KAN Regression

This example trains a small Kolmogorov-Arnold Network on the prepared Auto MPG tabular-regression CSV. The downloader normalizes the columns to [0, 1], so the piecewise-linear KAN basis uses inputScale = gridSize - 1 to spread its knots across the data interval.

KAN is a model constructor. The task is chosen by the general trainer surface:

let trainer := Trainer.new model { task := .regression, optimizer := optim.adam { lr := 1e-3 } }

The edge basis is a normal config field. This example uses triangular piecewise-linear hats; a spline, polynomial, or rational edge family would plug in through the same nn.models.KANEdgeFamily slot, while the trainer continues to choose the task.

def NN.Examples.Models.Supervised.Kan.exeName : String

CLI subcommand name used in terminal banners and error messages.

def NN.Examples.Models.Supervised.Kan.defaultLogJson : System.FilePath

Default JSON loss-curve path for this command.

def NN.Examples.Models.Supervised.Kan.batch : ℕ

Static minibatch size for the Auto MPG tabular loader.

def NN.Examples.Models.Supervised.Kan.inDim : ℕ

Auto MPG has seven normalized numeric predictors after dropping the car name.

def NN.Examples.Models.Supervised.Kan.outDim : ℕ

Scalar regression target.

def NN.Examples.Models.Supervised.Kan.cfg : TorchLean.nn.models.KANConfig

KAN configuration using triangular edge bases over normalized tabular features.

@[reducible, inline]

abbrev NN.Examples.Models.Supervised.Kan.σ : Shape

@[reducible, inline]

abbrev NN.Examples.Models.Supervised.Kan.τ : Shape

def NN.Examples.Models.Supervised.Kan.model : TorchLean.nn.M (TorchLean.nn.Sequential σ τ)

Generic KAN model. Regression/classification is selected by Trainer, not by the model name.

def NN.Examples.Models.Supervised.Kan.data (path : System.FilePath) (seed : ℕ) : TorchLean.Trainer.Dataset σ τ

Prepared Auto MPG CSV as a public trainer dataset.

def NN.Examples.Models.Supervised.Kan.train (opts : TorchLean.Options) (flags : TorchLean.ModelZoo.CsvTrainFlags) : IO (TorchLean.Trainer.TrainResult σ τ)

Train the Auto MPG KAN with the public Trainer surface.

def NN.Examples.Models.Supervised.Kan.main (args : List String) : IO UInt32

CLI entrypoint for Auto MPG regression with a KAN model.