Updated on 2024.12.23 DI-engine-v0.5.3

Documentation | 中文文档 | Tutorials | Feature | Task & Middleware | TreeTensor | Roadmap

DI-engine is a generalized decision intelligence engine for PyTorch and JAX.

It provides python-first and asynchronous-native task and middleware abstractions, and modularly integrates several of the most important decision-making concepts: Env, Policy and Model. Based on the above mechanisms, DI-engine supports various deep reinforcement learning algorithms with superior performance, high efficiency, well-organized documentation and unittest:

• Most basic DRL algorithms: such as DQN, Rainbow, PPO, TD3, SAC, R2D2, IMPALA
• Multi-agent RL algorithms: such as QMIX, WQMIX, MAPPO, HAPPO, ACE
• Imitation learning algorithms (BC/IRL/GAIL): such as GAIL, SQIL, Guided Cost Learning, Implicit BC
• Offline RL algorithms: BCQ, CQL, TD3BC, Decision Transformer, EDAC, Diffuser, Decision Diffuser, SO2
• Model-based RL algorithms: SVG, STEVE, MBPO, DDPPO, DreamerV3
• Exploration algorithms: HER, RND, ICM, NGU
• LLM + RL Algorithms: PPO-max, DPO, PromptPG, PromptAWR
• Other algorithms: such as PER, PLR, PCGrad
• MCTS + RL algorithms: AlphaZero, MuZero, please refer to LightZero
• Generative Model + RL algorithms: Diffusion-QL, QGPO, SRPO, please refer to GenerativeRL

DI-engine aims to standardize different Decision Intelligence environments and applications, supporting both academic research and prototype applications. Various training pipelines and customized decision AI applications are also supported:

On the low-level end, DI-engine comes with a set of highly re-usable modules, including RL optimization functions, PyTorch utilities and auxiliary tools.

BTW, DI-engine also has some special system optimization and design for efficient and robust large-scale RL training:

Have fun with exploration and exploitation.

• Introduction to DI-engine
• Outline
• Installation
• Quick Start
• Feature
  • Algorithm Versatility
  • Environment Versatility
  • General Data Container: TreeTensor
• Feedback and Contribution
• Supporters
  • ↳ Stargazers
  • ↳ Forkers
• Citation
• License

You can simply install DI-engine from PyPI with the following command:

pip install DI-engine

For more information about installation, you can refer to installation.

And our dockerhub repo can be found here，we prepare base image and env image with common RL environments.

The detailed documentation are hosted on doc | 中文文档.

3 Minutes Kickoff

3 Minutes Kickoff (colab)

DI-engine Huggingface Kickoff (colab)

How to migrate a new RL Env | 如何迁移一个新的强化学习环境

How to customize the neural network model | 如何定制策略使用的神经网络模型

测试/部署 强化学习策略 的样例

新老 pipeline 的异同对比

(Click to Collapse)

 discrete means discrete action space, which is only label in normal DRL algorithms (1-23)

 means continuous action space, which is only label in normal DRL algorithms (1-23)

 means hybrid (discrete + continuous) action space (1-23)

 Distributed Reinforcement Learning｜分布式强化学习

 Multi-Agent Reinforcement Learning｜多智能体强化学习

 Exploration Mechanisms in Reinforcement Learning｜强化学习中的探索机制

 Imitation Learning｜模仿学习

 Offiline Reinforcement Learning｜离线强化学习

 Model-Based Reinforcement Learning｜基于模型的强化学习

 means other sub-direction algorithms, usually as plugin-in in the whole pipeline

P.S: The .py file in Runnable Demo can be found in dizoo

No.	Algorithm	Label	Doc and Implementation	Runnable Demo
1	DQN		DQN doc DQN中文文档 policy/dqn	python3 -u cartpole_dqn_main.py / ding -m serial -c cartpole_dqn_config.py -s 0
2	C51		C51 doc policy/c51	ding -m serial -c cartpole_c51_config.py -s 0
3	QRDQN		QRDQN doc policy/qrdqn	ding -m serial -c cartpole_qrdqn_config.py -s 0
4	IQN		IQN doc policy/iqn	ding -m serial -c cartpole_iqn_config.py -s 0
5	FQF		FQF doc policy/fqf	ding -m serial -c cartpole_fqf_config.py -s 0
6	Rainbow		Rainbow doc policy/rainbow	ding -m serial -c cartpole_rainbow_config.py -s 0
7	SQL		SQL doc policy/sql	ding -m serial -c cartpole_sql_config.py -s 0
8	R2D2		R2D2 doc policy/r2d2	ding -m serial -c cartpole_r2d2_config.py -s 0
9	PG		PG doc policy/pg	ding -m serial -c cartpole_pg_config.py -s 0
10	PromptPG		policy/prompt_pg	ding -m serial_onpolicy -c tabmwp_pg_config.py -s 0
11	A2C		A2C doc policy/a2c	ding -m serial -c cartpole_a2c_config.py -s 0
12	PPO/MAPPO		PPO doc policy/ppo	python3 -u cartpole_ppo_main.py / ding -m serial_onpolicy -c cartpole_ppo_config.py -s 0
13	PPG		PPG doc policy/ppg	python3 -u cartpole_ppg_main.py
14	ACER		ACER doc policy/acer	ding -m serial -c cartpole_acer_config.py -s 0
15	IMPALA		IMPALA doc policy/impala	ding -m serial -c cartpole_impala_config.py -s 0
16	DDPG/PADDPG		DDPG doc policy/ddpg	ding -m serial -c pendulum_ddpg_config.py -s 0
17	TD3		TD3 doc policy/td3	python3 -u pendulum_td3_main.py / ding -m serial -c pendulum_td3_config.py -s 0
18	D4PG		D4PG doc policy/d4pg	python3 -u pendulum_d4pg_config.py
19	SAC/[MASAC]		SAC doc policy/sac	ding -m serial -c pendulum_sac_config.py -s 0
20	PDQN		policy/pdqn	ding -m serial -c gym_hybrid_pdqn_config.py -s 0
21	MPDQN		policy/pdqn	ding -m serial -c gym_hybrid_mpdqn_config.py -s 0
22	HPPO		policy/ppo	ding -m serial_onpolicy -c gym_hybrid_hppo_config.py -s 0
23	BDQ		policy/bdq	python3 -u hopper_bdq_config.py
24	MDQN		policy/mdqn	python3 -u asterix_mdqn_config.py
25	QMIX		QMIX doc policy/qmix	ding -m serial -c smac_3s5z_qmix_config.py -s 0
26	COMA		COMA doc policy/coma	ding -m serial -c smac_3s5z_coma_config.py -s 0
27	QTran		policy/qtran	ding -m serial -c smac_3s5z_qtran_config.py -s 0
28	WQMIX		WQMIX doc policy/wqmix	ding -m serial -c smac_3s5z_wqmix_config.py -s 0
29	CollaQ		CollaQ doc policy/collaq	ding -m serial -c smac_3s5z_collaq_config.py -s 0
30	MADDPG		MADDPG doc policy/ddpg	ding -m serial -c ptz_simple_spread_maddpg_config.py -s 0
31	GAIL		GAIL doc reward_model/gail	ding -m serial_gail -c cartpole_dqn_gail_config.py -s 0
32	SQIL		SQIL doc entry/sqil	ding -m serial_sqil -c cartpole_sqil_config.py -s 0
33	DQFD		DQFD doc policy/dqfd	ding -m serial_dqfd -c cartpole_dqfd_config.py -s 0
34	R2D3		R2D3 doc R2D3中文文档 policy/r2d3	python3 -u pong_r2d3_r2d2expert_config.py
35	Guided Cost Learning		Guided Cost Learning中文文档 reward_model/guided_cost	python3 lunarlander_gcl_config.py
36	TREX		TREX doc reward_model/trex	python3 mujoco_trex_main.py
37	Implicit Behavorial Cloning (DFO+MCMC)		policy/ibc model/template/ebm	python3 d4rl_ibc_main.py -s 0 -c pen_human_ibc_mcmc_config.py
38	BCO		entry/bco	python3 -u cartpole_bco_config.py
39	HER		HER doc reward_model/her	python3 -u bitflip_her_dqn.py
40	RND		RND doc reward_model/rnd	python3 -u cartpole_rnd_onppo_config.py
41	ICM		ICM doc ICM中文文档 reward_model/icm	python3 -u cartpole_ppo_icm_config.py
42	CQL		CQL doc policy/cql	python3 -u d4rl_cql_main.py
43	TD3BC		TD3BC doc policy/td3_bc	python3 -u d4rl_td3_bc_main.py
44	Decision Transformer		policy/dt	python3 -u d4rl_dt_mujoco.py
45	EDAC		EDAC doc policy/edac	python3 -u d4rl_edac_main.py
46	QGPO		QGPO doc policy/qgpo	python3 -u ding/example/qgpo.py
47	MBSAC(SAC+MVE+SVG)		policy/mbpolicy/mbsac	python3 -u pendulum_mbsac_mbpo_config.py \ python3 -u pendulum_mbsac_ddppo_config.py
48	STEVESAC(SAC+STEVE+SVG)		policy/mbpolicy/mbsac	python3 -u pendulum_stevesac_mbpo_config.py
49	MBPO		MBPO doc world_model/mbpo	python3 -u pendulum_sac_mbpo_config.py
50	DDPPO		world_model/ddppo	python3 -u pendulum_mbsac_ddppo_config.py
51	DreamerV3		world_model/dreamerv3	python3 -u cartpole_balance_dreamer_config.py
52	PER		worker/replay_buffer	rainbow demo
53	GAE		rl_utils/gae	ppo demo
54	ST-DIM		torch_utils/loss/contrastive_loss	ding -m serial -c cartpole_dqn_stdim_config.py -s 0
55	PLR		PLR doc data/level_replay/level_sampler	python3 -u bigfish_plr_config.py -s 0
56	PCGrad		torch_utils/optimizer_helper/PCGrad	python3 -u multi_mnist_pcgrad_main.py -s 0
57	AWR		policy/ibc	python3 -u tabmwp_awr_config.py

(Click to Collapse)

No	Environment	Label	Visualization	Code and Doc Links
1	Atari			dizoo link env tutorial 环境指南
2	box2d/bipedalwalker			dizoo link env tutorial 环境指南
3	box2d/lunarlander			dizoo link env tutorial 环境指南
4	classic_control/cartpole			dizoo link env tutorial 环境指南
5	classic_control/pendulum			dizoo link env tutorial 环境指南
6	competitive_rl			dizoo link 环境指南
7	gfootball			dizoo link env tutorial 环境指南
8	minigrid			dizoo link env tutorial 环境指南
9	MuJoCo			dizoo link env tutorial 环境指南
10	PettingZoo			dizoo link env tutorial 环境指南
11	overcooked			dizoo link env tutorial
12	procgen			dizoo link env tutorial 环境指南
13	pybullet			dizoo link 环境指南
14	smac			dizoo link env tutorial 环境指南
15	d4rl			dizoo link 环境指南
16	league_demo			dizoo link
17	pomdp atari			dizoo link
18	bsuite			dizoo link env tutorial 环境指南
19	ImageNet			dizoo link 环境指南
20	slime_volleyball			dizoo link env tutorial 环境指南
21	gym_hybrid			dizoo link env tutorial 环境指南
22	GoBigger			dizoo link env tutorial 环境指南
23	gym_soccer			dizoo link 环境指南
24	multiagent_mujoco			dizoo link 环境指南
25	bitflip			dizoo link 环境指南
26	sokoban			dizoo link env tutorial 环境指南
27	gym_anytrading			dizoo link env tutorial
28	mario			dizoo link env tutorial 环境指南
29	dmc2gym			dizoo link env tutorial 环境指南
30	evogym			dizoo link env tutorial 环境指南
31	gym-pybullet-drones			dizoo link 环境指南
32	beergame			dizoo link 环境指南
33	classic_control/acrobot			dizoo link 环境指南
34	box2d/car_racing			dizoo link 环境指南
35	metadrive			dizoo link 环境指南
36	cliffwalking			dizoo link env tutorial 环境指南
37	tabmwp			dizoo link env tutorial 环境指南
38	frozen_lake			dizoo link env tutorial 环境指南
39	ising_model			dizoo link env tutorial 环境指南
40	taxi			dizoo link env tutorial 环境指南

means discrete action space

means continuous action space

means hybrid (discrete + continuous) action space

means multi-agent RL environment

means environment which is related to exploration and sparse reward

means offline RL environment

means Imitation Learning or Supervised Learning Dataset

means environment that allows agent VS agent battle

P.S. some enviroments in Atari, such as MontezumaRevenge, are also the sparse reward type.

DI-engine utilizes TreeTensor as the basic data container in various components, which is ease of use and consistent across different code modules such as environment definition, data processing and DRL optimization. Here are some concrete code examples:

• TreeTensor can easily extend all the operations of torch.Tensor to nested data:

  (Click for Details)

  import treetensor.torch as ttorch
  
  
  # create random tensor
  data = ttorch.randn({'a': (3, 2), 'b': {'c': (3, )}})
  # clone+detach tensor
  data_clone = data.clone().detach()
  # access tree structure like attribute
  a = data.a
  c = data.b.c
  # stack/cat/split
  stacked_data = ttorch.stack([data, data_clone], 0)
  cat_data = ttorch.cat([data, data_clone], 0)
  data, data_clone = ttorch.split(stacked_data, 1)
  # reshape
  data = data.unsqueeze(-1)
  data = data.squeeze(-1)
  flatten_data = data.view(-1)
  # indexing
  data_0 = data[0]
  data_1to2 = data[1:2]
  # execute math calculations
  data = data.sin()
  data.b.c.cos_().clamp_(-1, 1)
  data += data ** 2
  # backward
  data.requires_grad_(True)
  loss = data.arctan().mean()
  loss.backward()
  # print shape
  print(data.shape)
  # result
  # <Size 0x7fbd3346ddc0>
  # ├── 'a' --> torch.Size([1, 3, 2])
  # └── 'b' --> <Size 0x7fbd3346dd00>
  #     └── 'c' --> torch.Size([1, 3])

• TreeTensor can make it simple yet effective to implement classic deep reinforcement learning pipeline

  (Click for Details)

  import torch
  import treetensor.torch as ttorch
  
  B = 4
  
  
  def get_item():
      return {
          'obs': {
              'scalar': torch.randn(12),
              'image': torch.randn(3, 32, 32),
          },
          'action': torch.randint(0, 10, size=(1,)),
          'reward': torch.rand(1),
          'done': False,
      }
  
  
  data = [get_item() for _ in range(B)]
  
  
  # execute `stack` op
  - def stack(data, dim):
  -     elem = data[0]
  -     if isinstance(elem, torch.Tensor):
  -         return torch.stack(data, dim)
  -     elif isinstance(elem, dict):
  -         return {k: stack([item[k] for item in data], dim) for k in elem.keys()}
  -     elif isinstance(elem, bool):
  -         return torch.BoolTensor(data)
  -     else:
  -         raise TypeError("not support elem type: {}".format(type(elem)))
  - stacked_data = stack(data, dim=0)
  + data = [ttorch.tensor(d) for d in data]
  + stacked_data = ttorch.stack(data, dim=0)
  
  # validate
  - assert stacked_data['obs']['image'].shape == (B, 3, 32, 32)
  - assert stacked_data['action'].shape == (B, 1)
  - assert stacked_data['reward'].shape == (B, 1)
  - assert stacked_data['done'].shape == (B,)
  - assert stacked_data['done'].dtype == torch.bool
  + assert stacked_data.obs.image.shape == (B, 3, 32, 32)
  + assert stacked_data.action.shape == (B, 1)
  + assert stacked_data.reward.shape == (B, 1)
  + assert stacked_data.done.shape == (B,)
  + assert stacked_data.done.dtype == torch.bool

• File an issue on Github

• Open or participate in our forum

• Discuss on DI-engine discord server

• Discuss on DI-engine slack communication channel

• Discuss on DI-engine's WeChat group (i.e. add us on WeChat: ding314assist)

  

• Contact our email (opendilab@pjlab.org.cn)

• Contributes to our future plan Roadmap

We appreciate all the feedbacks and contributions to improve DI-engine, both algorithms and system designs. And CONTRIBUTING.md offers some necessary information.

@misc{ding,
    title={DI-engine: A Universal AI System/Engine for Decision Intelligence},
    author={Niu, Yazhe and Xu, Jingxin and Pu, Yuan and Nie, Yunpeng and Zhang, Jinouwen and Hu, Shuai and Zhao, Liangxuan and Zhang,  Ming and Liu, Yu},
    publisher={GitHub},
    howpublished={\url{https://github.com/opendilab/DI-engine}},
    year={2021},
}

DI-engine released under the Apache 2.0 license.