The main purpose of this repo is to document how I was able to set up an EC2 instance through a JPMC sandbox and train model for the AWS DeepRacer competitions. I utilized the LarsLL deepracer for cloud github repo for training. By training on EC2 instances, my team and I have been able to win in the Houston Races, JPMC Global Races, Global Financial Services Races, and we made it to top 8 racers in the world at AWS reInvent 2020.
Training on an EC2 has many advantages:
- Being able to set up a customized action space
- Train much faster with up to triple the number of workers on a g4dn.4xl instance
- Ability to increment your training
- Improved log analysis tools
- Reduced cost: $1.25/hour cost of training versus $3.50/hour on amazon console
If you have any issues getting stuck please reach out to Tyler Wooten in slack (https://jpmc-deepracer.slack.com/archives/C01DC150R29) or via email [email protected].
Other useful links:
- Track names for DR_WORLD_NAME: https://github.com/aws-deepracer-community/deepracer-simapp/tree/master/bundle/deepracer_simulation_environment/share/deepracer_simulation_environment/routes
- Racing types (head to head, time trial, object avoidance) for DR_RACE_TYPE: https://aws-deepracer-community.github.io/deepracer-for-cloud/reference.html
- Pull new sagemaker/robomaker docker images: https://github.com/aws-deepracer-community/deepracer-simapp