Registry: NumPy Optimizers

NumPy Optimizer Registry

Factory for NumPy-based optimizers, wired to the project’s Optimizer API.

Usage:

from qmlhc.optim.registry_numpy import create_optimizer_numpy opt = create_optimizer_numpy(“spsa”, lr0=0.05, eps0=0.1)

Returned objects implement:

• initialize(params) -> state

• step_params(model, params, context) -> (new_params, state)

qmlhc.optim.registry_numpy.create_optimizer_numpy(name, **kwargs)

Create a NumPy-based optimizer by name.

Parameters:

• name (str) –

  One of {“finite-diff”,”spsa”,”adam”,”natural-grad”,”trust-kl”,

  ”dual-ascent”,”mpc”,”kfac”}.

• kwargs (dict) – Optimizer hyperparameters. For wrappers (“trust-kl”,”dual-ascent”), pass ‘base_opt’ (the underlying optimizer instance).

Returns:

Optimizer instance exposing initialize(…) and step_params(…).

Return type:

object