GAN CIFAR Example

Compact LSGAN-style executable path.

This trains:

• a generator z -> image toward the current CIFAR minibatch as a stable warm-up objective;
• a discriminator on real CIFAR images (1) and deterministic noise images (0).

The formal LSGAN objective decomposition lives in NN.Spec.Models.Gan and NN.MLTheory.Generative.Latent.GAN. A full alternating adversarial trainer can reuse the same generator/discriminator constructors and data path.

def NN.Examples.Models.Generative.Gan.exeName : String

CLI subcommand name used in terminal banners and error messages.

def NN.Examples.Models.Generative.Gan.defaultLogJson : System.FilePath

Default JSON loss-curve path for this command.

def NN.Examples.Models.Generative.Gan.cfg : TorchLean.nn.models.VectorGenerativeConfig

Shared vector-image configuration.

The generator, discriminator, latent batch, score batch, and CIFAR vector batch all derive from this record, so shape changes stay centralized.

@[reducible, inline]

abbrev NN.Examples.Models.Generative.Gan.Z : Shape

Latent-noise batch shape for the generator input.

@[reducible, inline]

abbrev NN.Examples.Models.Generative.Gan.X : Shape

Flattened CIFAR image-vector batch shape.

@[reducible, inline]

abbrev NN.Examples.Models.Generative.Gan.S : Spec.Shape

Discriminator score shape: one scalar score per batch row.

def NN.Examples.Models.Generative.Gan.mkGenerator : TorchLean.nn.M (TorchLean.nn.Sequential Z X)

Generator network mapping latent vectors to flattened image vectors.

def NN.Examples.Models.Generative.Gan.mkDiscriminator : TorchLean.nn.M (TorchLean.nn.Sequential X S)

Discriminator network mapping flattened image vectors to scalar real/fake scores.

def NN.Examples.Models.Generative.Gan.sampleMse {σ τ : Shape} (predict : TorchLean.Tensor.T Float σ → IO (TorchLean.Tensor.T Float τ)) (sample : TorchLean.SupervisedSample Float σ τ) : IO Float

Mean-squared error for one supervised sample evaluated through a public prediction closure.

def NN.Examples.Models.Generative.Gan.totalLoss (predictGen : TorchLean.Tensor.T Float Z → IO (TorchLean.Tensor.T Float X)) (predictDisc : TorchLean.Tensor.T Float X → IO (TorchLean.Tensor.T Float S)) (genSample : TorchLean.SupervisedSample Float Z X) (discReal discFake : TorchLean.SupervisedSample Float X S) : IO Float

Aggregate generator and discriminator scalar losses for one LSGAN reporting step.

The metric receives public prediction functions rather than raw modules. That is the whole point of this example after the trainer cleanup: the GAN-specific code still defines the task objective, but the model state, optimizer stepping, and backend details stay inside trainer.trainPairStreamFloat.

def NN.Examples.Models.Generative.Gan.trainCurve (opts : TorchLean.Options) (xPath yPath : System.FilePath) (nRows seed steps cudaMemWatch : ℕ) : IO TorchLean.Training.Curve

Train the compact LSGAN-style pair and return a total-loss curve.

The update rule is chosen for stable public runs: the generator first learns toward a fixed CIFAR minibatch, while the discriminator separates that minibatch from deterministic noise. The imported spec/theory modules carry the adversarial objective statements; this runtime path checks that both networks, optimizers, CUDA memory reporting, and real-data loading work together.

def NN.Examples.Models.Generative.Gan.main (args : List String) : IO UInt32

Executable entrypoint for the compact GAN-style run.

The command loads CIFAR vectors, trains generator and discriminator updates for --steps, and writes the combined loss curve to the requested logging destination.