sweet-rl 0.1

Why Sweet-RL

It exists dozens of Reinforcement Learning frameworks and algorithms implementations. Yet, most of them suffer of poor modularity and ease of understanding. This is why, I started to implement my own: Sweet-RL. It's so sweet that you can switch from Tensorflow 2.1 to PyTorch 1.4 with a single configuration line:

Getting started

Install sweet-rl

First, create a virtualenv:

python3.x -m venv ~/.virtualenvs/sweet/ 
# or: virtualenv ~/.virtualenvs/sweet/ -p python3
source ~/.virtualenvs/sweet/bin/activate

And then, install project dependancies:

make install # or pip install -e .

First execution

Run a DQN training:

python -m sweet.run --env=CartPole-v0 --algo=dqn --ml=tf

# Parameters:
#   -h, --help            show this help message and exit
#   --env ENV             Environment to play with
#   --algo ALGO           RL agent
#   --ml ML               ML platform (tf or torch)
#   --model MODEL         Model (dense, pi_actor_critic)
#   --timesteps TIMESTEPS
#                         Number of training steps
#   --output OUTPUT       Output directory (eg. './target/')

Custom neural network

If you want to specify your own model instead of default ones, take a look to sweet.agents.dqn.experiments.train_custom_model

Features, algorithms implemented

Algorithms

IO: Logs, model, tensorboard events

Outputs are configurable in training function:

targets: dict = {
        'output_dir': Path('./target/'), # Main directory to store your outputs
        'models_dir': 'models_checkpoints', # Saving models (depending on model_checkpoint_freq)
        'logs_dir': 'logs', # Saving logs (info, debug, errors)
        'tb_dir': 'tb_events' # Saving tensorboard events
}

Models are saved depending on model_checkpoint_freq parameter set in train function.

Benchmark

To reproduce benchmark, execute:

python -m sweet.benchmark.benchmark_runner

Here is an example of benchmark between TF 2.0 and Torch 1.4 with CartPole-v0 environment:

Troubleshootings

• Tensorflow 2.x doesn't work with Python 3.8 so far, so only Python versions 3.6 and 3.7 are supported
• GPU is not used. See https://www.tensorflow.org/install/gpu

History/Author

I started this open-source RL framework in january 2020, at first to take benefit of tensorflow 2.x readability without sacrifying the performance. Besides coding open-source project, i work for both Airbus and IRT Saint-Exupéry on Earth Observation satellites. Our team is focus on mission planning for satellites and Reinforcement Learning is an approach to solve it. Feel free to discuss with me: Adrien HADJ-SALAH @linkedin

You are welcome to participate to this project

RL related topics

• What is Reinforcement Learning It is supposed that you have knowledge in RL, if it is not the case, take a look to the spinningup from OpenAI