Build Status

Cherry is a reinforcement learning framework for researchers built on top of PyTorch.

Unlike other reinforcement learning implementations, cherry doesn't implement a single monolithic interface to existing algorithms. Instead, it provides you with low-level, common tools to write your own algorithms. Drawing from the UNIX philosophy, each tool strives to be as independent from the rest of the framework as possible. So if you don't like a specific tool, you don’t need to use it.

Features

To learn more about the tools and philosophy behind cherry, check out our Getting Started tutorial.

Example

The following snippet showcases some of the tools offered by cherry.

import cherry as ch

# Wrap environments
env = gym.make('CartPole-v0')
env = ch.envs.Logger(env, interval=1000)
env = ch.envs.Torch(env)

policy = PolicyNet()
optimizer = optim.Adam(policy.parameters(), lr=1e-2)
replay = ch.ExperienceReplay()  # Manage transitions

for step in range(1000):
    state = env.reset()
    while True:
        mass = Categorical(policy(state))
        action = mass.sample()
        log_prob = mass.log_prob(action)
        next_state, reward, done, _ = env.step(action)

        # Build the ExperienceReplay
        replay.append(state, action, reward, next_state, done, log_prob=log_prob)
        if done:
            break
        else:
            state = next_state

    # Discounting and normalizing rewards
    rewards = ch.td.discount(0.99, replay.reward(), replay.done())
    rewards = ch.normalize(rewards)

    loss = -th.sum(replay.log_prob() * rewards)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    replay.empty()

Many more high-quality examples are available in the examples/ folder.

Installation

Note Cherry is considered in early alpha release. Stuff might break.

pip install cherry-rl

Documentation

Documentation and tutorials are available on cherry’s website: http://cherry-rl.net.

Contributing

First, thanks for your consideration in contributing to cherry. Here are a couple of guidelines we strive to follow.

We don't have forums, but are happy to discuss with you on slack. Make sure to send an email to smr.arnold@gmail.com to get an invite.

Acknowledgements

Cherry draws inspiration from many reinforcement learning implementations, including

Why 'cherry' ?

Because it's the sweetest part of the cake.