docker.md

Running in Docker

In order to build Docker image you have to checkout GFootball git repository first:

git clone https://github.com/google-research/football.git
cd football

For rendering the game on macOS and Windows, we recommend installing the game according to the instructions for your platform in README.

Configure Docker for Rendering (Linux only)

In order to see rendered game you need to allow Docker containers access X server:

xhost +"local:docker@"

This command has to be executed after each reboot. Alternatively you can add this command to /etc/profile to not worry about it in the future.

Build Docker image

Tensorflow without GPU-training support version

docker build --build-arg DOCKER_BASE=ubuntu:20.04 . -t gfootball

Tensorflow with GPU-training support version

docker build --build-arg DOCKER_BASE=tensorflow/tensorflow:1.15.2-gpu-py3 . -t gfootball

Start the Docker image

docker run --gpus all -e DISPLAY=$DISPLAY -it -v /tmp/.X11-unix:/tmp/.X11-unix:rw gfootball bash

If you get errors related to --gpus all flag, you can replace it with --device /dev/dri/[X] adding this flag for every file in the /dev/dri/ directory. It makes sure that GPU is visible inside the Docker image. You can also drop it altogether (environment will try to perform software rendering).

Run environment

Inside the Docker image you can interact with the environment the same way as in case of local installation. For example, to play the game yourself you can run:

python3 -m gfootball.play_game --action_set=full

To run example PPO2 training you need to install OpenAI Baselines in addition.

Building a docker image under MacOS

You may need to increase memory for building. Go to the docker menu, then Preferences, then Advanced/Memory and set memory to the 4GB.