Hsuanwu: Long-Term Evolution Project of Reinforcement Learning is inspired by the long-term evolution (LTE) standard project in telecommunications, which aims to track the latest research progress in reinforcement learning (RL) and provide stable and efficient baselines. In Hsuanwu, you can find everything you need in RL, such as training, evaluation, deployment, etc. The highlight features of Hsuanwu:

⏱️ Latest algorithms and tricks;
🧱 Highly modularized design for complete decoupling of RL algorithms;
🚀 Optimized workflow for full hardware acceleration;
⚙️ Support for custom environments;
🖥️ Support for multiple computing devices like GPU and NPU;
🛠️ Support for RL model engineering deployment (TensorRT, CANN, ...);
💾 Large number of reusable bechmarks (See HsuanwuHub);
📋 Elegant experimental management powered by Hydra.

Join the developer community for issues and discussions:

Slack	QQ	GitHub

Quick Start
- Installation
- Build your first Hsuanwu application
Implemented Modules
Model Zoo
API Documentation
How To Contribute
Acknowledgment

Quick Start

Installation

Prerequisites

Currently, Hsuanwu requires Python>=3.8, user can create an virtual environment by

conda create -n hsuanwu python=3.8

with pip recommended

Open up a terminal and install Hsuanwu with pip:

pip install hsuanwu # basic installation
pip install hsuanwu[envs] # for pre-defined environments
pip install hsuanwu[tests] # for project tests
pip install hsuanwu[all] # install all the dependencies

with git

Open up a terminal and clone the repository from GitHub with git:

git clone https://github.com/RLE-Foundation/Hsuanwu.git

After that, run the following command to install package and dependencies:

pip install -e . # basic installation
pip install -e .[envs] # for pre-defined environments
pip install -e .[tests] # for project tests
pip install -e .[all] # install all the dependencies

For more detailed installation instruction, see https://docs.hsuanwu.dev/getting_started.

Build your first Hsuanwu application

For example, we want to use DrQ-v2 to solve a task of DeepMind Control Suite, and we only need the following two steps:

Write a config.yaml file in your working directory like:

experiment: drqv2_dmc     # Experiment ID.
device: cuda:0            # Device (cpu, cuda, ...) on which the code should be run.
seed: 1                   # Random seed for reproduction.
num_train_steps: 250000   # Number of training steps.

agent:
  name: DrQv2             # The agent name.

Write a train.py file like:

import hydra # Use Hydra to manage experiments

from hsuanwu.env import make_dmc_env # Import DeepMind Control Suite
from hsuanwu.common.engine import HsuanwuEngine # Import Hsuanwu engine

train_env = make_dmc_env(env_id='cartpole_balance') # Create train env
test_env = make_dmc_env(env_id='cartpole_balance') # Create test env

@hydra.main(version_base=None, config_path='./', config_name='config')
def main(cfgs):
    engine = HsuanwuEngine(cfgs=cfgs, train_env=train_env, test_env=test_env) # Initialize engine
    engine.invoke() # Start training

if __name__ == '__main__':
    main()

Run train.py and you will see the following output:

For more detailed tutorials, see https://docs.hsuanwu.dev/tutorials.

Implemented Modules

Roadmap

Hsuanwu evolves based on reinforcement learning algorithms and integrates latest tricks. The following figure demonstrates the main evolution roadmap of Hsuanwu:

Project Structure

See the project structure below:

Common: Auxiliary modules like trainer and logger.
- Engine: Engine for building Hsuanwu application.
- Logger: Logger for managing output information.
Xploit: Modules that focus on exploitation in RL.
- Encoder: Neural nework-based encoder for processing observations.
- Agent: Agent for interacting and learning.
- Storage: Storage for storing collected experiences.
Xplore: Modules that focus on exploration in RL.
- Augmentation: PyTorch.nn-like modules for observation augmentation.
- Distribution: Distributions for sampling actions.
- Reward: Intrinsic reward modules for enhancing exploration.
Evaluation: Reasonable and reliable metrics for algorithm evaluation.
Env: Packaged environments (e.g., Atari games) for fast invocation.
Pre-training: Methods of pre-training in RL.
Deployment: Methods of model deployment in RL.

For more detiled descriptions of these modules, see https://docs.hsuanwu.dev/api

RL Agents

Module	Recurrent	Box	Discrete	MultiBinary	Multi Processing	NPU	Paper	Citations
SAC	❌	✔️	❌	❌	❌	🐌	Link	5077⭐
DrQ	❌	✔️	❌	❌	❌	🐌	Link	433⭐
DDPG	❌	✔️	❌	❌	❌	🐌	Link	11819⭐
DrQ-v2	❌	✔️	❌	❌	❌	🐌	Link	100⭐
PPO	❌	✔️	✔️	✔️	✔️	🐌	Link	11155⭐
DAAC	❌	✔️	✔️	✔️	✔️	🐌	Link	56⭐
DrAC	❌	✔️	✔️	✔️	✔️	🐌	Link	29⭐
PPG	❌	✔️	✔️	❌	✔️	🐌	Link	82⭐
IMPALA	✔️	✔️	✔️	❌	✔️	🐌	Link	1219⭐

🐌: Developing.

NPU: Support Neural-network processing unit.

Recurrent: Support recurrent neural network.

Box: A N-dimensional box that containes every point in the action space.

Discrete: A list of possible actions, where each timestep only one of the actions can be used.

MultiBinary: A list of possible actions, where each timestep any of the actions can be used in any combination.

Intrinsic Reward Modules

Module	Remark	Repr.	Visual	Reference
PseudoCounts	Count-Based exploration	✔️	✔️	Never Give Up: Learning Directed Exploration Strategies
ICM	Curiosity-driven exploration	✔️	✔️	Curiosity-Driven Exploration by Self-Supervised Prediction
RND	Count-based exploration	❌	✔️	Exploration by Random Network Distillation
GIRM	Curiosity-driven exploration	✔️	✔️	Intrinsic Reward Driven Imitation Learning via Generative Model
NGU	Memory-based exploration	✔️	✔️	Never Give Up: Learning Directed Exploration Strategies
RIDE	Procedurally-generated environment	✔️	✔️	RIDE: Rewarding Impact-Driven Exploration for Procedurally-Generated Environments
RE3	Entropy Maximization	❌	✔️	State Entropy Maximization with Random Encoders for Efficient Exploration
RISE	Entropy Maximization	❌	✔️	Rényi State Entropy Maximization for Exploration Acceleration in Reinforcement Learning
REVD	Divergence Maximization	❌	✔️	Rewarding Episodic Visitation Discrepancy for Exploration in Reinforcement Learning

🐌: Developing.

Repr.: The method involves representation learning.

Visual: The method works well in visual RL.

See Tutorials: Use intrinsic reward and observation augmentation for usage examples.

Model Zoo

Hsuanwu provides a large number of reusable bechmarks, see https://hub.hsuanwu.dev/ and https://docs.hsuanwu.dev/benchmarks/

API Documentation

View our well-designed documentation: https://docs.hsuanwu.dev/

How To Contribute

Welcome to contribute to this project! Before you begin writing code, please read CONTRIBUTING.md for guide first.

Acknowledgment

This project is supported by FUNDING.yml. Some code of this project is borrowed or inspired by several excellent projects, and we highly appreciate them. See ACKNOWLEDGMENT.md.

CndAir / Hsuanwu