cvignac/DiGress

What it solves

DiGress is designed to generate synthetic graphs and molecular structures. It addresses the challenge of creating realistic discrete graph data by applying denoising diffusion models to discrete spaces, allowing for the generation of complex networks and chemical molecules.

How it works

The project implements discrete denoising diffusion models. It can operate in both discrete and continuous modes to generate graphs. The system uses a Graph Transformer architecture to process and generate these structures, supporting multi-GPU training and automatic dataset downloading.

Who it’s for

It is intended for researchers and developers working in graph generation, molecular discovery, and generative AI for discrete structures.

Highlights

• Supports both discrete and continuous diffusion models for graph generation.
• Compatible with molecular datasets (like Guacamol and MOSES) and abstract graph datasets (like QM9 and SBM).
• Includes a Graph Transformer implementation.
• Supports multi-GPU acceleration for training and inference.