Terraform & Packer configuration to bring up agents to train against OpenAI's Universe
HCL
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
.gitignore
README.md
single_host.json
single_host.tf

README.md

Quick infra for OpenAI Universe agents

This is a collection of Terraform & Packer files to create and manage running agents against OpenAI's Universe.

Why?

Building an RL agent is tough enough without having to deal with the infra.

Setup

NOTE: This is not super useful yet. Right now, it's just a faster way to set up a single host to run experiments on in EC2.

Prerequsites: Have Terraform, Packer installed. Have an AWS account.

  1. Create an AWS IAM role for Terraform. I gave it full admin rights, YMMV.
  2. Put the key & secret into single_host.json & terraform.tfvars.
  3. Bake an AMI using packer build single_host.json, take the AMI ID from that and put it into single_host.tf (and edit the subnet_id, key_name, and instance_type to appropriate values as well.)
  4. Run terraform apply to create a host.

At the end, you should have a host, suitable for ssh'ing into, and for example running the Universe Starter Agent on.

Teardown

  1. Run terraform destroy & confirm to destroy the host.

TODO

  • Add VPC, subnet & key setup to TF so that it handles that case.
  • Set up Kubernetes in TF so that it can be used as a job processor.
  • Set up Tensorflow / Tensorboard better.
  • Archive/preserve data when done / scaling down hosts.