https://github.com/embarkstudios/emote
Reinforcement learning library from Embark Studios
https://github.com/embarkstudios/emote
embark reinforcement-learning
Last synced: about 1 year ago
JSON representation
Reinforcement learning library from Embark Studios
- Host: GitHub
- URL: https://github.com/embarkstudios/emote
- Owner: EmbarkStudios
- License: apache-2.0
- Created: 2022-02-02T20:31:15.000Z (over 4 years ago)
- Default Branch: main
- Last Pushed: 2024-09-25T18:34:14.000Z (almost 2 years ago)
- Last Synced: 2025-06-24T19:14:23.010Z (about 1 year ago)
- Topics: embark, reinforcement-learning
- Language: Python
- Homepage: https://embarkstudios.github.io/emote/
- Size: 3.28 MB
- Stars: 19
- Watchers: 4
- Forks: 0
- Open Issues: 22
-
Metadata Files:
- Readme: README.md
- Changelog: CHANGELOG.md
- Contributing: CONTRIBUTING.md
- License: LICENSE-APACHE
- Code of conduct: CODE_OF_CONDUCT.md
- Codeowners: .github/CODEOWNERS
Awesome Lists containing this project
README
# `🍒 emote`
**E**mbark's **Mo**dular **T**raining **E**ngine - a flexible framework for
reinforcement learning
[](https://embark.dev)
[](https://discord.gg/dAuKfZS)
[](https://buildkite.com/embark-studios/emote)
[](https://embarkstudios.github.io/emote/)
🚧 This project is a **work in progress**. Things can and will change. 🚧
## What it does
Emote provides a way to build reusable components for creating reinforcement learning algorithms, and a
library of premade componenents built in this way. It is strongly inspired by the callback setup used
by Keras and FastAI.
As an example, let us see how the SAC, the Soft Actor Critic algorithm by
[Haarnoja et al.](https://arxiv.org/abs/1801.01290) can be written using Emote. The main algorithm in
SAC is given in [Soft Actor-Critic Algorithms and Applications](https://arxiv.org/abs/1812.05905) and
looks like this:
Using the components provided with Emote, we can write this as
```python
device = torch.device("cpu")
env = DictGymWrapper(AsyncVectorEnv(10 * [HitTheMiddle]))
table = DictObsTable(spaces=env.dict_space, maxlen=1000, device=device)
memory_proxy = MemoryTableProxy(table)
dataloader = MemoryLoader(table, 100, 2, "batch_size")
q1 = QNet(2, 1)
q2 = QNet(2, 1)
policy = Policy(2, 1)
ln_alpha = torch.tensor(1.0, requires_grad=True)
agent_proxy = FeatureAgentProxy(policy, device)
callbacks = [
QLoss(name="q1", q=q1, opt=Adam(q1.parameters(), lr=8e-3)),
QLoss(name="q2", q=q2, opt=Adam(q2.parameters(), lr=8e-3)),
PolicyLoss(pi=policy, ln_alpha=ln_alpha, q=q1, opt=Adam(policy.parameters())),
AlphaLoss(pi=policy, ln_alpha=ln_alpha, opt=Adam([ln_alpha]), n_actions=1),
QTarget(pi=policy, ln_alpha=ln_alpha, q1=q1, q2=q2),
SimpleGymCollector(env, agent_proxy, memory_proxy, warmup_steps=500),
FinalLossTestCheck([logged_cbs[2]], [10.0], 2000),
]
trainer = Trainer(callbacks, dataloader)
trainer.train()
```
Here each callback in the `callbacks` list is its own reusable class that can readily be used
for other similar algorithms. The callback classes themselves are very straight forward to write.
As an example, here is the `PolicyLoss` callback.
```python
class PolicyLoss(LossCallback):
def __init__(
self,
*,
pi: nn.Module,
ln_alpha: torch.tensor,
q: nn.Module,
opt: optim.Optimizer,
max_grad_norm: float = 10.0,
name: str = "policy",
data_group: str = "default",
):
super().__init__(
name=name,
optimizer=opt,
network=pi,
max_grad_norm=max_grad_norm,
data_group=data_group,
)
self.policy = pi
self._ln_alpha = ln_alpha
self.q1 = q
self.q2 = q2
def loss(self, observation):
p_sample, logp_pi = self.policy(**observation)
q_pi_min = self.q1(p_sample, **observation)
# using reparameterization trick
alpha = torch.exp(self._ln_alpha).detach()
policy_loss = alpha * logp_pi - q_pi_min
policy_loss = torch.mean(policy_loss)
assert policy_loss.dim() == 0
return policy_loss
```
## Installation
For package management and environment handling we use `pants`. Install it from [pants](https://v1.pantsbuild.org/install.html). After `pants` is set up, verify that it is setup by running
```bash
pants tailor ::
```
### Common problems
**Box2d complains:** Box2d needs swig and python bindings. On apt-based systems try
```bash
sudo apt install swig
sudo apt install python3.10-dev
```
**Python 3.10 is tricky to install:** For Ubuntu based distros try adding the deadsnakes PPA.
## Contribution
[](../main/CODE_OF_CONDUCT.md)
We welcome community contributions to this project.
Please read our [Contributor Guide](CONTRIBUTING.md) for more information on how to get started.
Please also read our [Contributor Terms](CONTRIBUTING.md#contributor-terms) before you make any contributions.
Any contribution intentionally submitted for inclusion in an Embark Studios project, shall comply with the Rust standard licensing model (MIT OR Apache 2.0) and therefore be dual licensed as described below, without any additional terms or conditions:
### License
This contribution is dual licensed under EITHER OF
* Apache License, Version 2.0, ([LICENSE-APACHE](LICENSE-APACHE) or )
* MIT license ([LICENSE-MIT](LICENSE-MIT) or )
at your option.
For clarity, "your" refers to Embark or any other licensee/user of the contribution.