Construct a live stat from any StatSpec, dispatching on its modality. Useful for code paths (like StatSchemaDef.materialize) that iterate over an erased Map<String, StatSpec> and don't statically know the modality.

minMaxScaleFeatures

fun <R : Result> RegressionStatSpec<R>.minMaxScaleFeatures(targetLow: Double = 0.0, targetHigh: Double = 1.0): RegressionStatSpec<R>

Element-wise min-max scale a regression spec's feature vector.

sample

fun <R : Result> RegressionStatSpec<R>.sample(rate: Double, seed: Long): RegressionStatSpec<R>

Wrap this regression spec to keep each update with probability rate; seed feeds the PRNG.

standardScaleFeatures

fun <R : Result> RegressionStatSpec<R>.standardScaleFeatures(): RegressionStatSpec<R>

Element-wise standardise a regression spec's feature vector.

throttle

fun <R : Result> RegressionStatSpec<R>.throttle(every: Int): RegressionStatSpec<R>

Wrap this regression spec so it only sees one in every every updates.

transformX

fun <R : Result> RegressionStatSpec<R>.transformX(expr: VectorExpr): RegressionStatSpec<R>

Wrap this regression spec so x is remapped by expr before the inner stat sees it.

transformY

fun <R : Result> RegressionStatSpec<R>.transformY(expr: ScalarExpr): RegressionStatSpec<R>

Wrap this regression spec so y is remapped by expr before the inner stat sees it.

weightBy

fun <R : Result> RegressionStatSpec<R>.weightBy(expr: ScalarExpr): RegressionStatSpec<R>

Wrap this regression spec so every update's weight is multiplied by expr.eval(0, y, v).

withWeight

fun <R : Result> RegressionStatSpec<R>.withWeight(weight: Double): RegressionStatSpec<R>

Wrap this regression spec so every update uses weight regardless of caller input.

SoftmaxRegression

constructor(featureSize: Int, numClasses: Int, optimizer: OptimizerSpec = Sgd(), biasOptimizer: OptimizerSpec = optimizer)(source)

biasOptimizer

val biasOptimizer: OptimizerSpec(source)

Bias optimizer; defaults to optimizer.

featureSize

val featureSize: Int(source)

Number of input features.

numClasses

val numClasses: Int(source)

Number of classes; the input y must round to [0, numClasses).

optimizer

val optimizer: OptimizerSpec(source)

Per-class weight-matrix optimizer; one instance is materialised per class.