nerva_jax.optimizers

Optimizers used to adjusts the model’s parameters based on the gradients.

Only SGD, Momentum and Nesterov variants are provided. The parser creates factory callables from textual specifications like “Momentum(mu=0.9)”.

Functions

parse_optimizer(text)

Parse a textual optimizer specification into a factory function.

Classes

CompositeOptimizer(optimizers)	Combines multiple optimizers to update different parameter groups.
GradientDescentOptimizer(obj, attr_x, attr_Dx)	Standard gradient descent optimizer: x -= eta * grad.
MomentumOptimizer(obj, attr_x, attr_Dx, mu)	Gradient descent with momentum for accelerated convergence.
NesterovOptimizer(obj, attr_x, attr_Dx, mu)	Nesterov accelerated gradient descent optimizer.
Optimizer()	Minimal optimizer interface used by layers to update parameters.

class nerva_jax.optimizers.Optimizer

Bases: object

Minimal optimizer interface used by layers to update parameters.

update(eta)

class nerva_jax.optimizers.CompositeOptimizer(optimizers: List[Optimizer])

Bases: Optimizer

Combines multiple optimizers to update different parameter groups.

update(eta)

Update all contained optimizers with the given learning rate.

class nerva_jax.optimizers.GradientDescentOptimizer(obj, attr_x: str, attr_Dx: str)

Bases: Optimizer

Standard gradient descent optimizer: x -= eta * grad.

update(eta)

Apply gradient descent update step.

class nerva_jax.optimizers.MomentumOptimizer(obj, attr_x: str, attr_Dx: str, mu: float)

Bases: GradientDescentOptimizer

Gradient descent with momentum for accelerated convergence.

update(eta)

Apply momentum update step.

class nerva_jax.optimizers.NesterovOptimizer(obj, attr_x: str, attr_Dx: str, mu: float)

Bases: MomentumOptimizer

Nesterov accelerated gradient descent optimizer.

update(eta)

Apply Nesterov accelerated gradient update step.

nerva_jax.optimizers.parse_optimizer(text: str) → Callable[[Any, str, str], Optimizer]

Parse a textual optimizer specification into a factory function.

Returns a callable that takes (x, Dx) and produces an Optimizer. Supported names: GradientDescent, Momentum(mu=…), Nesterov(mu=…).