Overview

This repository contains code for the paper:

Mulamba, M., Mandi, J., Canoy, R., & Guns, T. (2020, September). Hybrid classification and reasoning for image-based constraint solving. In International Conference on Integration of Constraint Programming, Artificial Intelligence, and Operations Research (pp. 364-380). Springer, Cham.

@inproceedings{mulamba2020hybrid,
  title={Hybrid classification and reasoning for image-based constraint solving},
  author={Mulamba, Maxime and Mandi, Jayanta and Canoy, Rocsildes and Guns, Tias},
  booktitle={International Conference on Integration of Constraint Programming, Artificial Intelligence, and Operations Research},
  pages={364--380},
  year={2020},
  organization={Springer}
}

Neural Networks are implemented in PyTorch.

We use the CP-SAT solver from OR-Tools and its python API for constraint solvers. An implementation in CPMpy is also available.

Requirements

Make sure to have all required packages installed:

pandas
numpy
scipy
ortools
torch
torchvision
tqdm

Dataset

Download and extract the Sudoku dataset

wget -cq powei.tw/sudoku.zip && unzip -qq sudoku.zip 

or simply run the get_data.sh script.

Run

1. Run mnist.py with --save to train a (un)calibrated CNN on MNIST (set --help for more info)
2. Run viz_sudoku.py to solve visual sudokus (with --help for more info)