Owner import path¶
jaxstro.params
Purpose¶
Selective Equinox PyTree-to-vector parameter bridges and bijectors for bounded physical leaves.
Public records and callables¶
Parameterization, AbstractBijector, Identity, Exp, Softplus, and
Sigmoid.
Shape and dtype expectations¶
Free leaves are JAX arrays flattened into one floating vector. Static and fixed leaves remain in the model PyTree.
JAX transforms and AD classification¶
Vector conversion and bijectors compose with JIT and AD for fixed leaf selection. Selection metadata is static.
Failure behavior¶
Selection and vector-shape mismatches raise. from_vector replaces leaves and
does not rerun model initialization, so cached derived leaves may remain stale.
Contract and evidence links¶
See Parameters, constraints, and transforms and ADR 0009 - jaxstro.params selective inference.
Canonical import example¶
from jaxstro.params import Parameterization