Generative Graph Transformer

PyTorch implementation of Image-Conditioned Graph Generation for Road Network Extraction (https://arxiv.org/abs/1910.14388)

Overview

This library contains a PyTorch implementation of the Generative Graph Transformer (GGT): an autoregressive, attention-based model for image-to-graph generation as presented in [1](https://arxiv.org/abs/1910.14388), in addition to other baselines discussed in the paper. Find out more about this project in our blog post.

Dependencies

See requirements.txt

scipy==1.2.1
scikit_image==0.15.0
numpy==1.14.2
seaborn==0.9.0
networkx==1.11
torch==1.1.0
matplotlib==2.2.2
Pillow==6.1.0
skimage==0.0
tensorboardX==1.8
torchvision==0.4.0

Structure

data/: Should contain the Toulouse Road Network dataset. If you run download_dataset.sh the script will download the dataset introduced in our paper (Toulouse Road Network dataset).
models/: Contains the implementation of encoder and decoder models and baselines discussed in the paper, including but not only: GGT, GraphRNN extended to node features, simple RNN, simple MLP.
metrics/: Class for StreetMover distance in : streetmover_distance.py. Also contains different methods to compute statistics for the evaluation the models.
utils/: Contains hyper-parameter configuration for the different models, the dataset class and other utils.
main.py: Main script for training and testing of all the models.
arguments.py: Configuration specifying which model, experiment and hyper-parameter setting to be used in main.py.
pretrain_encoder.py: Script to pre-train the CNN encoder for image reconstruction as part of an auto-encoder.

Usage

First download Toulouse Road Network dataset using data/download_dataset.sh.
Then, configure arguments.py to choose which model to train/test and finally run main.py.
The output plots, logs, tensorboard and statistics will be automatically generated in output_graph/

Find out more about this project in our blog post. Please cite [1] in your work when using this library in your experiments.

Feedback

For questions and comments, feel free to contact Davide Belli.

Citation

[1] Belli, Davide and Kipf, Thomas (2019). Image-Conditioned Graph Generation for Road Network Extraction. NeurIPS 2019 workshop on Graph Representation Learning.

BibTeX format:

@article{belli2019image,
  title={Image-Conditioned Graph Generation for Road Network Extraction},
  author={Belli, Davide and Kipf, Thomas},
  journal={NeurIPS 2019 workshop on Graph Representation Learning},
  year={2019}
}

Copyright

This project is distributed under the MIT license. This was developed as part of a master thesis supervised by Thomas Kipf at the University of Amsterdam, and presented as a paper at the Graph Representation Learning workshop in NeurIPS 2019, Vancouver, Canada.