FrechetInceptionDistance<T>

Models & Types · AiDotNet.Metrics

Fréchet Inception Distance (FID) - A metric for evaluating the quality of generated images.

FID measures how similar generated images are to real images by comparing their statistical properties in a feature space. Lower FID scores indicate better quality.

The algorithm:

  1. Extract features from images using a pre-trained neural network
  2. Compute statistics (mean and covariance) for real and generated image features
  3. Compute the Fréchet distance between the two Gaussian distributions

Formula: FID = ||μ₁ - μ₂||² + Tr(Σ₁ + Σ₂ - 2√(Σ₁Σ₂)) where μ is mean, Σ is covariance, Tr is trace

Typical FID scores:

  • FID less than 10: Excellent quality
  • FID 10-20: Good quality
  • FID 20-50: Moderate quality
  • FID greater than 50: Poor quality

Based on “GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium” by Heusel et al. (2017)

Constructors

Constructor Summary
FrechetInceptionDistance(ConvolutionalNeuralNetwork<T> featureNetwork, int featureDimension) Initializes a new instance of FID 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 FeatureLayer { get; set; } Gets or sets the layer index from which to extract features.

Methods

Method Summary
T ComputeFID(Tensor<T> realImages, Tensor<T> generatedImages) Computes the FID score between real and generated images.
T ComputeFIDWithStats(Vector<T> realMean, Matrix<T> realCov, Tensor<T> generatedImages) Computes FID using pre-computed statistics for the real distribution.
ValueTuple<Vector<T>, Matrix<T>> PrecomputeStatistics(Tensor<T> images) Pre-computes statistics for a set of images.