GraphMaker

[Paper]

Table of Contents

• Installation
• Usage
  • Train
  • Sample
• Frequently Asked Questions
  • Q1: libcusparse.so
  • Q2: Wandb
  • Q3: Other Requests
• Citation

Installation

conda create -n GraphMaker python=3.8 -y
conda activate GraphMaker
pip install torch==1.12.0+cu116 torchvision==0.13.0+cu116 torchaudio==0.12.0 --extra-index-url https://download.pytorch.org/whl/cu116
conda install -c conda-forge cudatoolkit=11.6
pip install dgl==1.1.0+cu116 -f https://data.dgl.ai/wheels/cu116/repo.html
pip install pandas scikit-learn pydantic wandb

You also need to compile orca.cpp (https://file.biolab.si/biolab/supp/orca/orca.html).

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

Usage

Train

# The GraphMaker-Sync variant simultaneously generates node attributes and graph structure.
python train_sync.py -d D
# The trained model checkpoint will be saved to {D}_cpts/Sync_XXX.pth

# The GraphMaker-Async variant first generates node attributes, and then generates graph structure.
python train_async.py -d D
# The trained model checkpoint will be saved to {D}_cpts/Async_XXX.pth

D can be one of the three built-in datasets, including cora, amazon_photo, amazon_computer.

Sample

python sample.py --model_path P

P is the path to a model checkpoint saved in the training stage.

Frequently Asked Questions

Q1: libcusparse.so

An error occurs that the program cannot find libcusparse.so.

To search for the location of it on linux,

find /path/to/directory -name libcusparse.so.11 -exec realpath {} \;

where /path/to/directory is the directory you want to search. Assume that the search returns home/miniconda3/envs/GraphMaker/lib/libcusparse.so.11. Then you need to manually specify the environment variable as follows.

export LD_LIBRARY_PATH=home/miniconda3/envs/GraphMaker/lib:$LD_LIBRARY_PATH

Q2: Wandb

What is WandB?

WandB is a tool for visualizing and tracking your machine learning experiments. It's free to use for open source projects. You may also use our code without it.

Q3: Other Requests

I have a question or request not listed here.

• It's generally recommended to open a GitHub issue. This allows us to track the progress, and the discussion might help others who have the same question.
• Otherwise, you can also send an email to [email protected].

Citation

@article{li2023graphmaker,
    title={GraphMaker: Can Diffusion Models Generate Large Attributed Graphs?},
    author={Mufei Li and Eleonora Kreačić and Vamsi K. Potluru and Pan Li},
    year={2023},
    journal={arXiv preprint arXiv:2310.13833}
}