sudokuNet

This project uses PyTorch to create a ResNet-based Convolutional Neural Network with the goal of solving sudoku puzzles. SudokuSolver.py initializes and trains this model with easy modification of parameters through command line arguments. test_model.py allows for easy testing of any pretrained model on a random selection of puzzles, printing useful evaluation statistics and displaying model architecture with Tensorboard.

My best model was trained on an RTX 2080 Ti with 20 epochs, 2000000 example puzzles, and a batch size of 256.
Recent output for test_model.py with this model:

==================================================
SUMMARY STATISTICS
==================================================
Average accuracy for empty cells: 96.78%
Median accuracy for empty cells: 100.00%
Complete puzzles solved correctly: 72/100 (72.00%)
Minimum accuracy: 57.14%
Maximum accuracy: 100.00%

Loosely inspired by https://www.geeksforgeeks.org/sudoku-solver-using-tensorflow/
Training data from https://www.kaggle.com/datasets/rohanrao/sudoku/data
Due to issues with GitHub storage limits for Git LFS, pretrained models cannot be stored in this repository. I have uploaded some of my models to my personal OneDrive here: https://1drv.ms/u/c/e775e84bf4865dfa/EaFXLLj_c_pNkJkjrcsw8y0BH-msC_9qxptkYgHMg85MxA?e=J5iwLj