Probabilistic Graph Circuits (PGCs)

This repository contains the code for the following paper: Papež M, Rektoris M, Šmídl V, Pevný T. Probabilistic Graph Circuits: Deep Generative Models for Tractable Probabilistic Inference over Graphs.

An example of a PGC for undirected acyclic graphs. (a) We consider a graph $\mathbf{G}$ represented by a feature matrix, $\mathbf{X}$, and an adjacency tensor, $\mathbf{A}$, such that each instance of $\mathbf{G}$ (highlighted in green and blue) has a random number of nodes, $n\in(0,1,\ldots,m)$, where $m$ is a fixed maximum number of nodes. The empty places (white) are not included in the training data. (b) The main building block of PGCs is the $n$-conditioned joint distribution over $\mathbf{X}^n$ and $\mathbf{L}^n$, the latter of which is a flattened lower triangular part of $\mathbf{A}^n$. $\mathbf{X}^n$ and $\mathbf{L}^n$ are used as input into the node-PC and edge-PC, respectively. The empty places are marginalized out (grey). The outputs of these two PCs are passed through the product layer with $n_c$ units and the sum layer with a single unit.

@inproceedings{papez2025probabilistic,
title={Probabilistic Graph Circuits: Deep Generative Models for Tractable Probabilistic Inference over Graphs},
author={Milan Pape{\v{z}} and Martin Rektoris and Vaclav {\v{S}}m{\'i}dl and Tom{\'a}{\v{s}} Pevn{\'y}},
booktitle={The 41st Conference on Uncertainty in Artificial Intelligence},
year={2025},
url={https://openreview.net/forum?id=lqNtJRjlT1}
}

1. Install

Clone this repository.

git clone https://github.com/mlnpapez/PGC PGC

Go to the PGCs directory.

cd PGC

Set up the environment.

conda create --name pgc python=3.10

conda activate pgc

pip install torch==2.5.1 --index-url https://download.pytorch.org/whl/cu124
pip install rdkit==2024.3.6
pip install tqdm==4.67.0
pip install pandas==2.2.3
pip install pylatex==1.4.2
pip install scipy==1.14.1
pip install fcd_torch==1.0.7
pip install scikit-learn==1.6.0
pip install git+https://github.com/fabriziocosta/EDeN.git

2. Preprocess

The following command will download and preprocess the QM9 or Zinc250k dataset, each for five different orderings.

python -m utils.datasets

For example, the script will produce qm9_canonical.pt, which contains molecules in the canonical ordering of the atoms. To select the dataset, change dataset in utils.datasets.py.

3. Train

config/qm9/ contains JSON files with the hyper-parameters of different PGC variants. Change the hyper-parameters based on your preferences and then run the following command.

python -m train

It will train all the PGC variants (or only the selected ones if you change the list of names in train.py).

The resulting models will be stored in results/training/model_checkpoint/, and the corresponding illustrations of unconditional molecule generation, along with the metrics assessing the performance of the models, will be stored in results/training/model_evaluation/.

Unconditional samples of molecular graphs from the PT-S variant of PGCs (pgc_marg).

4. Gridsearch

gridsearch_hyperpars.py contains hyper-parameter grids for finding suitable architectures of the PGCs variants. Change the hyper-parameter grids based on your preferences, and then run the following command.

nohup python -m gridsearch > gridsearch.log &

This command will run the script in the background, submitting jobs to your SLURM cluster. The resulting models, metrics, and output logs will be stored in results/gridsearch/model_checkpoint/, results/gridsearch/model_evaluation/, and results/gridsearch/model_outputlogs/, respectively.

After completing all the SLURM jobs, run the following command.

python -m scripts.gridsearch_evaluate

It will produce a table comparing the PGC variants with the baselines (both in the .pdf and .tex formats).

4. Conditional Molecule Generation

Run the following command to generate new molecules conditionally on a known molecule.

python -m scripts.conditional_sampling

To impose a known structure of the generated molecules, change patt_smls in conditional_sampling.py. Similarly, to select a model from which to generate the samples, change model_path.

Conditional samples of molecular graphs from the PT-S variant of PGCs (pgc_marg). The known part of a molecule is highlighted in blue.