So during my downtime (now officially cancer-free btw!) I’ve gotten more involved with the project to build a functional local training environment and have also taken on a role as a GitHub maintainer for various AWS ML community projects.
We’ve had local training on the 2019 setup for some time now largely thanks to the awesome work done last year by Chris Rhodes and Alex Schultz but the new 2020 environment is a totally new beast with exciting new racing modes including obstacle avoidence and head-to-head racing. Not only that there is a new car with rear-facing lidar and stereo front-facing cameras!
Under the hood there have been huge changes to the way that the training environment runs which means we basically needed a complete rebuild.
For the last few weeks a small group of AWS ML Community members have taken on the task of rebuilding many of the sub-components that make up a local training stack. Particular kudos to Lars Lorentz Ludvigsen and Richard Fan who (along with myself) have now released updated Sagemaker and Robomaker community docker images utilising the most up to date AWS simulation bundles.
In addition I am now releasing the first version of my 2020 DeepRacer local training environment which uses the previously mentioned images.
Currently things are generally optimised for nVidia GPU’s on linux hosts but we have successfully tested both on local machines and AWS EC2 instances (although LarsLL’s Deepracer for Cloud is probably an easier way to train on an EC2 instance today).
This is still an early release and should be viewed as a beta at best, but it should be enough to enable local training for those of us who want more control over the internals of DeepRacer training. Future updates will hopefully include a web UI and a batch training management system.
Some documentation is available in the README.md however additional support is usually available from the many talented people in
#dr-local-training-setup on the AWS Machine Learning Community Slack at https://deepracing.io