README.md

RecGNN

This repo contains the code to reproduce all results of the paper Learning Graph Algorithms with Recurrent Graph Neural Networks.

To cite our work:

@article{AlgorithmicRecurrentGNN2022,
  title = {Learning Graph Algorithms With Recurrent Graph Neural Networks},
  author = {Grötschla, Florian and Mathys, Joël and Wattenhofer, Roger},
  url = {https://arxiv.org/abs/2212.04934},
  publisher = {arXiv},
  year={2022}
}

Setup

The necessary conda environment can be set up as follows:

conda env create -f environment.yml
conda activate RecGNN

Training

To run all the configurations used in the paper, we first create configuration files for every run:

python create_configs.py

This will create three different directories with json configs for the RecGNN, IterGNN and GIN baseline runs. Once the configs are generated, they can be passed to the run_experiment.py script that will run them. Model checkpoints will be stored in models/ and stdout of the script can be redirected to a separate file for every run. This enables us to easily collect the results later on. First, create the models/ directory and directories for the output of every model:

mkdir models
mkdir runs_recGNN
mkdir runs_itergnn
mkdir runs_gin

Afterwards, all experiments for RecGNN can be run with:

CONFIG_ID=0
for FILE in configs_recGNN/* 
do 
	echo Run RecGNN config $CONFIG_ID
	python run_experiment.py --config $FILE > runs_recGNN/output_$CONFIG_ID.out 
	(( CONFIG_ID++ ))
done

For IterGNN, you can do:

CONFIG_ID=0
for FILE in configs_itergnn/* 
do 
	echo Run IterGNN config $CONFIG_ID
	python run_experiment.py --config $FILE > runs_itergnn/output_$CONFIG_ID.out 
	(( CONFIG_ID++ ))
done

And for GIN:

CONFIG_ID=0
for FILE in configs_gin/* 
do 
	echo Run GIN config $CONFIG_ID
	python run_experiment.py --config $FILE > runs_gin/output_$CONFIG_ID.out 
	(( CONFIG_ID++ ))
done

Now that the run directories are populated with the output from the script we can collect the results as follows:

python collect_results.py runs_recGNN output_recGNN # For RecGNN
python collect_results.py runs_itergnn output_itergnn # For IterGNN
python collect_results.py runs_gin output_gin # For GIN

The collect_results.py script scans all files in the provided directory and creates summary csv's. After running the code above, we get output_recGNN.csv, output_itergnn.csv and output_gin.csv with information for every run and training epoch.

Extrapolation runs

To run the extrapolation tests, do (this requires the output files and models from the trainings):

# RecGNN
python run_extrapolation.py output_recGNN.csv gin-mlp
python run_extrapolation.py output_recGNN.csv gru-mlp

# IterGNN
python run_extrapolation_itergnn.py

# GIN
python run_extrapolation_gin.py

These scripts will run the extrapolation on bigger graphs and write the results to output_extrapolation_gin-mlp.csv, output_extrapolation_gru-mlp.csv, output_extrapolation_itergnn.csv and output_extrapolation_gin.csv respectively.

Stabilization runs

For the stabilization runs, we have the following scripts:

python run_stabilization.py gin-mlp
python run_stabilization.py gru-mlp

Results will be written to output_stabilization_gin-mlp.csv and output_stabilization_gru-mlp.csv.

Creating the figures

After the above runs finished successfully, the figures presented in the paper can be generated by running the figures notebook.