2.0.13 - Model comparison, latent diffusion, and diffusion transformers

BayesFlow 2.0.13 Release Notes

BayesFlow 2.0.13 introduces substantial QoL improvements for Bayesian model comparison and high-dimensional inference, alongside several diagnostics. workflow and maintenance updates.

Major Changes

Bayesian Model Comparison

• Revamped ModelComparisonApproximator with a cleaner and more flexible design.
• Added a brand-new ModelComparisonWorkflow for end-to-end Bayesian model comparison. Check out the corresponding tutorial
• Introduced support for multiple discrete scoring rules generalizing the losses from Jeffrey & Wandelt (2023).
• Added new diagnostics for evaluating model-comparison performance.

Posterior Diagnostics

• Added TARP, a popular diagnostic for assessing posterior estimation quality.

Free-Form Networks & Diffusion Models

• Added Diffusion Transformers as flexible backbones for free-form networks. Can be tested by passing subnet=DiffusionTransformer() to any of our free-form inference backbones (DiffusionModel, FlowMatching, ConsistencyModel, StableConsistencyModel).
• Introduced latent diffusion for efficient compression and inference in high-dimensional settings. Can be accessed from the experimental module like so: LatentInferenceNetwork(autoencoder=..., inference_network=...). It is compatible with any inference network.
• Improved masking for free-form networks, enabling cleaner and more flexible conditioning and inference with arbitrary masks.

Minor Changes & Fixes

• Fixed absolute time indices in TimeSeriesTransformer when no explicit time axis is provided.
• Added an option to restore checkpoints directly from workflows.
• Improved checkpoint-related status and error messages.
• Exposed more keyword arguments for UNet backbones
• Updated the dependency list.
• Retargeted Dependabot updates to the dev branch.
• Included various minor bug fixes and maintenance improvements.