Module nmf_hals 

Implementation of NMF, dense and sparse paths. This version implements HALS with random or NNDSVD initialisation.

Modules

consensus

Consensus NMF: turns the random restarts from super::stabilised_nmf into a stability score and a consensus factor, plus a driver loop over k so the rank can be chosen from the stability and error curves.

dense

Dense backend for NMF. Borrows V as a faer MatRef and routes both data products through utils::gemm::gemm.

nmf_preprocessing

Pre-processing methods for NMF on sparse and dense data.

refit

Fixed-factor NNLS refits. Solves for one factor while the other is frozen, which is what a consensus NMF needs once the consensus factor has been assembled from the clustered restarts.

sparse

Sparse backend for NMF. Holds the chosen layer of V as CSC and CSR with storage type S; widens to compute type F per element inside two hand-rolled spMM kernels.

Structs

HalsOpts

HALS solver options.

NmfResult

Result of a single NMF run.

StabilisedNmfResult

Result of stabilised NMF across random restarts.

Enums

NmfInit

Initialisation strategy for HALS.

Traits

NmfInput

Backend-agnostic view of V for HALS NMF. Implementations write the two data-touching products into caller-owned buffers.

Functions

compute_objective

Frobenius squared reconstruction error.

gram_h_ht

Gram matrix H H^T.

gram_wt_w

Gram matrix W^T W.

hals_sweep_cols

HALS column sweep for W.

hals_sweep_rows

HALS row sweep for H.

nmf_hals

Frobenius HALS NMF.

nndsvd_from_svd

Boutsidis-Gallopoulos NNDSVD from a top-k truncated SVD.

normalise_w_cols

Normalise columns of W to unit L2 norm.

parse_nmf_init

Parse the NMF initialisation

random_init

Random initialisation

stabilised_nmf

Stabilised NMF via random restarts.