Shallow Guess Trainer

A neural network trainer (using candle) for the ShallowGuess chess engine, featuring quantization-aware training with simulated int8 quantization.

Prerequisites

• Rust Stable - Install from rustup.rs
• CUDA Toolkit (optional) - For GPU acceleration

Training

Basic Usage

./target/release/train <hidden_layer_size> <data_dir> <validation_file> <model_export_path> <max_epochs> <sample_size> <total_steps>

Arguments

Argument	Description
hidden_layer_size	Number of neurons in the hidden layer (e.g., 512)
data_dir	Directory containing training data files
validation_file	Path to validation data file
model_export_path	Directory to save trained models
max_epochs	Maximum number of training epochs
sample_size	Number of files to sample per epoch
total_steps	Total training steps for scheduler

Options

Option	Default	Description
--batch-size <N>	32768	Batch size for training
--learning-rate <LR>	0.01	Initial learning rate
--warmup-steps <N>	16	Warmup steps for scheduler
--existing-model <PATH>	-	Path to existing model to continue training

Exporting Weights

Export trained weights to CSV format for embedding into the ShallowGuess engine. The fc1 (input-to-hidden layer) weights are exported in int8 quantized form.

cargo run --release --bin export <hidden_layer_size> <model_file> <export_file>