Status: Archive (code is provided as-is, no updates expected)

gym-recording

A Python package to capture the sequences of actions and observations on a Gym environment by wrapping it in a TraceRecordingWrapper, like this:

import gym
from gym_recording.wrappers import TraceRecordingWrapper

def main():
    env = gym.make('CartPole-v0')
    env = TraceRecordingWrapper(env)
    # ... exercise the environment

It will save recorded traces in a directory, which it will print with logging. You can get the directory name from your code as env.directory.

Later you can play back the recording with code like the following, which runs a callback for every episode.

from gym_recording import playback

def handle_ep(observations, actions, rewards):
  # ... learn a model

playback.scan_recorded_traces(directory, handle_ep)

You can also use the storage_s3 module to upload and download traces from S3, so you it can run across machines.

from gym_recording import playback, storage_s3

def main():
    env = gym.make('CartPole-v0')
    env = TraceRecordingWrapper(env)
    # ... exercise the environment

    s3url = storage_s3.upload_recording(env.directory, env.spec.id, 'openai-example-traces')
    # ... store s3url in a database

    # ... Switch to another machine

    # ... load s3url from a database

    def handle_ep(observations, actions, rewards):
      # ... learn a model
    playback.scan_recorded_traces(storage_s3.download_recording(s3url), handle_ep)