Tables and Trainers.
We use docker to bypass the ray/mac M4 issue: ray-project/ray#61495
To start the container for PPO self-play training. Adjust the settings in src/ppo_self_play/global_settings.py, then:
docker compose up --build
To end the container
docker compose down
If you want to pretrain the models using behavior cloning. Use the docker-compose file in the run/behavior_cloning/ folder
to generate the data and the run/behavior_cloning/train.py script to pre-train the PPO actors. Make sure to properly specify
the model path in run/main.py to ensure it is using the pre-trained model.
To display the live leaderboard, simply run in a separate shell
uv run leaderboard_gui.py
To view the tensorboard logs, run in a separate shell
uv run tensorboard --logdir=./results --reload_multifile=true --reload_interval=5
and open http://localhost:6006 on your browser
Open http://localhost:8265 on your browser
The get_from_docker fetches the run results and models.
The run\app\run.py script launches a small poker app UI to play poker with one of your pre-trained AIs.