Combinatorial optimization with physics-inspired graph neural networks

This is my implementation of the PI-GNN paper, using torch_geometric.

Data

Data here is from this publicly available datset, the same as that being used in the numerical benchmark of the original paper. You can download it using the helper script:

python3 data-helper.py

The default directory hierarchy is like below:

.
├── data/
│   ├── dataset/
│   ├── links.txt
│   └── raw/
├── log/
├── model/
├── model.py
├── params.py
└── utils.py

where:

• model.py is the main file for the model.
• params.py contains the parameters for the model that one can tweak.
• utils.py contains the manipulation of data and a logger class.
• Processed data are stored in data/dataset/ in the form of nx.Graph objects, and the Hamiltonian matrix is only calculated when the data is loaded for the sake of memory efficiency (This could be improved if replaced with a sparse matrix along with sparse multiplication in place of corresponding standard matrix multiplication).

Environment

For some reason I cannot get torch and torch_geometric run on Python 3.10, so I am using Python 3.9 instead, with torch==1.12.0 and torch_geometric==2.0.4. Installation instructions can be found here for Pytorch and here for PyG, or you can use following command to install them:

# For Pytorch
pip3 install torch torchvision torchaudio
# This is for PyG, note that torch==1.12.0 is not officialy supported yet.
# Use at your own risk.
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.12.0+cpu.html