’Hello, TensorFlow!’
运行成功。我们可以运行的命令记录在.bashrc中:
1 | echo 'alias tf='"'"'$HOME/software/glibc-2.21-install/lib/ld-2.21.so --library-path $HOME/anaconda3/envs/tensorflow/lib:$HOME/software/glib-2.21-install/lib:/lib64:$LD_LIBRARY_PATH `which python`'"'">>$HOME/.bashrcsource $HOME/.bashrc |
即可用 tf 代替装了 TensorFlow 的 Python。