Dense - Klyn API

Dense

classin packageklyn.math.nn.layers

public class Dense<T extends AbstractFloat = Float> extends Layer<T>

Dense

An affine layer: output = inputs @ weights + bias, in row-major layout. Batches have shape [samples, inputs]; outputs have shape [samples, units]. Optional relu, sigmoid, tanh or softmax activation is included in the layer. Training contexts retain pre-activations only for the current batch, avoiding a second matrix product.

Example

layer = Dense<Float>(128, inputSize=784, seed=17)

Properties

Modifier and Type	Member	Description
public readonly property	activation `activation as String`	Configured activation identifier.
public readonly property	bias `bias as NDArray<T>`	Mutable zero-initialized bias [1, units].
public override readonly property	displayName `displayName as String`	Affine stage and configured activation name.
public override readonly property	inputShapeHint `inputShapeHint as IList<Long>`	Known input width, or an empty shape until build.
public readonly property	inputSize `inputSize as ULong`	Number of input features; zero until a deferred layer is built.
public override readonly property	parameterCount `parameterCount as ULong`	Two trainable matrices: weights, then bias.
public readonly property	seed `seed as Long`	Initialization seed, retained in checkpoints.
public readonly property	units `units as ULong`	Number of output features.
public readonly property	weights `weights as NDArray<T>`	Mutable parameters [inputs, units].

Constructors

Modifier and Type	Member	Description
public	Dense `Dense(units as Long, inputSize as Long = -1l, seed as Long = 1l, activation as String = "linear")`	Allocates parameters using seeded Glorot uniform initialization.

Methods

Modifier and Type	Member	Description
public	backward `backward(inputs as NDArray<T>, upstream as NDArray<T>) as DenseGradients<T>`	Computes exact affine derivatives using the current weights.
public override	backwardTraining `backwardTraining(context as LayerPass<T>, upstream as NDArray<T>) as LayerGradients<T>`	Differentiates the saved affine result without repeating its matrix product.
public	build `build(inputSize as Long) as Long`	Allocates seeded parameters once, or validates an existing input width.
public override	buildShape `buildShape(inputShape as IList<Long>) as IList<Long>`	Infers the input width from vector sample dimensions.
public override	differentiate `differentiate(inputs as NDArray<T>, upstream as NDArray<T>) as LayerGradients<T>`	Adapts affine derivatives to the common layer contract.
public override	forward `forward(inputs as NDArray<T>) as NDArray<T>`	Applies the affine transform followed by the configured activation.
public override	forwardTraining `forwardTraining(inputs as NDArray<T>) as LayerPass<T>`	Saves pre-activations in an explicit, batch-local backward context.
public override	parameter `parameter(index as Long) as NDArray<T>`	Returns weights (0) or bias (1) without copying.