AutoCast Container

A containerized dev/test environment for AutoCast

What does it include?

• Ubuntu 20.04 base image
• code-server (on port 9001 by default) for web-based file editing and terminal access
• Everything needed to run AutoCast, in /opt/autocast
• GPU-enabled PyTorch and Minkowski Engine
• CARLA 0.9.11 (note: running CARLA itself in a container can be difficult due to X server issues. For this reason, we recommend running CARLA natively outside of the container)

Requirements

• Docker Engine
• NVIDIA container runtime and a compatible NVIDIA GPU. We've tested this on a GTX 1080Ti and a RTX 3070.

How to run it?

Instant images (just add hot water!)

If you just want to quickly run the image, there's a public pre-built release available here. To pull, simply run:

docker pull quay.io/adrlab/autocast:latest
docker run --gpus=all quay.io/adrlab/autocast:latest

Building the image yourself from scratch

Make sure you have Docker and the NVIDIA container runtime installed. Note that you'll also need to set Docker's default runtime to nvidia for the build to work (the build requires an active NVIDIA GPU).

Your /etc/docker/daemon.json should look somewhat like this:

{
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
         } 
    },
    "default-runtime": "nvidia" 
}

(Make sure to restart the Docker daemon after modifying the config so that it takes effect!)

After that, simply run docker build . from the root directory of this repository, and if all goes well, ~30 minutes later you'll get your very own AutoCast container image.