d3rlpy: An offline deep reinforcement learning library

d3rlpy is an offline deep reinforcement learning library for practitioners and researchers.

import d3rlpy

dataset, env = d3rlpy.datasets.get_dataset("hopper-medium-v0")

# prepare algorithm
sac = d3rlpy.algos.SAC()

# train offline
sac.fit(dataset, n_steps=1000000)

# train online
sac.fit_online(env, n_steps=1000000)

# ready to control
actions = sac.predict(x)

• Documentation: https://d3rlpy.readthedocs.io
• Paper: https://arxiv.org/abs/2111.03788

⚠️ Next generation of d3rlpy is under development in nightly branch. It's not fully supporting all of features in the stable versions and there is no documentation yet. However, if you're interested, you can sneak to see what's there.

key features

⚡ Most Practical RL Library Ever

• offline RL: d3rlpy supports state-of-the-art offline RL algorithms. Offline RL is extremely powerful when the online interaction is not feasible during training (e.g. robotics, medical).
• online RL: d3rlpy also supports conventional state-of-the-art online training algorithms without any compromising, which means that you can solve any kinds of RL problems only with d3rlpy.
• advanced engineering: d3rlpy is designed to implement the faster and efficient training algorithms. For example, you can train Atari environments with x4 less memory space and as fast as the fastest RL library.

🔰 User-friendly API

• zero-knowledge of DL library: d3rlpy provides many state-of-the-art algorithms through intuitive APIs. You can become a RL engineer even without knowing how to use deep learning libraries.
• extensive documentation: d3rlpy is fully documented and accompanied with tutorials and reproduction scripts of the original papers.

🚀 Beyond State-of-the-art

• distributional Q function: d3rlpy is the first library that supports distributional Q functions in the all algorithms. The distributional Q function is known as the very powerful method to achieve the state-of-the-performance.
• many tweek options: d3rlpy is also the first to support N-step TD backup and ensemble value functions in the all algorithms, which lead you to the place no one ever reached yet.

installation

d3rlpy supports Linux, macOS and Windows.

PyPI (recommended)

$ pip install d3rlpy

Anaconda

$ conda install -c conda-forge d3rlpy

Docker

$ docker run -it --gpus all --name d3rlpy takuseno/d3rlpy:latest bash

supported algorithms

algorithm	discrete control	continuous control	offline RL?
Behavior Cloning (supervised learning)	✅	✅
Neural Fitted Q Iteration (NFQ)	✅	⛔	✅
Deep Q-Network (DQN)	✅	⛔
Double DQN	✅	⛔
Deep Deterministic Policy Gradients (DDPG)	⛔	✅
Twin Delayed Deep Deterministic Policy Gradients (TD3)	⛔	✅
Soft Actor-Critic (SAC)	✅	✅
Batch Constrained Q-learning (BCQ)	✅	✅	✅
Bootstrapping Error Accumulation Reduction (BEAR)	⛔	✅	✅
Conservative Q-Learning (CQL)	✅	✅	✅
Advantage Weighted Actor-Critic (AWAC)	⛔	✅	✅
Critic Reguralized Regression (CRR)	⛔	✅	✅
Policy in Latent Action Space (PLAS)	⛔	✅	✅
TD3+BC	⛔	✅	✅
Implicit Q-Learning (IQL)	⛔	✅	✅

supported Q functions

• standard Q function
• Quantile Regression
• Implicit Quantile Network

benchmark results

d3rlpy is benchmarked to ensure the implementation quality. The benchmark scripts are available reproductions directory. The benchmark results are available d3rlpy-benchmarks repository.

examples

MuJoCo

import d3rlpy

# prepare dataset
dataset, env = d3rlpy.datasets.get_d4rl('hopper-medium-v0')

# prepare algorithm
cql = d3rlpy.algos.CQL(use_gpu=True)

# train
cql.fit(
    dataset,
    eval_episodes=dataset,
    n_epochs=100,
    scorers={
        'environment': d3rlpy.metrics.evaluate_on_environment(env),
        'td_error': d3rlpy.metrics.td_error_scorer,
    },
)

See more datasets at d4rl.

Atari 2600

import d3rlpy
from sklearn.model_selection import train_test_split

# prepare dataset
dataset, env = d3rlpy.datasets.get_atari('breakout-expert-v0')

# split dataset
train_episodes, test_episodes = train_test_split(dataset, test_size=0.1)

# prepare algorithm
cql = d3rlpy.algos.DiscreteCQL(
    n_frames=4,
    q_func_factory='qr',
    scaler='pixel',
    use_gpu=True,
)

# start training
cql.fit(
    train_episodes,
    eval_episodes=test_episodes,
    n_epochs=100,
    scorers={
        'environment': d3rlpy.metrics.evaluate_on_environment(env),
        'td_error': d3rlpy.metrics.td_error_scorer,
    },
)

See more Atari datasets at d4rl-atari.

Online Training

import d3rlpy
import gym

# prepare environment
env = gym.make('Hopper-v3')
eval_env = gym.make('Hopper-v3')

# prepare algorithm
sac = d3rlpy.algos.SAC(use_gpu=True)

# prepare replay buffer
buffer = d3rlpy.online.buffers.ReplayBuffer(maxlen=1000000, env=env)

# start training
sac.fit_online(env, buffer, n_steps=1000000, eval_env=eval_env)

tutorials

Try cartpole examples on Google Colaboratory!

• offline RL tutorial:
• online RL tutorial:

More tutorial documentations are available here.

contributions

Any kind of contribution to d3rlpy would be highly appreciated! Please check the contribution guide.

The release planning can be checked at milestones.

community

Channel	Link
Chat	Gitter
Issues	GitHub Issues

family projects

Project	Description
d4rl-atari	A d4rl-style library of Google's Atari 2600 datasets

roadmap

The roadmap to the future release is available in ROADMAP.md.

citation

The paper is available here.

@article{d3rlpy,
  author  = {Takuma Seno and Michita Imai},
  title   = {d3rlpy: An Offline Deep Reinforcement Learning Library},
  journal = {Journal of Machine Learning Research},
  year    = {2022},
  volume  = {23},
  number  = {315},
  pages   = {1--20},
  url     = {http://jmlr.org/papers/v23/22-0017.html}
}

acknowledgement

This work is supported by Information-technology Promotion Agency, Japan (IPA), Exploratory IT Human Resources Project (MITOU Program) in the fiscal year 2020.