ContrastiveLoss<T>

Models & Types · AiDotNet.LossFunctions

Implements the Contrastive Loss function for learning similarity metrics.

Inheritance: LossFunctionBase<T> → ContrastiveLoss<T>

For Beginners

Contrastive Loss helps a model learn to identify whether two examples are similar or dissimilar. It works with pairs of examples and their similarity label (1 for similar, 0 for dissimilar).

For similar pairs, the loss penalizes distance between them, encouraging them to be close together. For dissimilar pairs, the loss penalizes proximity below a certain margin, encouraging them to be at least that far apart.

The formula has two components:

  • For similar pairs (y=1): distance²
  • For dissimilar pairs (y=0): max(0, margin - distance)²

Contrastive Loss is commonly used in:

  • Siamese neural networks
  • Face verification systems (determining if two faces are the same person)
  • Signature verification
  • Any situation where you need to learn a similarity metric between pairs

This approach is simpler than Triplet Loss as it only requires pairs of examples rather than triplets.

Example

using AiDotNet.LossFunctions;
using AiDotNet.Tensors.LinearAlgebra;

var loss = new ContrastiveLoss<float>();
var predicted = new Vector<float>(new[] { 0.9f, 0.2f, 0.7f });
var actual = new Vector<float>(new[] { 1.0f, 0.0f, 1.0f });

float value = loss.CalculateLoss(predicted, actual);
Console.WriteLine($"ContrastiveLoss = {value:F4}");

Constructors

Constructor Summary
ContrastiveLoss(double margin) Initializes a new instance of the ContrastiveLoss class.

Methods

Method Summary
T CalculateLoss(Vector<T> output1, Vector<T> output2, T similarityLabel) Calculates the Contrastive Loss between two output vectors based on their similarity.
T CalculateLoss(Vector<T> predicted, Vector<T> actual) This method is not used for Contrastive Loss as it requires two input vectors and a similarity label.
Tensor<T> ComputeTapeLoss(Tensor<T> output1, Tensor<T> output2, T similarityLabel) Computes the contrastive loss as a tape-differentiable scalar from an embedding PAIR and its similarity label.
Tensor<T> ComputeTapeLoss(Tensor<T> predicted, Tensor<T> target)
ValueTuple<T, Tensor<T>, Tensor<T>> CalculateLossAndGradientGpu(Tensor<T> output1, Tensor<T> output2, Tensor<T> labels) Calculates Contrastive Loss on GPU for batched input tensors.