Changelog#

Notable changes to sequince-density are documented in this file.

The format follows Keep a Changelog.

Unreleased #

Optional Zuko gradient clipping, cosine learning-rate annealing, early stopping and progress display, available through flow adapters’ fit_kwargs.
Zuko training-data checks before and after particle transforms, reporting affected dimensions for non-finite values.
Software archive DOI and method references for flow preconditioning and masked autoregressive flows.
Documentation sidebar links to the main Sequince documentation and sequince-examples under Related Packages.

Zuko’s validation_fraction=0 skips validation and keeps final parameters, with an empty validation-loss history. Combining it with patience raises ValueError.

First public beta release.

Normalizing-flow adapters for Zuko (PyTorch) and FlowJAX (JAX), supporting fitting, sampling, density evaluation, and forward/inverse transformations.
Gaussian-mixture proposals fitted with scikit-learn, using full covariance matrices and equal-weight training populations. Fitting and density evaluation use NumPy on the CPU.
Integration with Sequince’s FlowProposal and FlowTransform for fitted starting proposals and flow-preconditioned mutation.
Explicit RNG arguments for fitting and sampling, and checkpoint restoration of fitted models, particle transforms, and training histories. FlowJAX checkpoints also preserve the internal random key.