Bandwidth Restricted Graph Generation

Repo for Improving Graph Generation by Restricting Graph Bandwidth. Implemented in python 3.10.

setup

Here's how to build the conda environment:

conda env create --file env.yml
conda activate graph_gen
pip install -e .

Here's how to build orca to do orbit calculations for MMD evaluations:

cd graph_gen/analysis
g++ -std=c++11 -o orca.exe orca.cpp

Here's how to prepare the molecular datasets:

cd datasets
unzip zinc.tab.zip
unzip peptide_multi_class_dataset.csv.zip

layout

• graph_gen/data has synthetic and molecular datasets
• graph_gen/models contains models and model training + hyperoptimization scripts
• graph_gen/analysis has MMD calculations

hyperparameters

The hyperparameters we found using hyperoptimization are in hyperparameters.

hyperoptimizing models

Hyperoptimization scripts are in graph_gen/models/hyperoptimize. The scripts have help for all of their arguments. GraphRNN example:

conda activate graph_gen
python graph_gen/graph_gen/hyperoptimize_graphRNN.py --epochs 10 \ 
    --order C-M --data_name PROTEINS --version TEST --count 3

VAE example:

conda activate graph_gen
python graph_gen/graph_gen/hyperoptimize_gine_vae.py --order C-M --data_name PROTEINS \
    --edge_augmentation none --hidden_dim 32 --epochs 10 --version TEST --count 2 \ 
    --empirical_bw

Diffusion example:

conda activate graph_gen
python graph_gen/graph_gen/hyperoptimize_gine_diffusion.py --order BFS --data_name zinc250k \ 
    --hidden_dim 128 --version TEST-v4 --count 2 --epochs 5

train-evaluate

Train-evaluate scripts which use the hyperparameters found in hyperopt are in graph_gen/models/train_evaluate/. The scripts have help for all of their arguments. GraphRNN example:

conda activate graph_gen
python graph_gen/graph_gen/train_evaluate_graphRNN.py \ 
    --lr 0.0011 --wd 0.007 --order BFS --data_name ENZYMES --temperature 0.4 \ 
    --epochs 10 --version GraphRNNevalTEST --replicate 0

VAE example:

conda activate graph_gen

python graph_gen/graph_gen/train_evaluate_gine_vae.py \ 
    --kl_weight 0.0003 --lr 0.005 --order BFS --data_name ENZYMES --version \ 
    gine_vae_eval_TEST --sigma 1 --epochs 10 --replicate 0 --hidden_dim 32

Diffusion example:

conda activate graph_gen
python graph_gen/graph_gen/train_evaluate_gine_diffusion.py --data_name DD \
    --order C-M --lr 0.004 --hidden_dim 64 --empirical_bw --version diffusion_eval_TEST \ 
     --epochs 5 --replicate 0