Generative Models

Config-style constructors for runnable generative examples.

These models act on a trailing feature axis and preserve the caller's batchShape. Examples may flatten structured observations before applying them, while convolutional or operator-based models can use their own shape-specific constructors.

structure TorchLean.nn.models.Generative.Config : Type

Widths shared by dense generative models.

• dataWidth : ℕ

  Width of the data feature axis.

• hiddenWidth : ℕ

  Width of the hidden layers.

• latentWidth : ℕ

  Width of the latent representation.

@[instance_reducible]

instance TorchLean.nn.models.Generative.instReprConfig : Repr Config

def TorchLean.nn.models.Generative.instReprConfig.repr : Config → ℕ → Std.Format

def TorchLean.nn.models.Generative.Internal.validateSharedWidths (kind : String) (config : Config) : Except String Unit

Validate the data and hidden widths used by every dense generative model.

def TorchLean.nn.models.Generative.Internal.validateConfig (kind : String) (config : Config) : Except String Unit

Validate all widths while naming the public constructor being built.

def TorchLean.nn.models.Generative.Config.validate (config : Config) : Except String Unit

Validate every feature width before constructing or seeding a generative model.

@[reducible, inline]

abbrev TorchLean.nn.models.Generative.Config.dataShape (config : Config) (batchShape : Shape := []) : Shape

Data tensor shape with an arbitrary batch shape.

@[reducible, inline]

abbrev TorchLean.nn.models.Generative.Config.latentShape (config : Config) (batchShape : Shape := []) : Shape

Latent tensor shape with an arbitrary batch shape.

@[reducible, inline]

abbrev TorchLean.nn.models.Generative.Config.scoreShape (_config : Config) (batchShape : Shape := []) : Shape

Scalar-score tensor shape with an arbitrary batch shape.

def TorchLean.nn.models.Generative.autoencoder (config : Config) (batchShape : Shape := []) : Builder (Sequential (config.dataShape batchShape) (config.dataShape batchShape))

Autoencoder backbone: x -> hidden -> latent -> hidden -> reconstruction.

The reconstruction is unconstrained. Append an output activation such as nn.sigmoid when the data domain requires one.

def TorchLean.nn.models.Generative.generator (config : Config) (batchShape : Shape := []) : Builder (Sequential (config.latentShape batchShape) (config.dataShape batchShape))

Generator backbone z -> x.

The generated values are unconstrained. Choose an output activation at the call site to match the training data and objective.

def TorchLean.nn.models.Generative.discriminator (config : Config) (batchShape : Shape := []) : Builder (Sequential (config.dataShape batchShape) (config.scoreShape batchShape))

Discriminator x -> logits.

Returning logits keeps the model compatible with numerically stable objectives such as TorchLean.Loss.bceWithLogits. Append nn.sigmoid only when probabilities are required. Only the data and hidden widths are used; the latent width does not affect this model.