RLtools Python Interface

A Python wrapper for RLtools (https://rl.tools). The only dependency is pybind11 to bind the C++ functions. Note that pybind11 requires the Python headers (which can e.g. be installed by apt install python3-dev on Ubuntu). The RLtools training code needs to be compiled at runtime because properties like the observation and action dimensions are not known at compile time. One of the fundamental principles of RLtools is that the sizes of all data structures and loops are known at compile-time so that the compiler can maximally reason about the code and heavily optimize it. Hence this wrapper takes an environment (Gymnasium interface) factory function as an input to infer the observation and action shapes and compile a bridge environment that is compatible with RLtools.

This wrapper is work in progress and for now just exposes the SAC training loop and does not allow much modification of hyperparameters etc. yet. Stay tuned.

Installation:

pip install rltools gymnasium

Note: On macOS you might have to install and accept the Xcode CLI license terms by sudo xcodebuild -license because we use clang in the background to compile RLtools operations just-in-time.

Example:

from rltools import SAC
import gymnasium as gym
from gymnasium.wrappers import RescaleAction

seed = 0xf00d
def env_factory():
    env = gym.make("Pendulum-v1")
    env = RescaleAction(env, -1, 1)
    env.reset(seed=seed)
    return env

sac = SAC(env_factory)
state = sac.State(seed)

finished = False
while not finished:
    finished = state.step()

Evaluating the Trained Policy

pip install gymnasium[classic-control]

env_replay = gym.make("Pendulum-v1", render_mode="human")

while True:
    observation, _ = env_replay.reset(seed=seed)
    finished = False
    while not finished:
        env_replay.render()
        action = state.action(observation)
        observation, reward, terminated, truncated, _ = env_replay.step(action)
        finished = terminated or truncated

Saving and Loading Checkpoints

# Save
with open("pendulum_sac_checkpoint.h", "w") as f:
    f.write(state.export_policy())
# Load
from rltools import load_checkpoint_from_path
policy = load_checkpoint_from_path("pendulum_sac_checkpoint.h")
action = policy.evaluate(observation) # Note that e.g. SAC's policies output mean and std (concatenated)

Custom C++ Enviroments

To get the maximum performance you should rewrite your environment in C++. Don't be scared it is actually quite straightforward and similar to creating a Gym environment. For an example of a custom pendulum environment see examples/custom_environment (just 105 lines of code).

Acceleration

On macOS the RLtools Python interface automatically uses Accelerate. To use MKL on linux you can install RLtools with the mkl option:

pip install rltools[mkl]

Windows

The RLtools Python interface also works on Windows but MKL is not integrated, yet. Please make sure to install Python from the installer from the Python website and not using the Windows Store Python version. The latter resides in a directory that requires admin privileges even for read access. Due to the just-in-time compilation of RLtools we need to be able to read the Python header and library files. After installing the right Python version the easies way to run RLtools is by opening the cloned folder in Visual Studio Code and launching the preconfigured targets. Make sure to start Visual Studio Code from the Visual Studio Prompt (e.g. Developer Command Prompt for VS 2022) by running code so that cl.exe (MSVC) is available in the environment.

Acknowledgements

Thanks a lot to @omtinez for letting us use the PyPI rltools handle! The PyPI rltools handle originally contained the v0.1.0, v0.2.0, and v0.2.1 releases of omtinez/rltools. Please refer to the source repository omtinez/rltools if you were looking for the original releases.