Decorators

Decorators#

Model decorators for blayers.

  • autoreshape: Auto-reshape 1D arrays to (n, 1)

  • autoreparam: Auto-reparameterize LocScale distributions for VI and MCMC

Usage:

@autoreshape
@autoreparam
def my_model(x, y=None):
    ...
blayers.decorators.autoreshape(fn)[source]#

Decorator that ensures all array inputs have shape (n, d), not (n,).

blayers expects arrays with explicit trailing dimensions. This decorator auto-reshapes 1D arrays to (n, 1) so you don’t have to do it manually.

Usage:

@autoreshape def model(x, y=None):

mu = AdaptiveLayer()(‘mu’, x) return gaussian_link(mu, y)

Parameters:

fn (Callable[[...], Any])

Return type:

Callable[[…], Any]

blayers.decorators.autoreparam(model_fn=None, *, centered=0.0)[source]#

Auto-reparameterize LocScale distributions in a model for VI and MCMC.

Automatically applies LocScaleReparam to all LocScale distributions (Normal, LogNormal, StudentT, Cauchy, Laplace, Gumbel) found in the model, which can improve NUTS mixing and diagonal-guide VI by removing prior funnel geometries. Strongly informed coefficients may favor centering.

Works with or without parentheses:

@autoreparam
def model(x, y): ...

@autoreparam()
def model(x, y): ...

@autoreparam(centered=0.5)
def model(x, y): ...
Parameters:

  • model_fn (Callable[[...], Any] | None) – The model function (when used without parentheses).

  • centered (float) – Degree of centering for LocScaleReparam. 0.0 = fully non-centered (default, best for weak data); 1.0 = fully centered (better when data is informative).

Return type:

Any