Chess-GPU-Hackathon-Strong-Compute

This is the codebase in which we developed a chess bot that combines convolutional neural networks with self-attention and squeeze-excitation blocks. Unlike traditional chess engines relying on reinforcement learning, this approach treats board evaluation as a computer vision problem, processing 8×8 chess boards with piece encoding to produce a scalar position score. Trained on 350,000 Grandmaster games using a 48-GPU cluster, this innovative model won us and my team a $10,000 credit in a hackathon.

Team: THETA HAT

Gopi Maguluri

Venkatachalam Subramanian Periya Subbu

Download the model from here

Here is a demo video

Model architecture

Project Structure

Root Directory

• train_chessVision.py: Main training script for the chess vision model with distributed training capabilities
• model.py: Core model implementation with CNN, self-attention, and squeeze-excitation blocks
• model_config.yaml: Configuration file for model hyperparameters
• chess_gameplay.py: Script for playing chess using the trained model
• requirements.txt: Dependencies needed to run the project
• chessVision.isc: Configuration file for the vision system

Directories

theta_hat_submission/

Contains the final submission version of the project:

• chess_gameplay.py: Game playing script
• model.py: Model implementation
• model_config.yaml: Simplified configuration
• pre_submission_val.py: Validation script for pre-submission
• checkpoint.pt: Saved model weights

utils/

Utility functions and preprocessing tools:

• train_utils.py: Utilities for training
• transformer_utils.py: Transformer-related utility functions
• datasets.py: Dataset loaders and processors
• optimizers.py: Custom optimizers (including Lamb optimizer)
• data_utils.py: Data handling utilities
• constants.py: Project constants
• stockfish: Stockfish chess engine binary for evaluation
• data_preprocessing/: Notebooks and scripts for data preparation
  • eval_preproc.ipynb: Preprocessing for evaluation data
  • gm_preproc.ipynb: Preprocessing for Grandmaster games
  • lc_preproc.ipynb: Preprocessing for LichessChess data
  • preproc.isc: Preprocessing configuration
  • preproc_boardeval.py: Board evaluation preprocessing

logs/

Training logs from distributed training:

• rank_0_1st_training.txt: Logs from first training run
• rank_0_2nd_training.txt: Logs from second training run

eval_notebook/

Contains evaluation notebooks:

• gameplay.ipynb: Interactive notebook to evaluate and play with the model

sample_demo/

Demo images:

• demo.png: Sample visualization
• demo_2.png: Additional visualization