Adam - Klyn API

Adam

classin packageklyn.math.nn.optimizers

public class Adam<T extends AbstractFloat = Float> implements Optimizer<T>

Adam

All Implemented Interfaces: Optimizer

Adaptive moment estimation with per-parameter state and bias correction. Float16 parameters use FP32 master weights and moments; Float/Double retain their own precision. One instance retains its parameter identities across fit calls; use another optimizer to reset it. A fused buffer pass avoids per-element boxing and intermediate tensors.

Example

optimizer = Adam<Float>(learningRate=0.001)

Properties

Modifier and Type	Member	Description
public readonly property	beta1 `beta1 as Double`	First-moment decay.
public readonly property	beta2 `beta2 as Double`	Second-moment decay.
public readonly property	checkpointStates `checkpointStates as ImmutableList<AdamState<T>>`	Snapshot of parameter states for checkpoint transport, not tensor copies.
public readonly property	epsilon `epsilon as Double`	Stabilizer outside the square root.
public readonly property	learningRate `learningRate as Double`	Positive learning rate.

Constructors

Modifier and Type	Member	Description
public	Adam `Adam(learningRate as Double = 0.001, beta1 as Double = 0.9, beta2 as Double = 0.999, epsilon as Double = 1e-7)`	Configures Adam without weight decay or AMSGrad.

Methods

Modifier and Type	Member	Description
public override	apply `apply(parameter as NDArray<T>, gradient as NDArray<T>, gradientScale as Double = 1.0) as Void`	Updates moments and one parameter in place without altering the derivative.
public	restoreCheckpoint `restoreCheckpoint(states as IList<AdamState<T>>) as Void`	Restores checked per-parameter moments and bias-correction powers.

Inherited Methods

methodInherited Methods from Optimizer: apply