PGPS: A Neural Geometric Solver

Model Description

PGPS (Plane Geometry Problem Solver) is a neural geometric solver that uses multi-modal information through structural and semantic pre-training to solve plane geometry problems. This model was introduced in the IJCAI 2023 paper and represents the pre-trained language model component of the PGPSNet architecture.

Model Details

Model Type: Pre-trained Language Model for Geometric Problem Solving
Model File: LM_MODEL.pth
File Size: ~64MB
Framework: PyTorch
Paper: PGPS: A Neural Geometric Solver at IJCAI 2023
Original Repository: https://github.com/mingliangzhang2018/PGPS

Intended Use

This model is designed for:

Solving plane geometry problems
Parsing geometric diagrams
Understanding textual clauses in geometry problems
Generating solution programs for geometric problems

Requirements

Python 3.8
PyTorch 1.7.1
CUDA 10.2
One GTX-RTX or two TITAN Xp GPUs (for training)

Installation

Clone the original repository:

git clone https://github.com/mingliangzhang2018/PGPS.git
cd PGPS

Install dependencies:

pip install -r requirements.txt

Download this pre-trained model and place it in the appropriate directory.

Usage

Training with Pre-trained Model

python start.py --dataset Geometry3K --use_MLM_pretrain

Evaluation

python start.py --dataset Geometry3K --evaluate_only --eval_method completion

Dataset

The model works with the PGPS9K dataset, which contains:

Diagram annotations
Solution programs
Multi-modal geometric problem data

Download the dataset from the CASIA-PGPS9K homepage.

Citation

If you use this model in your research, please cite:

@inproceedings{zhang2023pgps,
  title={PGPS: A Neural Geometric Solver},
  author={Zhang, Mingliang and others},
  booktitle={Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI-23)},
  year={2023}
}

License

Apache 2.0

Authors

The original PGPS model was developed by Mingliang Zhang and colleagues. This Hugging Face repository is a mirror of the pre-trained model from the official GitHub repository.

Acknowledgments

Special thanks to the PGPS team for making their pre-trained models publicly available.