Class DETRSetLoss<T>

Namespace

AiDotNet.ComputerVision.Detection.Losses

Assembly

AiDotNet.dll

DETR Set Prediction Loss with Hungarian Matching for end-to-end object detection.

public class DETRSetLoss<T> : LossFunctionBase<T>, ILossFunction<T>

Type Parameters

T

The numeric type used for calculations.

Inheritance

object

LossFunctionBase<T>

DETRSetLoss<T>

Implements

ILossFunction<T>

Inherited Members

LossFunctionBase<T>.NumOps

LossFunctionBase<T>.Engine

LossFunctionBase<T>.CalculateLossAndGradientGpu(IGpuTensor<T>, IGpuTensor<T>)

LossFunctionBase<T>.ValidateVectorLengths(Vector<T>, Vector<T>)

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Extension Methods

LossFunctionExtensions.ComputeGradient<T>(ILossFunction<T>, Tensor<T>, Tensor<T>)

LossFunctionExtensions.ComputeLoss<T>(ILossFunction<T>, Tensor<T>, Tensor<T>)

Remarks

For Beginners: Unlike traditional detectors that use anchors and NMS, DETR treats detection as a set prediction problem. It uses Hungarian matching to find the optimal assignment between predicted and ground truth boxes, then computes loss on the matched pairs.

The loss has three components: - Classification loss: Cross-entropy for class predictions - Box loss: L1 loss for box coordinates - GIoU loss: For better box regression

Reference: Carion et al., "End-to-End Object Detection with Transformers", ECCV 2020

Constructors

DETRSetLoss(int, double, double, double)

Creates a new DETR set loss instance.

public DETRSetLoss(int numClasses = 91, double classWeight = 1, double boxL1Weight = 5, double boxGIoUWeight = 2)

Parameters

numClasses int

Number of object classes (including no-object class).

classWeight double

Weight for classification loss.

boxL1Weight double

Weight for L1 box loss.

boxGIoUWeight double

Weight for GIoU box loss.

Methods

CalculateDerivative(Tensor<T>, Tensor<T>)

Calculates the gradient of the DETR loss.

public Tensor<T> CalculateDerivative(Tensor<T> predicted, Tensor<T> targets)

Parameters

predicted Tensor<T>

targets Tensor<T>

Returns

Tensor<T>

CalculateDerivative(Vector<T>, Vector<T>)

Calculates the gradient of DETR loss with respect to predicted values.

public override Vector<T> CalculateDerivative(Vector<T> predicted, Vector<T> actual)

Parameters

predicted Vector<T>

actual Vector<T>

Returns

Vector<T>

CalculateLoss(Tensor<T>, Tensor<T>)

Calculates the DETR set loss.

public T CalculateLoss(Tensor<T> predicted, Tensor<T> targets)

Parameters

predicted Tensor<T>

Predicted tensor containing class logits and boxes.

targets Tensor<T>

Target tensor containing ground truth.

Returns

T

Combined loss value.

Remarks

Expected shapes:

• predicted: [batch, num_queries, num_classes + 4] (logits + boxes)
• targets: [batch, max_objects, 1 + 4] (class + boxes, padded)

CalculateLoss(Vector<T>, Vector<T>)

Calculates the DETR loss using flattened vectors (simplified for interface compatibility).

public override T CalculateLoss(Vector<T> predicted, Vector<T> actual)

Parameters

predicted Vector<T>

actual Vector<T>

Returns

T

Remarks

For DETR, the tensor-based CalculateLoss overload is preferred as it preserves the structured format needed for Hungarian matching. This vector-based method provides a basic L1 loss between predicted and actual values.