skainet-kan/sk.ainet.lang.kan

Package-level declarations

Types

Public NN module alias for Kolmogorov–Arnold Network (AKN). This exposes the same implementation the DSL uses (KanLayer) for direct creation without the DSL.

KAN

interface KAN<T : DType, V> : NetworkDslItem

DSL surface for configuring a Kolmogorov–Arnold Network layer.

KanConfig

data class KanConfig(val inputDim: Int, val outputDim: Int, val gridSize: Int = 16, val degree: Int = 3, val useBias: Boolean = true, val useResidual: Boolean = false, val regularization: KanRegularization = KanRegularization(), val gridMin: Float = 0.0f, val gridMax: Float = (PI / 2).toFloat())

Configuration for a Kolmogorov–Arnold Network layer.

KanDslImpl

class KanDslImpl<T : DType, V>(val executionContext: ExecutionContext, dtype: KClass<T>, initialOutputDim: Int, id: String) : KAN<T, V>

KanLayer

class KanLayer<T : DType, V>(val config: KanConfig, val baseActivation: (Tensor<T, V>) -> Tensor<T, V>, initMixingWeights: Tensor<T, V>, initBasis: Tensor<T, V>, initBias: Tensor<T, V>?, val name: String) : Module<T, V> , ModuleParameters<T, V>

Stub implementation for a Kolmogorov–Arnold Network layer.

KanRegularization

data class KanRegularization(val splineSmoothness: Double = 0.0, val sparsity: Double = 0.0)

Simple regularization hints for future spline/mixing penalties.

Functions

createAkn

fun <T : DType, V> createAkn(executionContext: ExecutionContext, dtype: KClass<T>, inputDim: Int, outputDim: Int, gridSize: Int = 16, degree: Int = 3, useBias: Boolean = true, useResidual: Boolean = false, name: String = "akn", baseActivation: (Tensor<T, V>) -> Tensor<T, V> = { it }, weightsInit: WeightsScope<T, V>.(Shape) -> Tensor<T, V> = { randn(std = 0.02f) }, basisInit: WeightsScope<T, V>.(Shape) -> Tensor<T, V> = { uniform(min = -0.5f, max = 0.5f) }, biasInit: BiasScope<T, V>.(Shape) -> Tensor<T, V> = { zeros() }): Akn<T, V>

Factory to create an AKN module directly (without the kanLayer DSL helper). Mirrors defaults and initializer hooks used by the DSL.

kanLayer

inline fun <T : DType, V> NeuralNetworkDsl<T, V>.kanLayer(outputDim: Int, gridSize: Int = 16, degree: Int = 3, useBias: Boolean = true, useResidual: Boolean = false, id: String = "", noinline baseActivation: (Tensor<T, V>) -> Tensor<T, V> = { it }, content: KAN<T, V>.() -> Unit = {})

Adds a KAN layer to the existing network DSL.