Loss Functions
37Comprehensive loss function library for every training objective
37 loss functions covering classification (CrossEntropy, Focal, ArcFace), regression (MSE, Huber, Quantile), segmentation (Dice, Tversky, Lovasz), and generative training (Perceptual, LPIPS, SSIM, InfoNCE). All with automatic differentiation and GPU support.
Classification Losses
Losses for categorical prediction tasks with class imbalance handling.
CrossEntropy
Standard cross-entropy with optional label smoothing and class weights.
BinaryCrossEntropy
Binary classification with pos_weight for imbalanced datasets.
FocalLoss
Down-weight easy examples to focus on hard cases (RetinaNet).
ArcFace / CosFace
Angular margin losses for discriminative face/embedding learning.
PolyLoss
Polynomial expansion of cross-entropy for improved tail-class accuracy.
Regression Losses
Losses for continuous value prediction with outlier robustness.
MSE
Mean Squared Error for standard regression objectives.
MAE / L1
Mean Absolute Error robust to outliers.
Huber
Smooth transition between MSE (small errors) and MAE (large errors).
LogCosh
Log of hyperbolic cosine, twice differentiable and outlier-robust.
Quantile
Asymmetric loss for predicting specific quantiles.
Segmentation Losses
Specialized losses for pixel-level prediction tasks.
Dice
Overlap-based loss for imbalanced segmentation masks.
Tversky
Generalized Dice with adjustable false positive/negative weighting.
Lovasz
Lovasz extension of IoU for direct mIoU optimization.
Boundary
Focus on boundary pixels for precise edge segmentation.
Generative & Contrastive Losses
Losses for image generation, representation learning, and similarity.
Perceptual (VGG)
Feature-space loss using VGG activations for perceptual quality.
LPIPS
Learned Perceptual Image Patch Similarity for generation quality.
SSIM
Structural Similarity Index for image quality assessment.
InfoNCE
Noise Contrastive Estimation for self-supervised representation learning.
NT-Xent
Normalized Temperature-scaled Cross Entropy for contrastive learning (SimCLR).
Triplet
Learn embeddings where positives are closer than negatives by a margin.
Loss functions with AiModelBuilder
using AiDotNet;
// Train with a custom loss function using AiModelBuilder
var result = await new AiModelBuilder<float, float[], float>()
.ConfigureModel(new NeuralNetwork<float>(
inputSize: 784, hiddenSize: 128, outputSize: 10))
.ConfigureLossFunction(new FocalLoss<float>(
gamma: 2.0f, alpha: 0.25f))
.ConfigureOptimizer(new AdamOptimizer<float>())
.BuildAsync(features, labels);
var prediction = result.Predict(newSample);Start building with Loss Functions
All 37 implementations are included free under Apache 2.0.