Dopamax is a library containing pure JAX implementations of common reinforcement learning algorithms. Everything is implemented in JAX, including the environments. This allows for extremely fast training and evaluation of agents, because the entire loop of environment simulation, agent interaction, and policy updates can be compiled as a single XLA program and executed on CPUs, GPUs, or TPUs. More specifically, the implementations in Dopamax follow the Anakin Podracer architecture -- see this paper for more details.
- Proximal Policy Optimization (PPO)
- Deep Q-Network (DQN)
- Deep Deterministic Policy Gradients (DDPG)
- Twin Delayed DDPG (TD3)
- Soft Actor Critic
- AlphaZero
Dopamax can be installed with:
pip install dopamaxThis will install the dopamax Python package, as well as a command-line interface (CLI) for training and evaluation.
Note that only the CPU version of JAX is installed by default. If you would like to use a GPU or TPU, you will need to
install the appropriate version of JAX. See the
JAX installation instructions.
Note
The above command will install the latest "release" of Dopamax, which may not necessarily align with the latest commit in the main branch. To install the version found in the main branch of this repository, you can use:
pip install git+https://github.com/rystrauss/dopamax.gitAfter installation, the Dopamax CLI can be used to train and evaluate agents:
dopamax --helpDopamax uses Weights and Biases (W&B) for logging and artifact management. Before using the CLI
for training and evaluation, you must first make sure you have a W&B account (it's free) and have authenticated
with wandb login.
Agent's can be trained using the dopamax train command, to which you must provide a configuration file. The
configuration file is a YAML file that specifies the agent, environment, and training hyperparameters. You can find
examples in the examples directory. For example, to train a PPO agent on the CartPole environment, you would
run:
dopamax train --config examples/ppo-cartpole/config.yamlNote that all of the example config files have a random seed specified, so you will get the same result every time you run the command. The seeds provided in the examples are known to result in a successful run (with the given hyperparameters). To get different results on each run, you can remove the seed from the config file.
Once you have trained some agents, you can evaluate them using the dopamax evaluate command. This will allow you to
specify a W&B agent artifact that you'd like to evaluate (these artifacts are produced by the training runs and
contain the agent hyperparameters and weights from the end of training). For example, to evaluate a PPO agent trained
on CartPole, you might use a command like:
dopamax evaluate --agent_artifact CartPole-PPO-agent:v0 --num_episodes 100where --num_episodes 100 signals that you would like to rollout the agent's policy for 100 episodes. The minimum,
mean, and maximum episode reward will be logged back to W&B. If you would additionally like to render the episodes and
have then logged back to W&B, you can provide the --render flag. But note that this will usually significantly slow
down the evaluation process since environment rendering is not a pure JAX function and requires callbacks to the host.
You should usually only use the --render flag with a small number of episodes.
Some of the JAX-native packages that Dopamax relies on: