Trans-Dimensional Generative Modeling via Jump Diffusion Models

Authors: Andrew Campbell, William Harvey, Christian Weilbach, Valentin De Bortoli, Thomas Rainforth, Arnaud Doucet

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate our approach on molecular and video datasets of varying dimensionality, reporting better compatibility with test-time diffusion guidance imputation tasks and improved interpolation capabilities versus fixed dimensional models that generate state values and dimensions separately.
Researcher Affiliation Academia 1Department of Statistics, University of Oxford, UK 2 Department of Computer Science, University of British Columbia, Vancouver, Canada 3CNRS ENS Ulm, Paris, France
Pseudocode Yes Algorithm 1: Sampling the Generative Process
Open Source Code Yes Our code is available at https://github.com/andrew-cr/jump-diffusion
Open Datasets Yes We model the QM9 dataset [35, 36] of 100K varying size molecules.
Dataset Splits No We train on the 100K molecules contained in the QM9 training split.
Hardware Specification Yes Training a model requires approximately 7 days on a single GPU which was done on an Academic cluster.
Software Dependencies No The paper mentions software like 'Adam optimizer', 'RDKit', 'UNet', 'Transformer', and 'EGNN' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We train our model for 1.3 million iterations at a batch size of 64. We use the Adam optimizer with learning rate 0.00003. We also keep a running exponential moving average of the network weights that is used during sampling as is standard for training diffusion models [2, 3, 16] with a decay parameter of 0.9999.