TSP Graph Model Training & Evaluation

Welcome to the TSP Graph Model repository! This project provides a full pipeline for training, evaluating, and visualizing the performance of various graph neural network (GNN) encoders and language models on the classic Traveling Salesman Problem (TSP). The codebase is designed for extensibility and reproducibility, making it easy to benchmark new models or encoders on TSP-style graph tasks.

---

Project Description

This repository explores the intersection of graph neural networks and large language models for solving combinatorial optimization problems, specifically the Traveling Salesman Problem (TSP). The pipeline includes:

• Dataset Generation: Automated creation of TSP instances of varying sizes and difficulties.
• Model Training: Training of different GNN encoders (e.g., SAGE, GAT, GIN) as graph preprocessors, paired with transformer-based language models.
• Evaluation: Automated inference and performance logging for all model/encoder pairs.
• Visualization: Generation of publication-quality plots and tables for accuracy, feasibility, and hallucination metrics, including detailed breakdowns by hallucination type.

The code is modular, GPU-accelerated, and designed for research and benchmarking.

---

Setup

1. Install Python dependencies

   pip install -r requirements.txt

2. Generate the TSP dataset

   python TSP.py

---

Training

Train all model/encoder combinations (this may take 4–6 hours on a high-end GPU):

mkdir -p saved_models
bash train.sh

---

Evaluation

Run inference and save model performance (about 30 minutes on a 4090):

mkdir -p outputs
bash eval.sh

---

Results & Visualization

After inference, you can:

1. Print results in the console

   python printer.py

2. Visualize results with graphs and tables

   python visualize.py

   • Plots and tables will be saved in the outputs/plots/ directory.
   • The script generates:
     • Accuracy and feasibility vs. node count (all model/encoder pairs)
     • Four hallucination-type plots vs. node count (all model/encoder pairs)
     • A summary table of results (results_table.csv)

---

Directory Structure

TSP.py              # Dataset generation
train.sh            # Trains all model/encoder pairs
train.py            # Training script (called by train.sh)
eval.sh             # Runs inference for all trained models
eval.py             # Evaluation script (called by eval.sh)
printer.py          # Prints summary results
visualize.py        # Generates plots and tables
outputs/            # Stores evaluation results and plots
saved_models/       # Stores trained model checkpoints
requirements.txt    # Python dependencies

---

Notes

• Training and evaluation are GPU-intensive. For best results, use a machine with a modern NVIDIA GPU.
• All scripts are designed to be run from the repository root.
• The codebase is modular—feel free to add new encoders, models, or tasks!

---

Contact

For questions, suggestions, or contributions, please open an issue or pull request. Feel free to contact the author at ethanreinhart@gmail.com directly if needed.