This repo contains a simple demonstration for the IEEE CG&A21 paper entitled "DeepGD: A Deep Learning Framework for Graph Drawing Using GNN". It includes:
- a dataloader for Rome Graphs dataset,
- a basic implementation of DeepGD model,
- and a demo notebook that shows how to train DeepGD models from scratch with minimal amount of codes.
This code has been tested on python3.10 + cuda11.8 + pytorch2.0 + pyg2.3. Anaconda is suggested for managing dependencies, as installing pyg with pip can be tricky.
The default hyper-parameters of the model have been configured to reproduce the best performance reported in the DeepGD paper.
However, the layout initializer for the dataset is by default nx.drawing.random_layout, which is not PivotMDS that yields the best results shown in the paper (random initialization already produces good enough results though). Feel free to modify this behavior by passing a different initializer with dataset = RomeDataset(layout_initializer=something_else). A custom initializer can be any function that follows the behavior of nx.drawing.random_layout.
With Nvidia V100, each training epoch takes 30s on average. It takes around 600 epochs to completly converge.
For evaluation on custom data, the easiest way is to subclass RomeDataset and override raw_file_names and process_raw methods.
Caveat: Even though the behavior of
processdo not need to be overriden, it is required to have a dummydef process(self): super().process()defined in the subclasses to make it work properly. For details, please refer topyg.data.InMemoryDatasetdocumentation.
If you used our code or find our work useful in your research, please consider citing:
@article{deepgd,
author={Wang, Xiaoqi and Yen, Kevin and Hu, Yifan and Shen, Han-Wei},
journal={IEEE Computer Graphics and Applications},
title={DeepGD: A Deep Learning Framework for Graph Drawing Using GNN},
year={2021},
volume={41},
number={5},
pages={32-44},
doi={10.1109/MCG.2021.3093908}
}