Unsupervised Discovery of Steerable Factors When Graph Deep Generative Models Are Entangled

TMLR 2024

Authors: Shengchao Liu, Chengpeng Wang, Jiarui Lu, Weili Nie, Hanchen Wang, Zuoxinran Li, Bolei Zhou, Jian Tang

This repository provides the source code for the paper GraphCG: Unsupervised Discovery of Steerable Factors When Graph Deep Generative Models Are Entangled, which aims at:

exploring the steerable factors in graphs
implementing the graph controllable generation in an unsupervised manner

[Project Page] [Paper] [ArXiv]
[NeurIPS Graph Learning Frontiers Workshop 2022]

1.1 Molecular Graph

1.1 Environment

conda create --name GraphCG python=3.7 pandas matplotlib
conda activate GraphCG

conda install -y -c pytorch pytorch=1.7.0 torchvision cudatoolkit=10.2
conda install -y -c rdkit rdkit=2019.03.4
conda install -y tabulate
conda install -y networkx
conda install -y scipy
conda install -y seaborn
conda install -y -c conda-forge opencv
pip install cairosvg
pip install orderedset
pip install pickle5
pip install git+https://github.com/bp-kelley/descriptastorus
pip install PyTDC
pip install scikit-learn==0.23
pip install gdown

pip install .

1.2 MoFlow

Go to directory, cd MoFlow.
Download datasets and pretrained models,

python step_01_download.py
unzip MoFlow.zip

Set up model weight path,

qm9_folder=./results_reported/qm9_64gnn_128-64lin_1-1mask_0d6noise_convlu1
zinc250k_folder=./results_reported/zinc250k_512t2cnn_256gnn_512-64lin_10flow_19fold_convlu2_38af-1-1mask
chembl_folder=./results_reported/chembl

Run testing scripts using bash test_GraphCG.sh.
Submit SLURM jobs using bash submit_*.sh.

1.3 HierVAE

Go to directory, cd HierVAE.
Download datasets and pretrained models,

python step_01_download.py
unzip HierVAE.zip

Set up model weight path,

data_name=qm9
model=results_reported/qm9/model.ckpt

Run testing scripts using bash test_GraphCG.sh. Notice that please make sure the GPU is enabled.
Submit SLURM jobs using bash submit_*.sh.

2 Point Clouds

2.1 Environment

conda create -n GraphCG python=3.6
conda activate GraphCG

conda install pytorch=1.9.1 torchvision -c pytorch -y
conda install numpy matplotlib pillow scipy tqdm scikit-learn -y
conda install tensorflow-gpu==1.13.1 -y
pip install tensorboardX==1.7
pip install pandas
pip install torchdiffeq==0.0.1
pip install cython
conda install -c sirokujira python-pcl --channel conda-forge
pip install gdown

pip install -e .

2.2 PointFlow

Go to directory, cd PointFlow.
Download datasets,

python step_01_download.py
unzip ShapeNetCore.v2.PC15k.zip
unzip pretrained_models.zip

Set up data path,

data_dir=ShapeNetCore.v2.PC15k

Run testing scripts using bash test_GraphCG.sh. Notice that please make sure the GPU is enabled.
Submit SLURM jobs using bash submit_*.sh.

3 Optimal Hyperparameters and Results

The optimal results and hyperparameters can be found at this HuggingFace link.

Please notice that in the archived scripts, we used hyperparameter contrastive_SSL (now changed to GraphCG_editing).

Cite Us

Feel free to cite this work if you find it useful to you!

@article{liu2024unsupervised,
    title={Unsupervised Discovery of Steerable Factors When Graph Deep Generative Models Are Entangled},
    author={Shengchao Liu and Chengpeng Wang and Jiarui Lu and Weili Nie and Hanchen Wang and Zhuoxinran Li and Bolei Zhou and Jian Tang},
    journal={Transactions on Machine Learning Research},
    issn={2835-8856},
    year={2024},
    url={https://openreview.net/forum?id=wyU3Q4gahM},
}

chao1224/GraphCG