Package io.github.kirstenali.deepj.optimisers

Class AdamW

java.lang.Object

io.github.kirstenali.deepj.optimisers.AdamW

All Implemented Interfaces:

ParameterOptimizer

public final class AdamW
extends Object
implements ParameterOptimizer

AdamW optimizer with per-parameter state.

State is keyed by Parameter identity, and updates are applied in-place.

Constructor Summary

Constructors

Constructor	Description
AdamW(float lr, float beta1, float beta2, float eps, float weightDecay)

Method Summary

Modifier and Type	Method	Description
static AdamW	defaultAdamW(float lr)
float	lr()	Current learning rate.
void	setLr(float lr)	Update the learning rate (e.g. from a warmup/cosine schedule).
void	step(List<Parameter> params)	Update the provided parameters (one optimizer step).

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Details

AdamW

public AdamW(float lr,
 float beta1,
 float beta2,
 float eps,
 float weightDecay)

Method Details

defaultAdamW

public static AdamW defaultAdamW(float lr)

lr

public float lr()

Current learning rate.

setLr

public void setLr(float lr)

Update the learning rate (e.g. from a warmup/cosine schedule). Per-parameter moment state (m, v) is preserved.

step

public void step(List<Parameter> params)

Description copied from interface: ParameterOptimizer

Update the provided parameters (one optimizer step).

Specified by:

step in interface ParameterOptimizer