Struct MaskedCsMat 

pub struct MaskedCsMat<'a, T, I = u32, Iptr = u64>
where
    I: SpIndex,
    Iptr: SpIndex,
{ /* private fields */ }

A view exposing a subset of the rows and/or columns of a CSR matrix.

Both selections keep ascending original order, so masked index i is original index selected_rows()[i] (resp. selected_columns()[i]).

use single_svdlib::{sprs::TriMatI, MaskedCsMat, SparseMat, SvdMat};

let mut t = TriMatI::<f64, u32>::new((6, 5));
for i in 0..6 { for j in 0..5 { t.add_triplet(i, j, (i * 5 + j) as f64); } }
let a: SvdMat<f64> = t.to_csr::<u64>();

// Cells 0, 2, 4 by genes 1, 3 — a 3x2 matrix, without touching `a`.
let view = MaskedCsMat::submatrix(&a, Some(&[0, 2, 4]), Some(&[1, 3]));
assert_eq!((view.rows(), view.cols()), (3, 2));

Implementations

impl<'a, T, I, Iptr> MaskedCsMat<'a, T, I, Iptr>

where T: SvdFloat, I: SpIndex, Iptr: SpIndex,

pub fn submatrix( matrix: &'a CsMatI<T, I, Iptr>, rows: Option<&[usize]>, cols: Option<&[usize]>, ) -> Self

A view of the given rows and/or columns. None keeps that axis whole.

Index lists may be in any order and may repeat; the view presents each selected index once, ascending.

Panics

If any index is out of bounds, or the matrix is not CSR.

pub fn with_columns(matrix: &'a CsMatI<T, I, Iptr>, columns: &[usize]) -> Self

A view of the given columns, every row.

pub fn with_rows(matrix: &'a CsMatI<T, I, Iptr>, rows: &[usize]) -> Self

A view of the given rows, every column.

pub fn from_masks( matrix: &'a CsMatI<T, I, Iptr>, row_mask: Option<&[bool]>, col_mask: Option<&[bool]>, ) -> Self

A view built from boolean masks, one entry per original row/column.

Panics

If a mask’s length does not match the corresponding dimension.

pub fn selected_rows(&self) -> Option<&[usize]>

Selected original row indices, ascending. Empty slice when every row is kept.

pub fn selected_columns(&self) -> Option<&[usize]>

Selected original column indices, ascending. None when every column is kept.

pub fn is_identity(&self) -> bool

Whether the view is the identity, in which case products delegate straight to the underlying matrix.

pub fn uses_all_columns(&self) -> bool

Whether every column is retained.

pub fn inner(&self) -> &'a CsMatI<T, I, Iptr>

The matrix being viewed.

pub fn to_sparse(&self) -> CsMatI<T, I, Iptr>

Copy the view into an owned sparse matrix. Still sparse — a subset copy, not a densification — and the source is untouched.

A view doesn’t make products cheaper: each one still walks every source non-zero and checks it against the column table. Since a solver issues hundreds of products, a restrictive mask usually repays the one-off O(nnz) copy almost immediately. Stay with the view if the mask keeps most columns, if the copy wouldn’t fit alongside the source, or if you only need a few products.

Panics

If the extracted index range would overflow the index type I.

Trait Implementations

impl<T, I, Iptr> SparseMat<T> for MaskedCsMat<'_, T, I, Iptr>

where T: SvdFloat, I: SpIndex, Iptr: SpIndex,

fn squared_frobenius(&self) -> f64

Only the selected entries, so the norm describes the view and not the matrix it borrows from.

fn centered_squared_frobenius(&self, means: ArrayView1<'_, T>) -> f64

Selected entries only, summed per entry so a column offset can’t cancel it away.

fn rows(&self) -> usize

fn cols(&self) -> usize

fn nnz(&self) -> usize

fn mul_vec(&self, x: &[T], y: &mut [T], trans: bool)

y = A·x when trans is false, y = Aᵀ·x when true.

impl<T, I, Iptr> SparseMatDense<T> for MaskedCsMat<'_, T, I, Iptr>

where T: SvdFloat, I: SpIndex, Iptr: SpIndex,

fn col_means(&self) -> Array1<T>

Column means of the view — averaged over the selected rows only, so PCA on a row subset centers on that subset’s means rather than the whole matrix’s.

fn mul_dense( &self, rhs: ArrayView2<'_, T>, out: ArrayViewMut2<'_, T>, trans: bool, )

out = A·rhs when trans is false, out = Aᵀ·rhs when true.

fn mul_dense_centered( &self, rhs: ArrayView2<'_, T>, out: ArrayViewMut2<'_, T>, trans: bool, means: ArrayView1<'_, T>, )

The product against A - 1·meansᵀ, without ever forming it — centering a sparse matrix would destroy its sparsity, so it goes in as the rank-1 update it is: