skainet-kan/sk.ainet.lang.kan.examples/SineKanPretrained

SineKanPretrained

A tiny, fixed KAN-based approximator for y = sin(x) on the interval 0, π/2. It consists of a single KAN layer whose basis, mixer weights and bias were exported from the accompanying PyTorch script in piekan/train.py, and are embedded here as constants. At runtime, it builds a Module with those values frozen, serving as a lightweight, dependency-free example of loading pretrained KAN parameters in Kotlin.

What it is:

A minimal, ready-to-use example demonstrating how to instantiate a KAN layer with predetermined basis/weights/bias to approximate sin(x) on a narrow domain.
Deterministic and non-trainable in this form; it does not modify parameters during execution.

What it is not:

Not a general-purpose sine model outside 0, π/2; quality may degrade beyond the trained range.
Not a full KAN stack or training pipeline; it omits optimizers, loss, and data handling, focusing solely on inference with fixed parameters.

Properties

basisValues

val basisValues: FloatArray

biasValues

val biasValues: FloatArray

mixingValues

val mixingValues: FloatArray

Functions

create

fun create(executionContext: ExecutionContext): Module<FP32, Float>

Build the pretrained KAN module with fixed weights/basis/bias.

toGraphvizDot

fun toGraphvizDot(rankdir: String = "LR"): String

Produce a minimal Graphviz/DOT visualization of this tiny pretrained KAN. This does not depend on any external Graphviz library — it just returns a DOT string that you can render with the dot tool or any Graphviz viewer.