thu-ml/tianshou

An elegant, flexible, and superfast PyTorch deep Reinforcement Learning platform.

repo name	thu-ml/tianshou
repo link	https://github.com/thu-ml/tianshou
homepage	https://tianshou.readthedocs.io
language	Python
size (curr.)	1626 kB
stars (curr.)	866
created	2018-04-16
license	MIT License

---

Tianshou (天授) is a reinforcement learning platform based on pure PyTorch. Unlike existing reinforcement learning libraries, which are mainly based on TensorFlow, have many nested classes, unfriendly API, or slow-speed, Tianshou provides a fast-speed framework and pythonic API for building the deep reinforcement learning agent with the least number of lines of code. The supported interface algorithms include:

• Policy Gradient (PG)
• Deep Q-Network (DQN)
• Double DQN (DDQN) with n-step returns
• Advantage Actor-Critic (A2C)
• Deep Deterministic Policy Gradient (DDPG)
• Proximal Policy Optimization (PPO)
• Twin Delayed DDPG (TD3)
• Soft Actor-Critic (SAC)

Tianshou supports parallel workers for all algorithms as well. All of these algorithms are reformatted as replay-buffer based algorithms. Our team is working on supporting more algorithms and more scenarios on Tianshou in this period of development, including model-based / atari / etc.

Installation

Tianshou is currently hosted on PyPI. You can simply install Tianshou with the following command:

pip3 install tianshou

You can also install with the newest version through GitHub:

pip3 install git+https://github.com/thu-ml/tianshou.git@master

After installation, open your python console and type

import tianshou as ts
print(ts.__version__)

If no error occurs, you have successfully installed Tianshou.

Documentation

The tutorials and API documentation are hosted on tianshou.readthedocs.io.

The example scripts are under test/ folder and examples/ folder.

Why Tianshou?

Fast-speed

Tianshou is a lightweight but high-speed reinforcement learning platform. For example, here is a test on a laptop (i7-8750H + GTX1060). It only uses 3 seconds for training an agent based on vanilla policy gradient on the CartPole-v0 task: (seed may be different across different platform and device)

python3 test/discrete/test_pg.py --seed 0 --render 0.03

We select some of famous reinforcement learning platforms: 2 GitHub repos with most stars in all RL platforms (OpenAI Baseline and RLlib) and 2 GitHub repos with most stars in PyTorch RL platforms (PyTorch DRL and rlpyt). Here is the benchmark result for other algorithms and platforms on toy scenarios: (tested on the same laptop as mentioned above)

RL Platform	Tianshou	Baselines	Ray/RLlib	PyTorch DRL	rlpyt
GitHub Stars
Algo - Task	PyTorch	TensorFlow	TF/PyTorch	PyTorch	PyTorch
PG - CartPole	9.03±4.18s	None	15.77±6.28s	None	?
DQN - CartPole	10.61±5.51s	1046.34±291.27s	40.16±12.79s	175.55±53.81s	?
A2C - CartPole	11.72±3.85s	*(~1612s)	46.15±6.64s	Runtime Error	?
PPO - CartPole	35.25±16.47s	*(~1179s)	62.21±13.31s (APPO)	29.16±15.46s	?
DDPG - Pendulum	46.95±24.31s	*(>1h)	377.99±13.79s	652.83±471.28s	172.18±62.48s
TD3 - Pendulum	48.39±7.22s	None	620.83±248.43s	619.33±324.97s	210.31±76.30s
SAC - Pendulum	38.92±2.09s	None	92.68±4.48s	808.21±405.70s	295.92±140.85s

*: Could not reach the target reward threshold in 1e6 steps in any of 10 runs. The total runtime is in the brackets.

?: We have tried but it is nontrivial for running non-Atari game on rlpyt. See here.

All of the platforms use 10 different seeds for testing. We erase those trials which failed for training. The reward threshold is 195.0 in CartPole and -250.0 in Pendulum over consecutive 100 episodes' mean returns.

Tianshou and RLlib’s configures are very similar. They both use multiple workers for sampling. Indeed, both RLlib and rlpyt are excellent reinforcement learning platform.

We will add results of Atari Pong / Mujoco these days.

Reproducible

Tianshou has its unit tests. Different from other platforms, the unit tests include the full agent training procedure for all of the implemented algorithms. It would be failed once if it could not train an agent to perform well enough on limited epochs on toy scenarios. The unit tests secure the reproducibility of our platform.

Check out the GitHub Actions page for more detail.

Modularized Policy

We decouple all of the algorithms into 4 parts:

• __init__: initialize the policy;
• process_fn: to preprocess data from replay buffer (since we have reformulated all algorithms to replay-buffer based algorithms);
• __call__: to compute actions over given observations;
• learn: to learn from a given batch data.

Within these API, we can interact with different policies conveniently.

Elegant and Flexible

Currently, the overall code of Tianshou platform is less than 1500 lines without environment wrappers for Atari and Mujoco. Most of the implemented algorithms are less than 100 lines of python code. It is quite easy to go through the framework and understand how it works. We provide many flexible API as you wish, for instance, if you want to use your policy to interact with the environment with (at least) n steps:

result = collector.collect(n_step=n)

If you have 3 environments in total and want to collect 1 episode in the first environment, 3 for the third environment:

result = collector.collect(n_episode=[1, 0, 3])

If you want to train the given policy with a sampled batch:

result = policy.learn(collector.sample(batch_size))

You can check out the documentation for further usage.

Quick Start

This is an example of Deep Q Network. You can also run the full script at test/discrete/test_dqn.py.

First, import some relevant packages:

import gym, torch, numpy as np, torch.nn as nn
from torch.utils.tensorboard import SummaryWriter
import tianshou as ts

Define some hyper-parameters:

task = 'CartPole-v0'
lr = 1e-3
gamma = 0.9
n_step = 3
eps_train, eps_test = 0.1, 0.05
epoch = 10
step_per_epoch = 1000
collect_per_step = 10
target_freq = 320
batch_size = 64
train_num, test_num = 8, 100
buffer_size = 20000
writer = SummaryWriter('log/dqn')  # tensorboard is also supported!

Make environments:

# you can also try with SubprocVectorEnv
train_envs = ts.env.VectorEnv([lambda: gym.make(task) for _ in range(train_num)])
test_envs = ts.env.VectorEnv([lambda: gym.make(task) for _ in range(test_num)])

Define the network:

class Net(nn.Module):
    def __init__(self, state_shape, action_shape):
        super().__init__()
        self.model = nn.Sequential(*[
            nn.Linear(np.prod(state_shape), 128), nn.ReLU(inplace=True),
            nn.Linear(128, 128), nn.ReLU(inplace=True),
            nn.Linear(128, 128), nn.ReLU(inplace=True),
            nn.Linear(128, np.prod(action_shape))
        ])
    def forward(self, s, state=None, info={}):
        if not isinstance(s, torch.Tensor):
            s = torch.tensor(s, dtype=torch.float)
        batch = s.shape[0]
        logits = self.model(s.view(batch, -1))
        return logits, state

env = gym.make(task)
state_shape = env.observation_space.shape or env.observation_space.n
action_shape = env.action_space.shape or env.action_space.n
net = Net(state_shape, action_shape)
optim = torch.optim.Adam(net.parameters(), lr=lr)

Setup policy and collectors:

policy = ts.policy.DQNPolicy(net, optim, gamma, n_step,
    use_target_network=True, target_update_freq=target_freq)
train_collector = ts.data.Collector(policy, train_envs, ts.data.ReplayBuffer(buffer_size))
test_collector = ts.data.Collector(policy, test_envs)

Let’s train it:

result = ts.trainer.offpolicy_trainer(
    policy, train_collector, test_collector, epoch, step_per_epoch, collect_per_step,
    test_num, batch_size, train_fn=lambda e: policy.set_eps(eps_train),
    test_fn=lambda e: policy.set_eps(eps_test),
    stop_fn=lambda x: x >= env.spec.reward_threshold, writer=writer, task=task)
print(f'Finished training! Use {result["duration"]}')

Save / load the trained policy (it’s exactly the same as PyTorch nn.module):

torch.save(policy.state_dict(), 'dqn.pth')
policy.load_state_dict(torch.load('dqn.pth'))

Watch the performance with 35 FPS:

collector = ts.data.Collector(policy, env)
collector.collect(n_episode=1, render=1 / 35)
collector.close()

Look at the result saved in tensorboard: (on bash script)

tensorboard --logdir log/dqn

You can check out the documentation for advanced usage.

Contributing

Tianshou is still under development. More algorithms and features are going to be added and we always welcome contributions to help make Tianshou better. If you would like to contribute, please check out CONTRIBUTING.md.

TODO

• More examples on [mujoco, atari] benchmark
• More algorithms
• Prioritized replay buffer
• RNN support
• Imitation Learning
• Multi-agent
• Distributed training

Citing Tianshou

If you find Tianshou useful, please cite it in your publications.

@misc{tianshou,
  author = {Jiayi Weng, Minghao Zhang, Dong Yan, Hang Su, Jun Zhu},
  title = {Tianshou},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/thu-ml/tianshou}},
}

Acknowledgment

Tianshou was previously a reinforcement learning platform based on TensorFlow. You can check out the branch priv for more detail. Many thanks to Haosheng Zou’s pioneering work for Tianshou before version 0.1.1.

We would like to thank TSAIL and Institute for Artificial Intelligence, Tsinghua University for providing such an excellent AI research platform.