Coral

A PyTorch-based neural network library for board game evaluation.

Overview

Coral provides a flexible framework for building and deploying neural networks to evaluate board game positions. It includes modular components for input conversion, neural network architectures, and output interpretation.

Features

• Multiple NN Architectures: Multi-layer perceptrons, transformers, and custom models
• Flexible Input/Output Conversion: Modular converters for different board representations and evaluation formats
• Point-of-View Evaluation: Support for evaluating positions from different player perspectives
• PyTorch Integration: Built on PyTorch with JIT compilation support for optimized inference

Installation

pip install algorhino-coral==0.2.0

Release 0.2.0 uses the published valanga==0.3.0. Run tox --workdir /path/to/private/coral-tox for complete validation. requirements/validation-py314.txt pins the verified public Valanga wheel and stable CPU PyTorch artifact by URL and SHA-256 for both local tox and CI.

Requirements

• Python >= 3.14; validated with standard CPython 3.14.10, GIL enabled, PYTHON_JIT=0
• PyTorch >= 2.14.1 (2.14.1+cpu is validated; no release candidate)
• Valanga 0.3.0

Project Structure

src/coral/
├── board_evaluation.py          # Point-of-view and evaluation types
├── chi_nn.py                     # Base neural network class
└── neural_networks/
    ├── factory.py                # NN factory pattern
    ├── nn_state_evaluator.py   # Board evaluation with NNs
    ├── models/                   # NN architectures (MLP, Transformer)
    ├── input_converters/         # Board to tensor conversion
    └── output_converters/        # NN output to evaluation conversion

License

GPL-3.0-only