GanOptions

Options & Configuration · AiDotNet.NeuralNetworks.Options

Shared configuration for generative adversarial networks (DCGAN, WGAN, WGAN-GP, BigGAN, SAGAN, StyleGAN, ProgressiveGAN, InfoGAN, Pix2Pix, CycleGAN).

Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → GanOptions

Known derived types: BigGANOptions, CycleGANOptions, DCGANOptions, InfoGANOptions, Pix2PixOptions, ProgressiveGANOptions, SAGANOptions, StyleGANOptions, WGANGPOptions, WGANOptions

For Beginners

A GAN trains two networks against each other. The generator invents images; the discriminator (sometimes called the critic) tries to tell real images from invented ones. Each gets better because the other does. The settings here describe how big each of the two networks is and how they are trained.

How It Works

Derived options classes assign their paper’s values in their parameterless constructor. See ModelHyperparameterOptions for why these properties are non-nullable.

Properties

Property Summary
double InitialLearningRate { get; set; } Gets or sets the learning rate the adversarial pair starts training at.
int CriticIterations { get; set; } Gets or sets how many discriminator or critic updates run per generator update.
int DiscriminatorChannels { get; set; } Gets or sets the base channel count of the discriminator or critic.
int GeneratorChannels { get; set; } Gets or sets the base channel count of the generator.
int ImageChannels { get; set; } Gets or sets the number of colour channels in the generated image.
int LatentSize { get; set; } Gets or sets the length of the random noise vector the generator starts from.