KernelInceptionDistance<T>

Models & Types · AiDotNet.Metrics

Kernel Inception Distance (KID) - A metric for evaluating the quality of generated images.

How It Works

KID measures how similar generated images are to real images using the Maximum Mean Discrepancy (MMD) with a polynomial kernel in the feature space of an Inception network.

Advantages over FID:

  • Unbiased estimator (FID is biased for small sample sizes)
  • Provides variance estimates for the metric
  • Works well with smaller datasets
  • More robust to sample size variations

Formula: KID = MMD^2(F_real, F_generated) using polynomial kernel k(x,y) = (x^T y / d + 1)^3

Typical KID scores (multiplied by 1000 for readability):

  • KID < 0.5: Excellent quality
  • KID 0.5-2.0: Good quality
  • KID 2.0-5.0: Moderate quality
  • KID > 5.0: Poor quality

Based on “Demystifying MMD GANs” by Binkowski et al. (2018)

Constructors

Constructor Summary
KernelInceptionDistance(ConvolutionalNeuralNetwork<T> featureNetwork, int featureDimension, int polynomialDegree, int numSubsets, int subsetSize) Initializes a new instance of KID calculator.

Properties

Property Summary
ConvolutionalNeuralNetwork<T> FeatureNetwork { get; } Gets the feature extraction network used for computing image representations.
int FeatureDimension { get; } Gets the dimensionality of extracted features.
int NumSubsets { get; set; } Gets or sets the number of subsets for computing variance estimates.
int PolynomialDegree { get; set; } Gets or sets the polynomial degree for the kernel.
int SubsetSize { get; set; } Gets or sets the subset size for variance computation.

Methods

Method Summary
Matrix<T> PrecomputeFeatures(Tensor<T> images) Pre-computes features for a set of images for efficient repeated evaluation.
ValueTuple<double, double> ComputeKIDWithVariance(Tensor<T> realImages, Tensor<T> generatedImages) Computes KID score with variance estimate using subset sampling.
double ComputeKID(Tensor<T> realImages, Tensor<T> generatedImages) Computes the KID score between real and generated images.
double ComputeKIDWithFeatures(Matrix<T> realFeatures, Tensor<T> generatedImages) Computes KID using pre-computed feature statistics.