Sample posterior weights¶
sample provides models, data loading, ensemble initialization, posterior
sampling and lossless trajectory storage for Bayesian neural networks. It uses
JAX and Flax, with declarative YAML or JSON configurations and explicit input
and output locations.
Start here¶
- Getting started: install the package and run a small regression example on one CPU device.
- Models and data: supported architectures, tasks, dataset sources, splits and preprocessing.
- Training and sampling: configuration, execution phases, sampler families, seeds and artifacts.
- Inference and evaluation: select stored draws, batch predictions and request predictive or chain diagnostics.
Package structure¶
| Module | Responsibility |
|---|---|
sample.config | Validated configurations for data, models, training and evaluation |
sample.architectures | Registered tabular, image and text models |
sample.dataset | Data loading, split selection and fitted preprocessing |
sample.bnns | Priors, likelihoods and the sampled model position |
sample.training | Ensemble initialization, sampling and evaluation stages |
sample.kernels | Sampler constructors and adaptation capabilities |
sample.artifacts | Declarative tree schemas and posterior trajectory storage |
sample.inference | Batched sample reading, prediction and evaluation |
sample.provenance | Acquisition identities and replay verification |
sample is independent of alignment and experiment orchestration. Applications
pass explicit locations and parameter trees; campaign management, fitting
selections and scientific publication belong to the consuming application.
The package is a fork of SMILE.
Attribution and citation entries are retained in packages/sample/README.md.