Class LabelMixingEventArgs<T>

Namespace

AiDotNet.Augmentation

Assembly

AiDotNet.dll

Event arguments for label mixing operations in augmentations like Mixup and CutMix.

public class LabelMixingEventArgs<T> : EventArgs

Type Parameters

T

The numeric type for calculations.

Inheritance

object

EventArgs

LabelMixingEventArgs<T>

Inherited Members

EventArgs.Empty

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

When augmentations like Mixup or CutMix are applied, they blend data from two samples together. The labels must also be blended proportionally. This event allows the training loop to handle the soft label generation appropriately.

For Beginners: In normal classification, an image is 100% one class (like "cat"). With Mixup/CutMix, we blend two images together, so the label becomes something like "70% cat, 30% dog". This helps the model learn smoother decision boundaries.

Constructors

LabelMixingEventArgs(Vector<T>, Vector<T>, T, int, int, MixingStrategy)

Creates a new label mixing event.

public LabelMixingEventArgs(Vector<T> originalLabels1, Vector<T> originalLabels2, T mixingLambda, int sampleIndex1, int sampleIndex2, MixingStrategy strategy)

Parameters

originalLabels1 Vector<T>

The labels for the first sample.

originalLabels2 Vector<T>

The labels for the second sample.

mixingLambda T

The mixing coefficient.

sampleIndex1 int

The index of the first sample.

sampleIndex2 int

The index of the second sample.

strategy MixingStrategy

The mixing strategy used.

Properties

Metadata

Gets additional metadata about the mixing operation.

public IDictionary<string, object> Metadata { get; }

Property Value

IDictionary<string, object>

Remarks

For CutMix, this may contain the bounding box coordinates of the cut region.

MixedLabels

Gets or sets the resulting mixed soft labels.

public Vector<T>? MixedLabels { get; set; }

Property Value

Vector<T>

Remarks

The training loop should set this to: lambda * OriginalLabels1 + (1 - lambda) * OriginalLabels2

MixingLambda

Gets the mixing coefficient (lambda).

public T MixingLambda { get; }

Property Value

T

Remarks

For Mixup: lambda determines the blend ratio. For CutMix: lambda represents the area ratio of the original image kept.

OriginalLabels1

Gets the original hard labels for the first sample.

public Vector<T> OriginalLabels1 { get; }

Property Value

Vector<T>

OriginalLabels2

Gets the original hard labels for the second (mixed) sample.

public Vector<T> OriginalLabels2 { get; }

Property Value

Vector<T>

SampleIndex1

Gets the index of the first sample in the batch.

public int SampleIndex1 { get; }

Property Value

int

SampleIndex2

Gets the index of the second sample in the batch.

public int SampleIndex2 { get; }

Property Value

int

Strategy

Gets the mixing strategy used.

public MixingStrategy Strategy { get; }

Property Value

MixingStrategy