Class LambGpuConfig

Namespace

AiDotNet.Interfaces

Assembly

AiDotNet.dll

Configuration for LAMB (Layer-wise Adaptive Moments) optimizer on GPU.

public class LambGpuConfig : IGpuOptimizerConfig

Inheritance

object

LambGpuConfig

Implements

IGpuOptimizerConfig

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

LAMB combines Adam's adaptive learning with LARS's layer-wise scaling. It was designed for training BERT and other large transformers with huge batches.

For Beginners: LAMB is like a combination of Adam and LARS. It uses Adam's moment estimates for adaptive learning AND applies layer-wise scaling like LARS. This enables training very large models (like BERT) with very large batch sizes efficiently.

Constructors

LambGpuConfig(float, float, float, float, float, int)

Creates a new LAMB GPU configuration.

public LambGpuConfig(float learningRate, float beta1 = 0.9, float beta2 = 0.999, float epsilon = 1E-06, float weightDecay = 0.01, int step = 0)

Parameters

learningRate float

Learning rate for parameter updates.

beta1 float

Exponential decay rate for first moment (default 0.9).

beta2 float

Exponential decay rate for second moment (default 0.999).

epsilon float

Numerical stability constant (default 1e-6).

weightDecay float

Weight decay coefficient (default 0.01).

step int

Current optimization step.

Properties

Beta1

Gets the exponential decay rate for the first moment estimates (typically 0.9).

public float Beta1 { get; init; }

Property Value

float

Beta2

Gets the exponential decay rate for the second moment estimates (typically 0.999).

public float Beta2 { get; init; }

Property Value

float

Epsilon

Gets the small constant for numerical stability (typically 1e-6).

public float Epsilon { get; init; }

Property Value

float

LearningRate

Gets the learning rate for parameter updates.

public float LearningRate { get; init; }

Property Value

float

OptimizerType

Gets the type of optimizer (SGD, Adam, AdamW, etc.).

public GpuOptimizerType OptimizerType { get; }

Property Value

GpuOptimizerType

Step

Gets the current optimization step (used for bias correction in Adam-family optimizers).

public int Step { get; init; }

Property Value

int

WeightDecay

Gets the weight decay (L2 regularization) coefficient.

public float WeightDecay { get; init; }

Property Value

float

Methods

ApplyUpdate(IDirectGpuBackend, IGpuBuffer, IGpuBuffer, GpuOptimizerState, int)

Applies the optimizer update to the given parameter buffer using its gradient.

public void ApplyUpdate(IDirectGpuBackend backend, IGpuBuffer param, IGpuBuffer gradient, GpuOptimizerState state, int size)

Parameters

backend IDirectGpuBackend

The GPU backend to execute the update.

param IGpuBuffer

Buffer containing the parameters to update (modified in-place).

gradient IGpuBuffer

Buffer containing the gradients.

state GpuOptimizerState

Optimizer state buffers (momentum, squared gradients, etc.).

size int

Number of parameters to update.

Remarks

For Beginners: This method applies the optimizer's update rule directly on the GPU. Each optimizer type (SGD, Adam, etc.) implements its own update logic using GPU kernels. The state parameter contains any auxiliary buffers needed (like velocity for SGD with momentum, or m/v buffers for Adam).

Design Note: Following the Open/Closed Principle, each optimizer config knows how to apply its own update, so adding new optimizers doesn't require modifying layer code.