WGAN<T>

Models & Types · AiDotNet.NeuralNetworks

Represents a Wasserstein Generative Adversarial Network (WGAN), which uses the Wasserstein distance (Earth Mover’s distance) to measure the difference between the generated and real data distributions.

Inheritance: NeuralNetworkBase<T> → DeclaredModelLayoutBase<T> → ImageGeneratorModelLayoutBase<T> → WGAN<T>

For Beginners

WGAN is an improved GAN that solves many training problems.

Key improvements over vanilla GAN:

  • More stable training (less likely to fail)
  • The loss value actually tells you how well training is going
  • No mode collapse issues (generating only a few types of outputs)
  • Can train the discriminator (critic) many times without problems

The main change is using a different mathematical way to measure the difference between real and fake images, which turns out to be much more stable.

Reference: Arjovsky et al., “Wasserstein GAN” (2017)

How It Works

WGAN addresses several training instabilities in vanilla GANs by:

  • Using Wasserstein distance instead of Jensen-Shannon divergence
  • Replacing the discriminator with a “critic” that doesn’t output probabilities
  • Enforcing a Lipschitz constraint through weight clipping
  • Providing a loss that correlates with image quality
  • Enabling more stable training and better convergence

Example

var options = new WGANOptions { };
var noise = Tensor<float>.CreateRandom(new[] { 1, 100 });
var trainX = Tensor<float>.CreateRandom(4, 8);
var trainY = Tensor<float>.CreateRandom(4, 2);
var generatorArchitecture = new NeuralNetworkArchitecture<float>(inputFeatures: 8, outputSize: 8);
var criticArchitecture = new NeuralNetworkArchitecture<float>(inputFeatures: 8, outputSize: 1);
var result = new AiModelBuilder<float, Tensor<float>, Tensor<float>>()
    .ConfigureModel(new WGAN<float>(
        generatorArchitecture, criticArchitecture,
        inputType: InputType.OneDimensional))
    .Build(trainX, trainY);
var generated = result.Predict(noise);

Constructors

Constructor Summary
WGAN(NeuralNetworkArchitecture<T> generatorArchitecture, NeuralNetworkArchitecture<T> criticArchitecture, InputType inputType, WGANOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> generatorOptimizer, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> criticOptimizer, ILossFunction<T> lossFunction) Initializes a new instance of the WGAN class.

Properties

Property Summary
NeuralNetworkBase<T> Critic { get; set; } Gets the critic network (called discriminator in vanilla GAN) that evaluates data.
NeuralNetworkBase<T> Generator { get; set; } Gets the generator network that creates synthetic data.

Methods

Method Summary
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input) Named activations of both subnetworks: the generator’s layers, then the critic’s over the image the generator produced.
Dictionary<string, double> EvaluateModel(int sampleSize) Evaluates the WGAN by generating images and calculating metrics.
ModelMetadata<T> GetModelMetadata()
Tensor<T> GenerateImages(Tensor<T> noise) Generates synthetic images using the generator.
Tensor<T> GenerateRandomNoiseTensor(int batchSize, int noiseSize) Generates a tensor of random noise for the generator.
ValueTuple<T, T> TrainStep(Tensor<T> realImages, Tensor<T> noise) Performs one training step for the WGAN using tensor batches.
void ResetOptimizerState() Resets both optimizer states for a fresh training run.