Rainbow DQN implementation that outperforms the paper's results on 40% of games using 20x less data 🌈

Overview

Rainbow 🌈

An implementation of Rainbow DQN which outperforms the paper's (Hessel et al. 2017) results on 40% of tested games while using 20x less data. This was developed as part of an undergraduate university course on scientific research and writing. The results are also available as a spreadsheet here. A selection of videos is available here.

Key Changes and Results

  • We implemented the large IMPALA CNN with 2x channels from Espeholt et al. (2018).
  • The implementation uses large, vectorized environments, asynchronous environment interaction, mixed-precision training, and larger batch sizes to reduce training time.
  • Integrations and recommended preprocessing for >1000 environments from gym, gym-retro and procgen are provided.
  • Due to compute and time constraints, we only trained for 10M frames (compared to 200M in the paper).
  • We implemented all components apart from distributional RL (we saw mixed results with C51 and QR-DQN).

When trained for only 10M frames, this implementation outperforms:

google/dopamine trained for 10M frames on 96% of games
google/dopamine trained for 200M frames on 64% of games
Hessel, et al. (2017) trained for 200M frames on 40% of games
Human results on 72% of games

Most of the observed performance improvements compared to the paper come from switching to the IMPALA CNN as well as some hyperparameter changes (e.g. the 4x larger learning rate).

Setup

Install necessary prerequisites with

sudo apt install zlib1g-dev cmake unrar
pip install wandb gym[atari]==0.18.0 imageio moviepy torchsummary tqdm rich procgen gym-retro torch stable_baselines3 atari_py==0.2.9

If you intend to use gym Atari games, you will need to install these separately, e.g., by running:

wget http://www.atarimania.com/roms/Roms.rar 
unrar x Roms.rar
python -m atari_py.import_roms .

To set up gym-retro games you should follow the instructions here.

How to use

To get started right away, run

python train_rainbow.py --env_name gym:Qbert

This will train Rainbow on Atari Qbert and log all results to "Weights and Biases" and the checkpoints directory.

Please take a look at common/argp.py or run python train_rainbow.py --help for more configuration options.

Some Notes

  • With a single RTX 2080 and 12 CPU cores, training for 10M frames takes around 8-12 hours, depending on the used settings
  • About 15GB of RAM are required. When using a larger replay buffer or subprocess envs, memory use may be much higher
  • Hyperparameters can be configured through command line arguments; defaults can be found in common/argp.py
  • For fastest training throughput use batch_size=512, parallel_envs=64, train_count=1, subproc_vecenv=True

Acknowledgements

We are very grateful to the TU Wien DataLab for providing the majority of the compute resources that were necessary to perform the experiments.

Here are some other implementations and resources that were helpful in the completion of this project:

Owner
Dominik Schmidt
I'm a computer science & math student at the Vienna University of Technology in Austria.
Dominik Schmidt
This repository contains answers of the Shopify Summer 2022 Data Science Intern Challenge.

Data-Science-Intern-Challenge This repository contains answers of the Shopify Summer 2022 Data Science Intern Challenge. Summer 2022 Data Science Inte

1 Jan 11, 2022
Sequential model-based optimization with a `scipy.optimize` interface

Scikit-Optimize Scikit-Optimize, or skopt, is a simple and efficient library to minimize (very) expensive and noisy black-box functions. It implements

Scikit-Optimize 2.5k Jan 04, 2023
This is the workbook I created while I was studying for the Qiskit Associate Developer exam. I hope this becomes useful to others as it was for me :)

A Workbook for the Qiskit Developer Certification Exam Hello everyone! This is Bartu, a fellow Qiskitter. I have recently taken the Certification exam

Bartu Bisgin 66 Dec 10, 2022
Official Pytorch and JAX implementation of "Efficient-VDVAE: Less is more"

The Official Pytorch and JAX implementation of "Efficient-VDVAE: Less is more" Arxiv preprint Louay Hazami   ·   Rayhane Mama   ·   Ragavan Thurairatn

Rayhane Mama 144 Dec 23, 2022
U-Time: A Fully Convolutional Network for Time Series Segmentation

U-Time & U-Sleep Official implementation of The U-Time [1] model for general-purpose time-series segmentation. The U-Sleep [2] model for resilient hig

Mathias Perslev 176 Dec 19, 2022
UniLM AI - Large-scale Self-supervised Pre-training across Tasks, Languages, and Modalities

Pre-trained (foundation) models across tasks (understanding, generation and translation), languages (100+ languages), and modalities (language, image, audio, vision + language, audio + language, etc.

Microsoft 7.6k Jan 01, 2023
Models, datasets and tools for Facial keypoints detection

Template for Data Science Project This repo aims to give a robust starting point to any Data Science related project. It contains readymade tools setu

girafe.ai 1 Feb 11, 2022
[CVPR 2021] Region-aware Adaptive Instance Normalization for Image Harmonization

RainNet — Official Pytorch Implementation Region-aware Adaptive Instance Normalization for Image Harmonization Jun Ling, Han Xue, Li Song*, Rong Xie,

130 Dec 11, 2022
Real-time multi-object tracker using YOLO v5 and deep sort

This repository contains a two-stage-tracker. The detections generated by YOLOv5, a family of object detection architectures and models pretrained on the COCO dataset, are passed to a Deep Sort algor

Mike 3.6k Jan 05, 2023
4th place solution to datafactory challenge by Intermarché.

Solution to Datafactory challenge by Intermarché. 4th place solution to datafactory challenge by Intermarché. The objective of the challenge is to pre

Raphael Sourty 11 Mar 19, 2022
Learning from Synthetic Humans, CVPR 2017

Learning from Synthetic Humans (SURREAL) Gül Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J. Black, Ivan Laptev and Cordelia Schmid,

Gul Varol 538 Dec 18, 2022
Graph Analysis From Scratch

Graph Analysis From Scratch Goal In this notebook we wanted to implement some functionalities to analyze a weighted graph only by using algorithms imp

Arturo Ghinassi 0 Sep 17, 2022
A PyTorch implementation for V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation

A PyTorch implementation of V-Net Vnet is a PyTorch implementation of the paper V-Net: Fully Convolutional Neural Networks for Volumetric Medical Imag

Matthew Macy 606 Dec 21, 2022
Official Pytorch implementation of Meta Internal Learning

Official Pytorch implementation of Meta Internal Learning

10 Aug 24, 2022
Analyses of the individual electric field magnitudes with Roast.

Aloi Davide - PhD Student (UoB) Analysis of electric field magnitudes (wp2a dataset only at the moment) and correlation analysis with Dynamic Causal M

Davide Aloi 7 Dec 15, 2022
VISSL is FAIR's library of extensible, modular and scalable components for SOTA Self-Supervised Learning with images.

What's New Below we share, in reverse chronological order, the updates and new releases in VISSL. All VISSL releases are available here. [Oct 2021]: V

Meta Research 2.9k Jan 07, 2023
An efficient 3D semantic segmentation framework for Urban-scale point clouds like SensatUrban, Campus3D, etc.

An efficient 3D semantic segmentation framework for Urban-scale point clouds like SensatUrban, Campus3D, etc.

Zou 33 Jan 03, 2023
Code for our CVPR 2021 paper "MetaCam+DSCE"

Joint Noise-Tolerant Learning and Meta Camera Shift Adaptation for Unsupervised Person Re-Identification (CVPR'21) Introduction Code for our CVPR 2021

FlyingRoastDuck 59 Oct 31, 2022
Code for the paper BERT might be Overkill: A Tiny but Effective Biomedical Entity Linker based on Residual Convolutional Neural Networks

Biomedical Entity Linking This repo provides the code for the paper BERT might be Overkill: A Tiny but Effective Biomedical Entity Linker based on Res

Tuan Manh Lai 24 Oct 24, 2022
We propose a new method for effective shadow removal by regarding it as an exposure fusion problem.

Auto-exposure fusion for single-image shadow removal We propose a new method for effective shadow removal by regarding it as an exposure fusion proble

Qing Guo 146 Dec 31, 2022