ml-agents/docs/Using-Docker.md

Using Docker For ML Agents (Experimental)

We are currently offering an experimental solution for Windows and Mac users who would like to do training or inference using Docker. This option may be appealing to users who would like to avoid dealing with Python and TensorFlow installation on their host machines. This setup currently forces both TensorFlow and Unity to rely on only the CPU for computation purposes. As such, we currently only support training using environments that only contain agents which use vector observations, rather than camera-based visual observations. For example, the GridWorld environment which use visual observations for training is not supported.

Requirements

• Unity Linux Standalone Player (Link)
• Docker (Link)

Setup

• Install Docker (see link above) if you don't have it setup on your machine.

• Since Docker runs a container in an environment that is isolated from the host machine, we will be using a mounted directory, e.g. unity-volume in your host machine in order to share data, e.g. the Unity executable, curriculum files and tensorflow graph.

Usage

• Docker typically runs a container sharing a (linux) kernel with the host machine, this means that the Unity environment has to be built for the linux platform. From the Build Settings Window, please select the architecture to be x86_64 and choose the build to be headless (This is important because we are running it in a container that does not have graphics drivers installed). Save the generated environment in the directory to be mounted (e.g. we have conveniently created an empty directory called at the top level unity-volume).

• Ensure that unity-volume/<environment-name>.x86_64 and unity-volume/environment-name_Data. So for example, <environment_name> might be 3Dball and you might want to ensure that unity-volume/3Dball.x86_64 and unity-volume/3Dball_Data are both present in the directory unity-volume.

• Make sure the docker engine is running on your machine, then build the docker container by running docker build -t <image_name> . . in the top level of the source directory. Replace <image_name> by the name of the image that you want to use, e.g. balance.ball.v0.1.

• Run the container:

docker run --mount type=bind,source="$(pwd)"/unity-volume,target=/unity-volume \
	 <tag-name>:latest <environment-name> \
	 --docker-target-name=unity-volume \
	 --train --run-id=<run-id>

For the 3DBall environment, for example this would be:

• Run the container:

docker run --mount type=bind,source="$(pwd)"/unity-volume,target=/unity-volume \
	 balance.ball.v0.1:latest 3Dball \
	 --docker-target-name=unity-volume \
	 --train --run-id=<run-id>

Notes on argument values

• source : Reference to the path in your host OS where you will store the Unity executable.
• target: Tells docker to mount the source path as a disk with this name.
• docker-target-name: Tells the ML-Agents python package what the name of the disk where it can read the Unity executable and store the graph.*This should therefore be identical to the target.
• train, run-id: ML-Agents arguments passed to learn.py. train trains the algorithm, run-id is used to tag each experiment with a unique id.

For more details on docker mounts, look at these docs from Docker.