InfoGAN<T>

Models & Types · AiDotNet.NeuralNetworks

Represents an Information Maximizing Generative Adversarial Network (InfoGAN), which learns disentangled representations in an unsupervised manner by maximizing mutual information between latent codes and generated observations.

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

For Beginners

InfoGAN learns to separate different features automatically.

Key concept:

  • Splits random input into two parts:
  1. Random noise (z): provides variety
  2. Latent codes (c): control specific features
  • Learns what each code controls WITHOUT labels
  • Example: For faces, might learn codes for:
  • Code 1: controls rotation
  • Code 2: controls width
  • Code 3: controls lighting

How it works:

  • Generator uses both z and c to create images
  • Auxiliary network Q tries to predict c from the generated image
  • If Q can predict c accurately, the codes are meaningful
  • This forces codes to represent interpretable features

Use cases:

  • Discover semantic features in datasets
  • Disentangled representation learning
  • Controllable image generation
  • Feature manipulation (change one aspect, keep others)

Reference: Chen et al., “InfoGAN: Interpretable Representation Learning by Information Maximizing Generative Adversarial Nets” (2016)

How It Works

InfoGAN extends the GAN framework by:

  • Decomposing the input noise into incompressible noise (z) and latent codes (c)
  • Maximizing the mutual information I(c; G(z,c)) between codes and generated images
  • Learning interpretable and disentangled representations automatically
  • Using an auxiliary network Q to approximate the posterior P(c|x)
  • Enabling control over semantic features without labeled data

Example

var architecture = new NeuralNetworkArchitecture<float>(inputFeatures: 8, outputSize: 4);
var noise = Tensor<float>.CreateRandom(new[] { 1, 74 });
var trainX = Tensor<float>.CreateRandom(4, 8);
var trainY = Tensor<float>.CreateRandom(4, 2);
var result = new AiModelBuilder<float, Tensor<float>, Tensor<float>>()
    .ConfigureModel(new InfoGAN<float>(architecture))
    .Build(trainX, trainY);
var generated = result.Predict(noise);

Constructors

Constructor Summary
InfoGAN(NeuralNetworkArchitecture<T> architecture, InfoGANOptions options) Creates an InfoGAN with default architectures derived from a single architecture.
InfoGAN(NeuralNetworkArchitecture<T> generatorArchitecture, NeuralNetworkArchitecture<T> discriminatorArchitecture, NeuralNetworkArchitecture<T> qNetworkArchitecture, int latentCodeSize, InputType inputType, InfoGANOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> generatorOptimizer, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> discriminatorOptimizer, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> qNetworkOptimizer, ILossFunction<T> lossFunction) Initializes a new instance of the InfoGAN class with the specified architecture and training parameters.

Properties

Property Summary
NeuralNetworkBase<T> Discriminator { get; set; } Gets the discriminator network.
NeuralNetworkBase<T> Generator { get; set; } Gets the generator network.
NeuralNetworkBase<T> QNetwork { get; set; } Gets the auxiliary Q network that predicts latent codes from images.

Methods

Method Summary
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input) Surfaces named activations from each sub-network so introspection invariants (NamedLayerActivations_ShouldBeNonEmpty) see real per-network state.
ModelMetadata<T> GetModelMetadata()
Tensor<T> ForwardForTraining(Tensor<T> input) Defines the InfoGAN forward graph for tape-based training.
Tensor<T> Generate(Tensor<T> noise, Tensor<T> latentCodes) Generates images with specific latent codes.
Tensor<T> GenerateRandomLatentCodes(int batchSize) Generates random latent codes: each categorical code one-hot over a uniformly drawn value, then each continuous code from Uniform(-1, 1).
Tensor<T> GenerateRandomNoiseTensor(int batchSize, int noiseSize) Generates random noise tensor using vectorized Gaussian noise generation with CPU/GPU acceleration.
ValueTuple<T, T, T> TrainStep(Tensor<T> realImages, Tensor<T> noise, Tensor<T> latentCodes) Performs one training step for InfoGAN.
void ResetOptimizerState() Resets the state of all optimizers to their initial values.