functional.svi.resolve_optimizer()

Normalize an optimizer specification into a NumPyro optimizer.

Usage

Source

functional.svi.resolve_optimizer(optim)

Accepted forms: None (Adam(0.01)); a finite positive scalar learning rate (float/int/NumPy scalar/0-d array) giving Adam(lr); an optax.GradientTransformation (wrapped via numpyro.optim.optax_to_numpyro, imported lazily so optax stays a soft dependency); a _NumPyroOptim (returned unchanged).

Parameters

optim: OptimizerLike

The optimizer specification (see ~numpyro_forecast.typing.OptimizerLike).

Returns

_NumPyroOptim

The resolved NumPyro optimizer.

Raises

OptimizerResolutionError

For boolean inputs of any form, including 0-d boolean arrays (bool is an int subclass, so a bool would silently mean Adam(1.0)), and for any other unrecognized type; the message lists the accepted forms.

ValueError

For a non-finite or non-positive learning rate.