gym-recording's Introduction

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)

gym-recording's Issues

The package last pushed by @tlbtlbtlb on https://pypi.python.org/pypi/gym_recording/0.0.1 is from 20/10/2016.
It lacks the 2a979d4 commit that actually installs the packages. Can someone please push the latest version?
Thanks!

P.S. After reviewing #4 and #5

ModuleNotFoundError: No module named 'gym_recording'

Hi, I need to record my Gym environment so I've put on my code:

import gym
from gym_recording.wrappers import TraceRecordingWrapper

When I run my code, I get: ModuleNotFoundError: No module named 'gym_recording'

I've done from prompt: pip install gym_recording

Everything goes well but when I run the code, I still get the same error.

What should I do?

Segmentation fault with playback.scan_recorded_traces

The following code results in a segmentation fault:

ep = {}
def handle_ep(o, a, r):
  ep['r'] = r

playback.scan_recorded_traces(directory)

print(len(ep['r'])) # OK
print(ep['r']) # Segfault

However, when I just copy over the body from playback.scan_recorded_traces and run this code without closing the reader first, it works just fine. The segfault results from closing the reader (i.e. its open binary files).

Edit: As I suspected, after reading around a bit this seems intended behaviour as the numpy arrays map directly into mmaps. Copying the arrays is sufficient to keep them around after closing the reader. I will leave this issue here for someone else to close, just in case this is not entirely intended.

Does this work with SubprocVecEnv

The openai baselines create multiple stacked environments for training a model and use the wrap_deepmind function. Is it possible to use the gym-recording module with the models created using the openai baselines?

Recommend Projects