一、配置百度源

将/etc/apt/sources.list的内容替换为如下：
以Ubuntu18.04为例

# 默认注释了源码镜像以提高 apt update 速度，如有需要可自行取消注释deb http://mirrors.baidu.com/file/ubuntu/ bionic main restricted universe multiverse# deb-src https:/mirrors.baidu.com/file/ubuntu/ bionic main restricted universe multiversedeb http://mirrors.baidu.com/file/ubuntu/ bionic-updates main restricted universe multiverse# deb-src https://mirrors.baidu.com/file/ubuntu/ bionic-updates main restricted universe multiversedeb http://mirrors.baidu.com/file/ubuntu/ bionic-backports main restricted universe multiverse# deb-src https://mirrors.baidu.com/file/ubuntu/ bionic-backports main restricted universe multiversedeb http://mirrors.baidu.com/file/ubuntu/ bionic-security main restricted universe multiverse# deb-src https://mirrors.baidu.com/file/ubuntu/ bionic-security main restricted universe multiverse# 预发布软件源，不建议启用# deb http://mirrors.baidu.com/file/ubuntu/ bionic-proposed main restricted universe multiverse# deb-src http://mirrors.baidu.com/file/ubuntu/ bionic-proposed main restricted universe multiverse

再执行 apt-get update更新

二、启动docker

创建start_docker.sh脚本

#/bin/bash# paddle1.7.1镜像地址# image_name=registry.baidu.com/paddlecloud/paddlecloud-runenv-centos6u3-online:paddlecloud-v1.7.1-gcc482-cuda10.0_cudnn7# image_name=registry.baidubce.com/paddlepaddle/paddle:2.0.0-gpu-cuda10.1-cudnn7#image_name=registry.baidu.com/vis-general-ocr/paddle:2.0.0-gpu-cuda10.1-cudnn7-pdc-base#image_nmame=iregistry.baidu-int.com/paddlecloud/pytorch1.4.0:ubuntu16.04-cuda10.1_cudnn7image_name=iregistry.baidu-int.com/paddlecloud/pytorch1.4.0:ubuntu16.04-cuda10.1_cudnn7#image_name=iregistry.baidu-int.com/paddlecloud/paddlecloud-runenv-ubuntu16.04-offline:paddlecloud-paddle-v2.2.2-gcc820-cuda10.1_cudnn7# 容器名称container_name=$1guazai_path= xxx# 挂载物理机GPU驱动库export CUDA_SO="$(\ls /usr/lib64/libcuda* | xargs -I{} echo '-v {}:{}') $(\ls /usr/lib64/libnvidia* | xargs -I{} echo '-v {}:{}')"# 挂载nvidia硬件export DEVICES=$(find /dev/nvidia* -maxdepth 1 -not -type d | xargs -I{} echo '--device {}:{}')# -d: 后台运行容器# -v: 格式为 {宿主机目录}:{容器目录}，将容器目录挂载到宿主机目录上# --name: 为容器命名# -it: 以交互式的方式启动容器# --network=host: 使用宿主机网络docker run ${CUDA_SO} ${DEVICES} -it -d --network=host --ipc=host \-v ${guazai_path}:${guazai_path} \-v /usr/bin/nvidia-smi:/usr/bin/nvidia-smi \--name ${container_name} ${image_name} bash

ps：docker镜像为百度内镜像，torch版本较低，后续手动安装
启动脚本

source start_docker.sh ${container_name}

三、docker基本操作

启动docker

image_name 为创建的容器名称

docker exec -it ${container_name} bash

删除docker

首先需要先关闭docker

docker stop ${image_id}

然后再删除

docker rm ${image_id}

四、安装环境

1、安装python3.6以上的版本

docker中自带python2.7 和python3.6，在这里我们将其升级为python3.6以上的版本，去官网下载对应的python版本，以python3.7为例。
a. 查看当前安装的python：

which pythonwhich python3Out:/opt/conda/envs/py27/bin/python/usr/bin/python3

b. 安装依赖，执行下列命令安装依赖过程中，如有提示，一律输入 y 。

sudo apt-get install python-dev libffi-dev libssl-devsudo apt-get install -y make build-essential libssl-dev zlib1g-dev libbz2-dev libreadline-dev libsqlite3-dev wget curl llvm libncurses5-dev libncursesw5-dev xz-utils tk-dev

c. 执行安装

./configure prefix=/usr/local/python3make && make install

d. 修改软连接（配置全局变量）

#备份现有的软连接mv /usr/bin/python3 /opt/conda/bin/python3.bak#添加python3的软链接ln -s /usr/local/python3/bin/python3.7 /opt/conda/bin/python3#测试是否安装成功了python3 -V

e. 安装/升级pip

apt-get install python3-pip# 执行升级pip3 install --upgrade pip

2、安装torch

去官网下载python对应的torch / vision版本，在用pip3进行安装，以python3.7，torch1.10的版本为例：

pip3 install torch-1.10.1+cu113-cp37-cp37m-linux_x86_64.whlpip3 install torchvision-0.11.2+cu113-cp37-cp37m-linux_x86_64.whl