Presentation

This repository contains main code features of the paper Deep Neural Network for DrawiNg Networks, (DNN)², appeared in the Proceedings of the 29th International Symposium on Graph Drawing and Network Visualization (GD 2021) and available at https://arxiv.org/abs/2108.03632.

It contains research code that was not written to be extensively maintained.

The repository does not provides the model and benchmarks from the paper. However, all the methods (DNN layers, data processing, metrics) code are shared for reproducibility.

Requirements

This project was conducted using some ``dependencies" whose behavior might change according to their version. Here follows the list of the versions that were used (and hopefully work) for these dependencies.

• Python 3.8.5
• Pyspark 3.0.1 (Scalable data processing) ; can require specific installation (on Windows OS)
• TF/Keras 2.4.1 and TF/Keras 2.9.0 (Deep Learning) ; requires additionnal dependencies (CUDA, cudNN)
• tulip-python 5.5.1 (Graph management)

All Python dependencies are listed in requirements.txt.

Organisation

• data contains various graphs, including Rome Graphs taken from the Graph Drawing website in format .tlpb.gz
• code python code of (DNN)²
• train.py Example of main file to train a model using DNN2_Wrapper class.
• predict.ipynb Notebook providing examples to generate DNN2 predictions, visualize them and compute aesthetic metrics.

More precisely, code contains:

• DNN2_Wrapper.py Contains a class that paritions the different stages of data preprocessing and model training.
• graph_preprocessing.py All the operations we need to do on/with graphs.
• spark_preprocessing.py Scalable handling of data generation.
• sequences.py Data structure used to feed the model during training.
• losses.py Custom loss functions. Contains stress and tsNET losses.
• graph_layers.py Our implementation of some Graph Neural Network layers
• models.py Code to build the Neural network architecture.
• custom_callbacks.py VisualSampleCallback, a custom callback that, at the end of each epoch, saves some graph layout examples of that epoch model. Used to keep visual tracking of the model layouts evolution throughut the training
• SOTA_layouts.py Interfaces to layout graphs with state of the art layout algorithms and trained (DNN)².
• custom_graphs.py Provides interface to easily generate some common graphs
• metrics.py Metrics used to evaluate layouts.

Hyper parameters

This version of the code enables some design choices.

Loss function : stress and tsNET loss functions are implemented. To reproduce tsNET training as explained in (DNN)² paper, the model has to be trained twice (stage 1 and stage 2). The lambda parameters of each stage is commented in DNN2_Wrapper.py. For transfer learning, simply give the weights to be loaded in the corresponding parameter in train.py.

Topological structure : Chebyshev filters and GCN filters are implemented. For Chebyshev filters, a max degree K has to be set (> 0 to be meaningful). For GCN filters the parameter K should be set to 0.

Nodes features : unique node id, random metric and 2D layout (supported by tulip-python API) are supported and several can be given as input to the model. More information are directly commented in train.py.

Other hyper-parameters : most other common hyper-parameters (e.g., batch size, optimizer) can be provided through the DNN2_Wrapper class, more specifically in its methods __init__, prepare_data and train.

Contacts

Loann Giovannangeli: [email protected]
Frederic Lalanne: [email protected]
David Auber: [email protected]
Romain Giot: [email protected]
Romain Bourqui: [email protected]

Cite

@inproceedings{giovannangeli2021dnn2,
  title={{Deep Neural Network for DrawiNg Networks, (DNN)\textsuperscript{2}}},
  author={{Giovannangeli, Loann and Lalanne, Frederic and Auber, David and Giot, Romain and Bourqui, Romain}},
  booktitle={{International Symposium on Graph Drawing and Network Visualization}},
  pages={375--390},
  year={2021},
  organization={Springer}
}