强化学习需要 atari 环境,使用 dockerfile 快速构建。
下载 ROMS
wget http://www.atarimania.com/roms/Roms.rar
unrar -e Roms.rar
unzip ROMS.zip
创建文件夹 dockerbuild/
mkdir dockerbuild && cd dockerbuild
cp -r ../ROMS/ .
vim my.dockerfile
复制以下内容,ubuntu20.04+cu112+python3.8,大小为 10 G。
FROM nvidia/cuda:11.2.2-cudnn8-runtime-ubuntu20.04
MAINTAINER Jiale Han
SHELL ["/bin/bash","-c"]
WORKDIR /root
ADD ROMS/ ROMS/
RUN apt update \
&& apt install -y python3 python3-pip tmux \
&& pip install opencv-python matplotlib gym atari-py ale-py \
&& python3 -m atari_py.import_roms ROMS/ \
&& apt install -y libgl1-mesa-glx \
&& DEBIAN_FRONTEND=noninteractive apt install -y libglib2.0-dev \
&& apt clean
或者下面这个,ubuntu18.04+cu102+python3.7+virtualenv+pytorch1.82,大小为 2 G。
FROM nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04
MAINTAINER Jiale Han
SHELL ["/bin/bash", "-c"]
WORKDIR /root/
ENV LANG C.UTF-8
RUN apt update && \
# install python3.7
apt install -y software-properties-common && \
add-apt-repository -y ppa:deadsnakes/ppa && \
apt install -y python3.7 python3.7-distutils python3.7-dev && \
apt install -y vim tmux wget && \
# install virtualenv & virtualenvwrapper
wget https://bootstrap.pypa.io/get-pip.py && \
python3.7 get-pip.py && \
pip install virtualenv virtualenvwrapper && \
sed -i 's/which python/which python3.7/g' `which virtualenvwrapper.sh` && \
echo -e "\n# virtualenvwrapper" >> .bashrc && \
echo "export WORKON_HOME=/root/.virtualenvs" >> .bashrc && \
echo "export VIRTUALENVWRAPPER_VIRTUALENV=`which virtualenv`" >> .bashrc && \
echo "export VIRTUALENVWRAPPER_PYTHON=/usr/bin/python3.7" >> .bashrc && \
echo "source `which virtualenvwrapper.sh`" >> .bashrc && \
apt clean && \
rm get-pip.py
COPY ROMS/ /root/ROMS/
RUN source `which virtualenvwrapper.sh` && \
mkvirtualenv py37 && \
workon py37 && \
pip install opencv-python matplotlib ipython && \
pip install gym ale-py && \
ale-import-roms ROMS/ && \
deactivate && \
apt install -y libgl1-mesa-glx libglib2.0-dev && \
apt clean && \
rm -rf ROMS/
COPY .oh-my-zsh/ /root/.oh-my-zsh/
RUN apt install -y zsh && \
cp .oh-my-zsh/templates/zshrc.zsh-template .zshrc && \
sed -i 's/plugins=(git)/plugins=(git zsh-autosuggestions zsh-syntax-highlighting)/g' .zshrc && \
echo -e "\n# virtualenvwrapper" >> .zshrc && \
echo "export WORKON_HOME=/root/.virtualenvs" >> .zshrc && \
echo "export VIRTUALENVWRAPPER_VIRTUALENV=`which virtualenv`" >> .zshrc && \
echo "export VIRTUALENVWRAPPER_PYTHON=/usr/bin/python3.7" >> .zshrc && \
echo "source `which virtualenvwrapper.sh`" >> .zshrc && \
chsh -s $(which zsh) && \
apt clean
COPY id_rsa.pub /root/.ssh/id_rsa.pub
RUN apt install -y openssh-server && \
cat /root/.ssh/id_rsa.pub >> /root/.ssh/authorized_keys && \
chmod 600 /root/.ssh/authorized_keys && \
sed -i 's/#PubkeyAuthentication yes/PubkeyAuthentication yes/g' \
/etc/ssh/sshd_config && \
sed -i 's/#PasswordAuthentication yes/PasswordAuthentication no/g' \
/etc/ssh/sshd_config && \
sed -i 's/#PermitRootLogin prohibit-password/PermitRootLogin prohibit-password/g' \
/etc/ssh/sshd_config && \
apt clean
EXPOSE 22
COPY init.sh /root/init.sh
ENTRYPOINT ["/bin/zsh", "init.sh"]
使用镜像运行容器后,
启动 ssh 服务
service ssh restart
安装 pytorch1.8.2
pip install torch==1.8.2+cu102 torchvision==0.9.2+cu102 \
-f https://download.pytorch.org/whl/lts/1.8/torch_lts.html
pip install 'ray[rllib]' tensorboard==1.15
或者把 nvidia/cuda:10.2
改成 nvidia/cuda:10.0
,使用镜像运行容器后,安装 tensorflow1.15
pip install tensorflow-gpu==1.15
pip install 'ray[rllib]'
如果要运行框架
export HOROVOD_GPU_OPERATIONS=NCCL
apt install git cmake g++
pip install influxdb psutil pyzmq pyarrow horovod
conda 镜像
FROM nvidia/cuda:10.2-cudnn7-devel-ubuntu18.04
MAINTAINER Jiale Han
SHELL ["/bin/bash","-c"]
WORKDIR /root
RUN apt update \
&& apt install -y vim cmake tmux wget \
&& wget https://repo.anaconda.com/miniconda/Miniconda3-py37_4.10.3-Linux-x86_64.sh \
&& echo -e "\nexport export HOROVOD_GPU_OPERATIONS=NCCL" >> .bashrc \
&& apt install -y libgl1-mesa-glx libglib2.0-dev git\
&& apt clean \
构建镜像
cd dockerbuild/
docker build -t nvidia/cuda:cu100-ubu18-rl-zsh -f tf1.dockerfile .
从镜像运行容器
docker run -it -v /data2/hanjl:/root/code/ \
-p 0.0.0.0:2200:22 -p 0.0.0.8000:80 \
--gpus '"device=1,2"' --name hjl_test \
nvidia/cuda:cu100-ubu18-rl-zsh /bin/zsh