Visual Sudoku Puzzle Classification

This repository contains the code necessary to create the collective neuro-symbolic benchmark suite described in the paper: Visual Sudoku Puzzle Classification: A Suite of Collective Neuro-Symbolic Tasks.

Pre-Generated Data

Pre-generated data is made available with 4x4 and 9x9 puzzles. 11 splits are provided for each configuration. The first 10 splits (01 - 10) are to be scored, while the final split (11) is free to be used as the experimenter sees fit. For convenience, the data has been split up by size, task, and datasets.

All the data is available in this directory: https://linqs-data.soe.ucsc.edu/public/datasets/ViSudo-PC/v01/

Shortcuts to the basic task datasets are provided here:

Data Source	4x4	9x9
MNIST	Basic 4x4 MNIST	Basic 9x9 MNIST
EMNIST	Basic 4x4 EMNIST	Basic 9x9 EMNIST
FMNIST	Basic 4x4 FMNIST	Basic 9x9 FMNIST
KMNIST	Basic 4x4 KMNIST	Basic 9x9 KMNIST

We also provide some sample datasets that only cover one specific setting and split:

Has Overlap?	4x4	9x9
Without Overlap	Sample 4x4 No Overlap	Sample 9x9 No Overlap
With Overlap	Sample 4x4 Overlap	Sample 9x9 Overlap

Generating Data

The ./scripts/generate-split.py script can be used to generate a split of puzzles. All available options can be viewed using --help:

./scripts/generate-split.py --help

Citations

To reference this work, please cite:

@inproceedings{augustine:nesy22,
    title = {Visual Sudoku Puzzle Classification: A Suite of Collective Neuro-Symbolic Tasks},
    author = {Eriq Augustine and Connor Pryor and Charles Dickens and Jay Pujara and William Yang Wang and Lise Getoor},
    booktitle = {International Workshop on Neural-Symbolic Learning and Reasoning (NeSy)},
    year = {2022},
    _publisher = {CEUR},
    address = {Windsor, United Kingdom},
}