AlphaZero inspired machine learning applied to the board game Pente.
Spiritual successor to my college thesis' source code.
- uv
curl -LsSf https://astral.sh/uv/install.sh | sh
uv sync
Run training starting from scratch:
uv run python main.py --model-dir=pente-model-v1.9 --batch-games=64 --batch-size=512 --arena --num-arena-games=35 --temp-threshold=5 --mcts-sim=40 --gpuRun training starting from a checkpoint:
uv run python main.py --model-dir=pente-model-v1.6 --model=pente-model-v1.5/checkpoint-30_19_361_5_128_1.pth.tar --batch-games=96 --batch-size=1024 --arena --num-arena-games=35 --temp-threshold=9 --mcts-sim=15 --gpuRun training starting from scratch and force processing of a raw dataset:
uv run python main.py --model-dir=pente-model-v1.5 --batch-games=1 --arena --raw-dataset=data/pente_dataset.txt --processed-dataset=data/pente-dataset-processed.pkl --force-dataset-processingStart inference and self-play evaluation
uv run python main.py --model-dir=PATH_TO_MODEL_DIR --model=PATH_TO_MODEL --infer --infer-mcts --batch-games=64 --batch-size=512 --arena --num-arena-games=35 --temp-threshold=5 --mcts-sim=40 --gpu