docker.md

Docker

If you prefer not to install dependencies locally, or if you want to train your models on a containerized remote machine, you can use the provided Dockerfile to build an image with all dependencies pre-installed.

The only prerequisites are Docker and, on your deployment machine, the NVIDIA Container Toolkit for GPU support.

To build the Docker image, run the following command from the root of the repository:

docker build -f Dockerfile -t r2dreamer:local .

You can replace the -t argument with any image name you like. The command above will build and tag the image as r2dreamer:local.

Then start a container from the built image with:

docker run -it -d --rm \
    --gpus=all \
    --network=host \
    --volume=$PWD:/workspace \
    --name=r2dreamer-container \
    r2dreamer:local

You can then connect to the running container and execute your training scripts. For example, to run R2-Dreamer on DMC Walker Walk:

# Connect to the running container
docker exec -it r2dreamer-container bash

# And then inside the container:
python3 train.py env=dmc_vision env.task=dmc_walker_walk

# Alternatively, combine it with the docker exec command and use the -d flag to run in detached mode:
docker exec -it -d r2dreamer-container bash -c "python3 train.py env=dmc_vision env.task=dmc_walker_walk"

To monitor training progress with TensorBoard, run the following command in a separate terminal on your host machine:

docker exec -it r2dreamer-container tensorboard --logdir ./logdir

The TensorBoard dashboard will then be available at http://localhost:6006/.

Docker documentation contributed by @MeierTobias.