This is the official Pytorch implementation of Deep Generative Model for Periodic Graphs accepted by NeurIPS 2022:

@article{wang2022deep,
  title={Deep Generative Model for Periodic Graphs},
  author={Wang, Shiyu and Guo, Xiaojie and Zhao, Liang},
  journal={arXiv preprint arXiv:2201.11932},
  year={2022}
}

The code for PGD-VAE is adapted from the code of GRAN: Efficient graph generation with graph recurrent attention networks.

Running environment:

Python 3.9; PyTorch 1.8.1, networkx 2.5, scipy, numpy, pyyaml

Datasets

1. Data: QMOF, MeshSeg and synthetic datasets have been processed and provided in the repository as:

• QMOF: MOFGraphs.p, MOFUnitCells.p
• MeshSeg: MeshSeqGraphs.p, MeshSegUnitCells.p
• Synthetic: SynGraphs.p, SynUnitCells.p

2. To download the original datasets:

• Original QMOF data can be downloaded from here.
• Original MeshSeg data can be downloaded from here.
• Synthetic data is synthesized manually by the paper.

3. The code to extract and generate data from original QMOF and MeshSeg datasets have been provided in the package as:

• QMOF: MOFDataGen.py
• MeshSeg: MeshSegDataGen.py

Arguments of code：

1. data = {synthetic, mesh, qmof}
2. lr: learning rate
3. epoch: epochs
4. train = {True, False}
5. eval = {True, False}

Training the model

To train the model using synthetic data and 100 epochs with learning rate of 0.001:

python train.py --data synthetic --train True --epoch 100 --lr 0.001

Note：

1. The trained model will be saved as "{data}_train.pt" in the working directory.
2. The default dataset to be trained is the synthetic dataset. To train the model on another datasets please specify the dataset after --data argument.
3. To train the model with different penalties on KL divergence terms, please manually modify the code in model.py , line 293.

Evaluating the model

To evaluate trained model:

python train.py --data synthetic --train False --eval True

Note:

1. Functions for evaluation purpose are wrapped up in evaluation.py:

• compute_kld(): degree, cluster, orbit
• check_uniqueness(): uniqueness
• novelty_metric(): novelty