Parax v0.7 brings parametric modeling to JAX with constrained parameters and Bayesian sampling
Bridge pure JAX PyTrees and object-oriented modeling with derived parameters and metadata.
Deep Dive
Parax v0.7 is out—a JAX library for parametric modeling that bridges pure PyTree manipulation with object-oriented approaches like Equinox. New features include derived/constrained parameters with metadata, computed PyTrees and callable parameterizations, plus abstract interfaces for fixed, bounded, and probabilistic PyTrees and parameters. Two fresh examples in the docs showcase bounded optimization (JAXopt) and Bayesian sampling (BlackJAX).
Key Points
- Derived/constrained parameters with automatic metadata updates simplify relationship modeling
- New computed PyTrees and callable parameterizations enable complex transformations within JAX
- Abstract interfaces support fixed, bounded, and probabilistic parameters, demonstrated with JAXopt and BlackJAX examples
Why It Matters
Bridges functional and OOP modeling in JAX, enabling cleaner code for scientific computing and Bayesian inference.