OpenAi's gym environment wrapper to vectorize them with Ray

Last update: Nov 10, 2022

Overview

Ray Vector Environment Wrapper

You would like to use Ray to vectorize your environment but you don't want to use RLLib ?
You came to the right place !

This package allows you to parallelize your environment using Ray
Not only does it allows to run environments in parallel, but it also permits to run multiple sequential environments on each worker
For example, you can run 80 workers in parallel, each running 10 sequential environments for a total of 80 * 10 environments
This can be useful if your environment is fast and simply running 1 environment per worker leads to too much communication overhead between workers

Installation

pip install RayEnvWrapper

If something went wrong, it most certainly is because of Ray
For example, you might have issue installing Ray on Apple Silicon (i.e., M1) laptop. See Ray's documentation for a simple fix
At the moment Ray does not support Python 3.10. This package has been tested with Python 3.9.

How does it work?

You first need to define a function that seed and return your environment:

Here is an example for CartPole:

import gym

def make_and_seed(seed: int) -> gym.Env:
    env = gym.make('CartPole-v0')
    env = gym.wrappers.RecordEpisodeStatistics(env) # you can put extra wrapper to your original environment
    env.seed(seed)
    return env

Note: If you don't want to seed your environment, simply return it without using the seed, but the function you define needs to take a number as an input

Then, call the wrapper to create and wrap all the vectorized environment:

from RayEnvWrapper import WrapperRayVecEnv

number_of_workers = 4 # Usually, this is set to the number of CPUs in your machine
envs_per_worker = 2

vec_env = WrapperRayVecEnv(make_and_seed, number_of_workers, envs_per_worker)

You can then use your environment. All the output for each of the environments are stacked in a numpy array

Reset:

vec_env.reset()

Output

[[ 0.03073904  0.00145001 -0.03088818 -0.03131252]
 [ 0.03073904  0.00145001 -0.03088818 -0.03131252]
 [ 0.02281231 -0.02475473  0.02306162  0.02072129]
 [ 0.02281231 -0.02475473  0.02306162  0.02072129]
 [-0.03742824 -0.02316945  0.0148571   0.0296055 ]
 [-0.03742824 -0.02316945  0.0148571   0.0296055 ]
 [-0.0224773   0.04186813 -0.01038048  0.03759079]
 [-0.0224773   0.04186813 -0.01038048  0.03759079]]

The i-th entry represent the initial observation of the i-th environment
Note: As environments are vectorized, you don't need explicitly to reset the environment at the end of the episode, it is done automatically However, you need to do it once at the beginning

Take a random action:

vec_env.step([vec_env.action_space.sample() for _ in range(number_of_workers * envs_per_worker)])

Notice how the actions are passed. We pass an array containing an action for each of the environments
Thus, the array is of size number_of_workers * envs_per_worker (i.e., the total number of environments)

Output

(array([[ 0.03076804, -0.19321568, -0.03151444,  0.25146705],
       [ 0.03076804, -0.19321568, -0.03151444,  0.25146705],
       [ 0.02231721, -0.22019969,  0.02347605,  0.3205903 ],
       [ 0.02231721, -0.22019969,  0.02347605,  0.3205903 ],
       [-0.03789163, -0.21850128,  0.01544921,  0.32693872],
       [-0.03789163, -0.21850128,  0.01544921,  0.32693872],
       [-0.02163994, -0.15310344, -0.00962866,  0.3269806 ],
       [-0.02163994, -0.15310344, -0.00962866,  0.3269806 ]],
      dtype=float32), 
 array([1., 1., 1., 1., 1., 1., 1., 1.], dtype=float32), 
 array([False, False, False, False, False, False, False, False]), 
 [{}, {}, {}, {}, {}, {}, {}, {}])

As usual, the step method returns a tuple, except that here both the observation, reward, dones and infos are concatenated
In this specific example, we have 2 environments per worker.
Index 0 and 1 are environments from worker 1; index 1 and 2 are environments from worker 2, etc.

License

Apache License 2.0

You might also like...

Comments

envs_per_worker
Hi！@ingambe. Thank you very much for your work! I have some questions. What does the "worker and envs" mean here? My understanding is as follows:

Worker represents a process. Two env in a worker belong to two threads.

I don't know if I understand this correctly. Thanks!
opened by Meta-YZ 2
how to wrap two DIFFERENT environments?

Thank you for upload the package. My question is is there a way to stack different environments together? For example I have ten or hundreds different race track environments and I want to train an agent simultaneously drive through this vectorized environment. In stable baseline I can stack them together and train a vectorized environment. Now I want to move to ray and try to speed up the training by using multiple gpu...but so far didn't figure out how to do this. Thanks in advance
enhancement

opened by superfan123 1

OpenAi's gym environment wrapper to vectorize them with Ray

Related tags

Overview

Ray Vector Environment Wrapper

Installation

How does it work?

License

You might also like...

A "gym" style toolkit for building lightweight Neural Architecture Search systems

Customizable RecSys Simulator for OpenAI Gym

Robot Servers and Server Manager software for robo-gym

Deep Q Learning with OpenAI Gym and Pokemon Showdown

Manipulation OpenAI Gym environments to simulate robots at the STARS lab

AI virtual gym is an AI program which can be used to exercise and can be used to see if we are doing the exercises

Multi-objective gym environments for reinforcement learning.

Pytorch Lightning Distributed Accelerators using Ray

Pytorch Lightning Distributed Accelerators using Ray

Comments

envs_per_worker

how to wrap two DIFFERENT environments?

Releases(v1.0)

v1.0(Jan 22, 2022)

Owner

Pierre TASSEL

基于AlphaPose的TensorRT加速

LIVECell - A large-scale dataset for label-free live cell segmentation

A PyTorch implementation of the continual learning experiments with deep neural networks

A graph-to-sequence model for one-step retrosynthesis and reaction outcome prediction.

PyTorch code of paper "LiVLR: A Lightweight Visual-Linguistic Reasoning Framework for Video Question Answering"

Neural Reprojection Error: Merging Feature Learning and Camera Pose Estimation

Official PyTorch implementation of the paper "Recycling Discriminator: Towards Opinion-Unaware Image Quality Assessment Using Wasserstein GAN", accepted to ACM MM 2021 BNI Track.

Multiple paper open-source codes of the Microsoft Research Asia DKI group

Train CPPNs as a Generative Model, using Generative Adversarial Networks and Variational Autoencoder techniques to produce high resolution images.

implementation for paper "ShelfNet for fast semantic segmentation"

Implementation of [Time in a Box: Advancing Knowledge Graph Completion with Temporal Scopes].

This repository is an implementation of our NeurIPS 2021 paper (Stylized Dialogue Generation with Multi-Pass Dual Learning) in PyTorch.

Anime Face Detector using mmdet and mmpose

The Implicit Bias of Gradient Descent on Generalized Gated Linear Networks

To model the probability of a soccer coach leave his/her team during Campeonato Brasileiro for 10 chosen teams and considering years 2018, 2019 and 2020.

Jittor implementation of Recursive-NeRF: An Efficient and Dynamically Growing NeRF

Pretrained Pytorch face detection (MTCNN) and recognition (InceptionResnet) models

The Easy-to-use Dialogue Response Selection Toolkit for Researchers

Self-Supervised Pillar Motion Learning for Autonomous Driving (CVPR 2021)

PyTorch implementations of the paper: "DR.VIC: Decomposition and Reasoning for Video Individual Counting, CVPR, 2022"