Run garage with Docker

Currently there are two types of garage images:

  • headless: garage without environment visualization.

  • nvidia: garage with environment visualization using an NVIDIA graphics card.

Headless image

Prerequisites

Be aware of the following prerequisites to build the image.

  • Install Docker CE version 19.03 or higher. Tested on version 19.03.12.

Tested on Ubuntu 16.04, 18.04 & 20.04.

Run a pre-compiled image

If you already have a source copy of garage, proceed to subsection Build and run the headless image, otherwise, keep reading.

To run an example launcher in the container, execute:

docker run -it --rm rlworkgroup/garage-headless python examples/tf/trpo_cartpole.py

This will run the latest image available. To use a stable release such as v2020.06, use rlworkgroup/garage-headless:2020.06.

To run environments using MuJoCo, pass the contents of the MuJoCo key in a variable named MJKEY in the same docker-run command using cat. For example, if your key is at ~/.mujoco/mjkey.txt, execute:

docker run \
  -it \
  --rm \
  -e MJKEY="$(cat ~/.mujoco/mjkey.txt)" \
  rlworkgroup/garage-headless python examples/tf/trpo_swimmer.py

To save the experiment data generated in the container, you need to specify a path where the files will be saved inside your host computer with the argument -v in the docker-run command. For example, if the path you want to use is at /home/tmp/data, execute:

docker run \
  -it \
  --rm \
  -v /home/tmp/data:/home/garage-user/code/garage/data \
  rlworkgroup/garage-headless python examples/tf/trpo_cartpole.py

This binds a volume between your host path and the path in garage at /home/garage-user/code/garage/data.

Note

Make sure the directory at the host path exists and is writable by the current user, otherwise docker will create it with user as root, but the garage container won’t be able to write to it.

Build and run the headless image

To build and run the headless image, first clone the garage repository, move to the root folder of your local repository and then execute;

make run-headless RUN_CMD="python examples/tf/trpo_cartpole.py"

Where RUN_CMD specifies the executable to run in the container.

The previous command adds a volume from the data folder inside your cloned garage repository to the data folder in the garage container, so any experiment results ran in the container will be saved in the data folder inside your cloned repository.

By default, docker generates random names for containers. If you want to specify a name for the container, you can do so with the variable CONTAINER_NAME. As a side effect, this will output the results in data/$CONTAINER_NAME directory instead of the data directory.

make run-headless RUN_CMD="..." CONTAINER_NAME="my_container_123"

This will output results in data/my_container_123 directory.

If you need to use MuJoCo, you need to place your key at ~/.mujoco/mjkey.txt or specify the corresponding path through the MJKEY_PATH variable:

make run-headless RUN_CMD="..." MJKEY_PATH="/home/user/mjkey.txt"

If you require to pass additional arguments to docker build and run commands, you can use the variables BUILD_ARGS and RUN_ARGS, for example:

make run-headless BUILD_ARGS="--build-arg MY_VAR=123" RUN_ARGS="-e MY_VAR=123"

NVIDIA image

The garage NVIDIA images come with CUDA 10.2 and cuDNN (required for tensorflow).

Prerequisites for NVIDIA image

Additional to the prerequisites for the headless image, make sure to have:

Tested on Ubuntu 18.04 & 20.04.

Run a pre-compiled garage-nvidia image

The same commands for the headless image mentioned above apply for the nvidia image, except that the image name is defined by rlworkgroup/garage-nvidia.

For example, to execute a launcher file:

docker run -it --rm rlworkgroup/garage-nvidia python examples/tf/trpo_cartpole.py

Build and run the NVIDIA image

The same rules for the headless image apply here, except that the target name is:

make run-nvidia

This make command builds the NVIDIA image and runs it in a non-headless mode. It will not work on headless machines. You can run the NVIDIA in a headless state using the following target:

make run-nvidia-headless

Expose GPUs for use

By default, garage-nvidia uses all of your gpus. If you want to customize which GPUs are used and/or want to set the GPU capabilities exposed, as described in official docker documentation here, you can pass the desired values to –gpus option using the variable GPUS. For example:

make run-nvidia GPUS="device=0,2"

Using a different driver version

The garage-nvidia docker image uses nvidia/cuda:10.2-cudnn7-runtime-ubuntu18.04 as the parent image which requires NVIDIA driver version 440.33+. If you need to use garage with a different driver version, you might be able to build the garage-nvidia image from scratch using a different parent image using the variable PARENT_IMAGE.

make run-nvidia PARENT_IMAGE="nvidia/cuda:10.1-cudnn7-runtime-ubuntu18.04"

You can find the required parent images at NVIDIA CUDA’s DockerHub

This page was authored by Angel Ivan Gonzalez (@gonzaiva), with contributions from Gitanshu Sardana (@gitanshu).