Preprocessing standardscaler partialfit

Development build. This page describes main, not a released package. The latest published Lodestar.Preprocessing is 0.1.0 — read its documentation.

Home › Preprocessing › Feature scaling

StandardScaler.PartialFit

Folds another batch into the mean, the variance and the scale, as partial_fit does.

public StandardScaler PartialFit(ReadOnlySpan<double> samples)

Parameters — samples is the next batch, row-major, with FeatureCount values per row.

Returns — a new scaler summarising every batch seen so far; the one it was called on is unchanged.

Exceptions — ArgumentException when samples holds no row, a partial one, or a non-finite value.

Example — two batches, and the fit they add up to.

using Lodestar.Preprocessing;

double[] first = [1.0, 2.0, 4.0];
double[] second = [8.0, 16.0];

StandardScaler scaler = StandardScaler.Fit(first, featureCount: 1).PartialFit(second);

int seen = scaler.SampleCount;   // => 5
double mean = scaler.Mean![0];   // => 6.2

Remarks — the statistics are the ones a single fit over the concatenation gives, to a couple of units in the last place: the update is _incremental_mean_and_var's, Chan, Golub and LeVeque's parallel form, correction term included. Measured, the reference's own two paths differ by at most 3.3e-16 relative, and the corpus freezes both to say so.

It returns a new scaler and leaves this one alone, which is the one place the spelling differs from partial_fit. scaler = scaler.PartialFit(batch) reads the same, and a fitted object here never changes answer under a reader.

SampleCount stays one number where the reference's n_samples_seen_ becomes an array once a NaN appears — these scalers refuse a non-finite value, so the per-feature counts cannot diverge.

Applies to — net10.0, netstandard2.0.

See also — StandardScaler, StandardScaler.Fit.