Dockerfiles for DeepRL research.

Dependency

• If you want to build docker images which support mujoco, you must need mjkey.txt for building the docker images. You can also customize the dockerfile upon our dockerfiles.
• I am grateful to DeepMind for making MuJoCo free. Now, I add the mjkey.txt to the corresponding folder, you can directly compile the dockerfile for running the MuJoCo environment. Also, you can customize the dockerfile upon our dockerfiles.
• See the directory dmc-mujoco-atari-torch1.4-tf1.14-cu100, dmc-mujoco-atari-torch1.7-cu110, etc, for more details.

Recommendation

• If your machine can work on CUDA 10.0 or CUDA 10.1, we recommend for reading the dmc-mujoco-atari-torch1.4-tf1.14-cu100 or dmc-mujoco-atari-torch1.5-tf1.14-cu101 dockerfile. The image built by dmc-mujoco-atari-torch1.4-tf1.14-cu100 or dmc-mujoco-atari-torch1.5-tf1.14-cu101 dockerfile contains:

  • DeepMind Control Suite
  • MuJoCo 200
  • Atari
  • CUDA 10.0 / CUDA 10.1
  • Torch 1.4 / Torch 1.5
  • Tensorflow 1.14

• If your machine only works on CUDA 11.x, we recommend for reading the dmc-mujoco-atari-torch1.7-cu110 dockerfile. The image built by dmc-mujoco-atari-torch1.7-cu110 dockerfile contains:

  • DeepMind Control Suite
  • MuJoCo 200
  • Atari
  • CUDA 11.0
  • Torch 1.7.1
  • Tensorflow

Ready-made Docker images

• We also provide our compiled docker images to help you run MuJoCo environment quickly.

  docker pull liqingya/mujoco:py36-torch1.4-tf1.14-cu100  # `mujoco-torch1.4-tf1.14-cu100`
  
  docker pull liqingya/mujoco:dmc-atari-py36-torch1.4-tf1.14-cu100  # `dmc-mujoco-atari-torch1.4-tf1.14-cu100`
  
  docker pull liqingya/mujoco:dmc-atari-py36-torch1.5-tf1.14-cu101  # `dmc-mujoco-atari-torch1.5-tf1.14-cu101`
  
  docker pull liqingya/mujoco:dmc-atari-py36-torch1.7-cu110  # dmc-mujoco-atari-torch1.7-cu110

Usage of Tmux in our Docker

• We have added .tmux.conf to the docker image for customizing tmux usage:

  # Pane splitting commands
  Ctrl-a + v # split pane along vertical direction
  Ctrl-a + h # split pane along horizontal direction
  
  # Mouse mode for fast Pane-Switching
  # We can switch pane by simply clicking the pane through the mouse.
  
  # Switch out window
  Ctrl-a + d

Start the Container

• As the root (not secure)
  Example script:

  docker run --gpus all -itd --rm --name [container_name] \
    -v [local-dir]:/share \
    -it [image_name] /bin/zsh

• As the user (non-root, recommend this way)
  Example script:

  docker run --gpus all -itd --user $(id -u ${USER}):$(id -g ${USER}) --rm --name [container_name] \
    -v /etc/passwd:/etc/passwd:ro -v /etc/group:/etc/group:ro -v /etc/shadow:/etc/shadow:ro -v /home/${USER}:/home/${USER}:ro \
    -v [local-dir]:/share \
    -it [image_name] /bin/bash

  For the usage of MuJoCo as the user in the docker container:

  • Firstly, you should create ~/.mujoco directory, and then move mujoco_200 which unzipped from mujoco200_linux.zip and mjkey.txt to ~/.mujoco:

    # cd to the `mujoco200_linux.zip` and 'mjkey.txt` location.
    mkdir -p ~/.mujoco && \
    unzip mujoco200_linux.zip && \
    mv mujoco200_linux ~/.mujoco/mujoco200_linux && \
    cp -r ~/.mujoco/mujoco200_linux ~/.mujoco/mujoco200 && \
    cp mjkey.txt ~/.mujoco/

  • Secondly, edit the bash file:

    vim ~/.bashrc

    And then add the following scripts to the end of the file:

    export LD_LIBRARY_PATH=~/.mujoco/mujoco200/bin${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
    export LD_LIBRARY_PATH=~/.mujoco/mujoco200_linux/bin${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
    export MUJOCO_KEY_PATH=~/.mujoco${MUJOCO_KEY_PATH}

  • Finally, after create container, enter the container and input the following script:

    sudo chmod 777 /usr/local/lib/python3.6/dist-packages/mujoco_py/generated/