Efficient and Scalable Graph Generation through Iterative Local Expansion

This repository contains the reference implementation of the paper Efficient and Scalable Graph Generation through Iterative Local Expansion.

Setup

To get started, follow these steps:

• Clone this repository

• Create the specified conda environment named graph-generation by running the following command on Linux or macOS:

  conda env create -f environment.yaml

  and on Windows:

  conda env create -f environment_win.yaml

  Note that the graph-tool library is not available on Windows. This library is used for SBM graph evaluation, which will consequently not work on Windows. Everything else will work regardless of the operating system.

• Activate the newly created environment:

  conda activate graph-generation

• Navigate to util/orca and compile the Orca library:

  g++ -O2 -std=c++11 -o orca orca.cpp

Usage

The main entry point is main.py with parameters managed by the Hydra framework. To reproduce the results from the paper, run:

python main.py +experiment=XXX

where XXX is one of the following experiments: planar, sbm, tree, protein, point_cloud, planar_extrapolation, tree_extrapolation, planar_interpolation, tree_interpolation.

To reproduce the results for our one-shot baseline, add the prefix oneshot_ to the experiment name when running the same experiments, for example, oneshot_planar.

New experiments can be added by creating a new config file in config/experiment/ or by passing the parameters directly through the command line. Please refer to the Hydra documentation for more information.

Checkpoints

When training.save_checkpoint in the configuration is set to True, checkpoints are saved. To resume training from a checkpoint, set training.resume to the step number of the checkpoint, or to True to resume from the latest checkpoint.

Wandb

To log the results to Wandb, set wandb.logging to True in the configuration.

Multirun

The config/experiment/multirun_sample.yaml file contains a sample configuration for launching multiple runs with hyperparameter sweeps.

Slurm

The slurm/cluster/my_cluster.yaml file contains a sample configuration for launching experiments using the Slurm workload manager. To launch experiment XXX on the cluster, adapt the my_cluster.yaml for your specific cluster and run:

python main.py +experiment=XXX +slurm=my_cluster

Citation

When using this code, please cite our paper:

@misc{bergmeister2023efficient,
      title={Efficient and Scalable Graph Generation through Iterative Local Expansion}, 
      author={Andreas Bergmeister and Karolis Martinkus and Nathanaël Perraudin and Roger Wattenhofer},
      year={2023},
      eprint={2312.11529},
      archivePrefix={arXiv},
      primaryClass={cs.SI}
}