CoRe-GD

Setup

Create the conda environment

CONDA_OVERRIDE_CUDA=11.7 conda env create -f environment.yml
conda activate CoRe-GD

# Install pyg
pip install torch==2.0.0
pip install torch_geometric pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.0.0+cu117.html

# Other dependencies
pip install test-tube wandb s_gd2 cupy-cuda11x ssgetpy

# Create directory to store models
mkdir models

# Install the neuraldrawer package that provides CoRe-GD
pip install -e .

Install the graph coarsening implementation used in the paper

git clone https://github.com/SergeiShumilin/graph-coarsening-up-to-date.git
cd graph-coarsening-up-to-date
pip install -e .
cd ..

If you want to use DeepGD, you need a valid R installation (including the python interface rpy2) with the igraph and graphlayouts packages. These are needed to compute PivotMDS initializations. You can do so as follows:

conda install r-essentials r-base
pip install rpy2
R   # This will let you enter the R shell
> install.packages("igraph")
> install.packages("graphlayouts")
> quit()

You also have to clone DeepGD-Demo into the root of the repo

git clone https://github.com/yolandalalala/DeepGD-Demo.git deepgd

Training

Training for CoRe-GD can be done by providing a config file. Configs for all dataset runs and ablation studies are provided in configs/:

python run_experiment.py --config configs/config_rome.json

Note: Logging is done through wandb

Assuming that you installed the necessary dependencies to compute PivotMDS initializations:

mkdir deepgd_models
python run_deepgd.py --dataset rome

Evaluation

During training, checkpoints are stored in the models/ directory. Pretrained checkpoints are available for all datasets in checkpoints/. In addition to the checkpoint, a config file is also needed. For the provided checkpoints you can use the default configs in configs/ for the datasets. Lastly, we specify the dataset to test on:

python eval_CoRe-GD.py checkpoints/core_rome.pt configs/config_rome.json rome 

For DeepGD, we provide checkpoints in deepgd_checkpoints/. They can be run on a dataset with:

python eval_dgd.py deepgd_checkpoints/rome.pt rome

To run the non-learned baselines on all datasets (5 runs with different random seeds each), do as follows:

python eval_baselines.py

This will store a csv containing all run information in baseline_runs.csv

The results should align with the following values:

Model	Rome	ZINC	MNIST	CIFAR10	PATTERN	CLUSTER
PivotMDS	388.77 ± 1.02	29.85 ± 0.00	173.32 ± 0.12	384.64 ± 0.20	3813.30 ± 1.89	3538.25 ± 0.73
neato	244.22 ± 0.55	5.76 ± 0.05	129.87 ± 0.19	263.44 ± 0.20	3188.34 ± 0.59	2920.30 ± 0.75
sfdp	296.05 ± 1.16	20.02 ± 0.27	172.63 ± 0.11	377.97 ± 0.11	3219.23 ± 1.31	2952.95 ± 1.81
(sgd)$^2$	233.49 ± 0.22	5.14 ± 0.01	129.19 ± 0.00	262.52 ± 0.00	3181.51 ± 0.05	2920.36 ± 0.78
DeepGD	235.22 ± 0.71	6.19 ± 0.07	129.23 ± 0.03	262.91 ± 0.13	3080.70 ± 0.24	2838.13 ± 0.08
CoRe-GD	233.17 ± 0.13	5.11 ± 0.02	129.10 ± 0.02	262.68 ± 0.08	3067.02 ± 0.79	2827.81 ± 0.36
CoRe-GD-mix	234.60 ± 0.10	5.21 ± 0.02	129.24 ± 0.01	262.90 ± 0.02	3066.31 ± 0.20	2828.13 ± 0.12

Scaling runs

The runtimes for the scaling experiment from the paper can be reproduced with

python runtime_comparison.py