Lodestar.Decomposition 0.1.1. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

NmfOptions

What Nmf.Fit is allowed to vary. Every property has an initialiser, so new NmfOptions() is scikit-learn's own default configuration and you set only what you are changing.

Properties

Property	Default	What it does
BetaLoss	Frobenius	What the factorization minimises — see NmfBetaLoss.
Initialization	NndSvd	Where the iteration starts — see NmfInitialization. Ignored by the overload handed W₀ and H₀.
MaxIterations	200	The iteration cap, and the exact iteration count when Tolerance is zero. scikit-learn's default is the same 200.
Tolerance	1e-4	The relative improvement below which the loop stops, measured every tenth iteration. Zero disables the stop.
Seed	0	Seeds this package's generator for the initialisation's Ω when RandomMatrix is null. It reproduces a run of Lodestar, never a run of NumPy.
RandomMatrix	null	Ω itself, row-major and FeatureCount × (componentCount + 10). Given, it replaces the draw entirely.

The ten in that shape is not this type's to choose. NNDSVD reaches for the randomized SVD with scikit-learn's own defaults rather than the ones TruncatedSvdOptions exposes, and ten oversamples is one of them — so the block an initialisation wants is wider than the rank by exactly that, and an Ω of any other length is refused rather than silently reshaped. See ADR 0072 for why Ω is an input at all.

Tolerance = 0 is a feature, not a way to disable a safety net. With the stop off, MaxIterations stops being a cap and becomes the number of updates that will run, which is what makes two implementations comparable: an early stop that fires one check apart leaves two correct runs disagreeing on every digit. It is also the slowest setting, since nothing can end the loop early.

MaxIterations below one is refused, and so is a negative or NaN Tolerance. A BetaLoss and an Initialization are enums, and a value outside either falls back to nothing: the loss that is not KullbackLeibler is Frobenius, and the initialisation that is not NndSvda leaves the zeros alone.