MolecularDiffusion.modules.models.midi

MiDi backbone: mixed graph + 3D denoising diffusion.

Vignac, Osman, Toni, Frossard, “MiDi: Mixed Graph and 3D Denoising Diffusion for Molecule Generation”, ECML 2023, arXiv:2302.09048.

Ported from the reference implementation with the training loop, PyG data loading, wandb/torchmetrics accumulators, molecular metrics, ExtraFeatures and the variational-NLL machinery left behind – the platform already covers those. What remains is the network, the noise process and the sampling math, which is all modules/tasks/diffusion_midi.py needs.

Submodules