paulfenton / rllib-gridworld-experiments Goto Github PK

Repository for holding environments and models for gridworld experiments with rllib, and deploying them with Ray & Tune on Google Cloud.

• pyvirtualenv
• google-api-python-client
• Service account keys for a GCP project: (https://cloud.google.com/iam/docs/creating-managing-service-account-keys)
  • Save the credentials file inside secrets/.
• Create an .env file as follows, substituting values between <> with your own:

  GOOGLE_REGION=us-west1
  GOOGLE_AVAILABILITY_ZONE=us-west1-a
  GOOGLE_PROJECT_ID=< your project id >
  GOOGLE_APPLICATION_CREDENTIALS=./secrets/< name of downloaded credentials.json >

source .env  # load environment variables
envsubst < cluster.yaml.tmpl > cluster.yaml  # interpolate placeholders in cluster config template
ray up cluster.yaml

View the cluster's dashboard:

ray dashboard cluster.yaml

TODO: Add instructions for usage without development-mode

Use python env from sibling ray directory (with dev symlinks)

export RAY_DEV_VENV='../ray/venv'
ln -s $RAY_DEV_VENV venv
source venv/bin/activate

Run the notebook run_experiment.ipynb.

ray down cluster.yaml