SAGAN<T>

Models & Types · AiDotNet.NeuralNetworks

Self-Attention GAN (SAGAN) — a deep convolutional GAN whose generator maps a 1D latent vector to an image and whose discriminator scores images as real or fake. SAGAN (Zhang et al. 2019) augments the DCGAN backbone with self-attention and spectral normalization for long-range coherence and training stability.

For Beginners: A GAN has two networks competing: a generator (the “artist”) that turns random noise into images, and a discriminator (the “critic”) that tells real images from generated ones. As in every GAN since Goodfellow et al. 2014, the generator’s INPUT is a 1D latent noise vector z (e.g. 128 numbers). Its first layer is a fully-connected projection that reshapes z into a small spatial feature map, which transposed convolutions then upsample into a full image. SAGAN’s distinctive ingredient is self-attention, letting each location attend to the whole feature map so global structure stays consistent.

This implementation derives from GenerativeAdversarialNetwork, which provides the proven adversarial training loop (tape-based generator and discriminator updates with BCE-with-logits, per Goodfellow 2014 §3 / Radford 2015), and builds paper-faithful generator/discriminator architectures: • Generator: latent[B, latentSize] → Dense → Reshape[C₀, s, s] → ⟨Deconv 4×4 stride 2 + BatchNorm + ReLU⟩×log₂(target/s) → Deconv → image[C, H, W] (Tanh). • Discriminator: image[C, H, W] → ⟨Conv 4×4 stride 2 (+ BatchNorm) + LeakyReLU⟩ → Conv → logit (no final sigmoid; BCE-with-logits is the criterion).

Based on “Self-Attention Generative Adversarial Networks” by Zhang et al. (2019).

Inheritance: NeuralNetworkBase<T> → DeclaredModelLayoutBase<T> → ImageGeneratorModelLayoutBase<T> → GenerativeAdversarialNetwork<T> → SAGAN<T>

Example

var noise = Tensor<float>.CreateRandom(new[] { 1, 128 });
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 SAGAN<float>(latentSize: 128, imageChannels: 3, imageHeight: 64, imageWidth: 64))
    .Build(trainX, trainY);
var generated = result.Predict(noise);

Constructors

Constructor Summary
SAGAN(NeuralNetworkArchitecture<T> architecture, SAGANOptions options) Convenience constructor deriving image dimensions from a single architecture.
SAGAN(NeuralNetworkArchitecture<T> generatorArchitecture, NeuralNetworkArchitecture<T> discriminatorArchitecture, SAGANOptions options, int[] attentionLayers, ILossFunction<T> lossFunction) Backward-compatible constructor that accepts explicit generator/discriminator architectures.
SAGAN(int latentSize, int imageChannels, int imageHeight, int imageWidth, SAGANOptions options, int[] attentionLayers, ILossFunction<T> lossFunction) Initializes a SAGAN with paper-faithful architectures built from the image dimensions and latent size.

Properties

Property Summary
bool UseSpectralNormalization { get; set; } Gets or sets whether spectral normalization is used.
int LatentSize { get; } Gets the size of the latent (noise) vector input to the generator.
int NumClasses { get; } Gets the number of classes for conditional generation (0 for unconditional).
int[] AttentionLayers { get; } Gets the indices of layers where self-attention is applied (Zhang et al.

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()