DiGress: Discrete Denoising diffusion models for graph generation

Update (Nov 20th, 2023): Working with large graphs (more than 100-200 nodes)? Consider using SparseDiff, a sparse version of DiGress: https://github.com/qym7/SparseDiff

Update (July 11th, 2023): the code now supports multi-gpu. Please update all libraries according to the instructions. All datasets should now download automatically

For the conditional generation experiments, check the guidance branch.
If you are training new models from scratch, we recommand to use the fixed_bug branch in which some neural network layers have been fixed. The fixed_bug branch has not been evaluated, but should normally perform better. If you train the fixed_bug branch on datasets provided in this code, we would be happy to know the results.

Environment installation

This code was tested with PyTorch 2.0.1, cuda 11.8 and torch_geometrics 2.3.1

Download anaconda/miniconda if needed
Create a rdkit environment that directly contains rdkit:

conda create -c conda-forge -n digress rdkit=2023.03.2 python=3.9
conda activate digress
Check that this line does not return an error:

python3 -c 'from rdkit import Chem'
Install graph-tool (https://graph-tool.skewed.de/):

conda install -c conda-forge graph-tool=2.45
Check that this line does not return an error:

python3 -c 'import graph_tool as gt'
Install the nvcc drivers for your cuda version. For example:

conda install -c "nvidia/label/cuda-11.8.0" cuda
Install a corresponding version of pytorch, for example:

pip3 install torch==2.0.1 --index-url https://download.pytorch.org/whl/cu118
Install other packages using the requirement file:

pip install -r requirements.txt
Run:

pip install -e .
Navigate to the ./src/analysis/orca directory and compile orca.cpp:

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

Note: graph_tool and torch_geometric currently seem to conflict on MacOS, I have not solved this issue yet.

Run the code

All code is currently launched through python3 main.py. Check hydra documentation (https://hydra.cc/) for overriding default parameters.
To run the debugging code: python3 main.py +experiment=debug.yaml. We advise to try to run the debug mode first before launching full experiments.
To run a code on only a few batches: python3 main.py general.name=test.
To run the continuous model: python3 main.py model=continuous
To run the discrete model: python3 main.py
You can specify the dataset with python3 main.py dataset=guacamol. Look at configs/dataset for the list of datasets that are currently available

Checkpoints

The following checkpoints should work with the latest commit:

Planar: https://drive.switch.ch/index.php/s/hRWLp8gOGOGFzgR \ Performance of this checkpoint:
- Test NLL: 1135.6080
- {'spectre': 0.006211824145982536, 'clustering': 0.0563302653184386, 'orbit': 0.00980205113753696, 'planar_acc': 0.85, 'sampling/frac_unique': 1.0, 'sampling/frac_unique_non_iso': 1.0, 'sampling/frac_unic_non_iso_valid': 0.85, 'sampling/frac_non_iso': 1.0}
MOSES (the model in the paper was trained a bit longer than this one): https://drive.switch.ch/index.php/s/DBbvfMmezjg6KUm \ Performance of this checkpoint:
- Test NLL: 203.8171
- {'valid': 0.86032, 'unique@1000': 1.0, 'unique@10000': 0.9999, 'FCD/Test': 0.6176261401223826, 'SNN/Test': 0.5493580505032953, 'Frag/Test': 0.9986637035374839, 'Scaf/Test': 0.8997144919185305, 'FCD/TestSF': 1.2799741890619032, 'SNN/TestSF': 0.5231424506655995, 'Frag/TestSF': 0.9968362360368359, 'Scaf/TestSF': 0.11830576038721641, 'IntDiv': 0.8550915438149056, 'IntDiv2': 0.8489191659624407, 'Filters': 0.9707550678817184, 'logP': 0.02719348046624242, 'SA': 0.05725088257521343, 'QED': 0.0043940205061221965, 'weight': 0.7913020095007184, 'Novelty': 0.9442790697674419}
SBM: https://drive.switch.ch/index.php/s/rxWFVQX4Cu4Vq5j \ Performance of this checkpoint:
- Test NLL: 4757.903
- {'spectre': 0.0060240439382095445, 'clustering': 0.05020166160905111, 'orbit': 0.04615866844490847, 'sbm_acc': 0.675, 'sampling/frac_unique': 1.0, 'sampling/frac_unique_non_iso': 1.0, 'sampling/frac_unic_non_iso_valid': 0.625, 'sampling/frac_non_iso': 1.0}

The following checkpoints require to revert to commit 682e59019dd33073b1f0f4d3aaba7de6a308602e and run pip install -e .:

Guacamol: https://drive.switch.ch/index.php/s/jjM3pIHdxWrUGOH

Generated samples

We provide the generated samples for some of the models. If you have retrained a model from scratch for which the samples are not available yet, we would be very happy if you could send them to us!

Troubleshooting

PermissionError: [Errno 13] Permission denied: '/home/vignac/DiGress/src/analysis/orca/orca': You probably did not compile orca.

Use DiGress on a new dataset

To implement a new dataset, you will need to create a new file in the src/datasets folder. Depending on whether you are considering molecules or abstract graphs, you can base this file on moses_dataset.py or spectre_datasets.py, for example. This file should implement a Dataset class to process the data (check PyG documentation), as well as a DatasetInfos class that is used to define the noise model and some metrics.

For molecular datasets, you'll need to specify several things in the DatasetInfos:

The atom_encoder, which defines the one-hot encoding of the atom types in your dataset
The atom_decoder, which is simply the inverse mapping of the atom encoder
The atomic weight for each atom atype
The most common valency for each atom type

The node counts and the distribution of node types and edge types can be computed automatically using functions from AbstractDataModule.

Once the dataset file is written, the code in main.py can be adapted to handle the new dataset, and a new file can be added in configs/dataset.

Cite the paper

@inproceedings{
vignac2023digress,
title={DiGress: Discrete Denoising diffusion for graph generation},
author={Clement Vignac and Igor Krawczuk and Antoine Siraudin and Bohan Wang and Volkan Cevher and Pascal Frossard},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=UaAD-Nu86WX}
}

cvignac / DiGress

readme