deepbox/ndarray

Sparse Matrices

Compressed Sparse Row (CSR) format for memory-efficient representation of matrices with many zero entries. Useful for large, sparse datasets and graph-based computations.

CSRMatrix.fromCOO

CSRMatrix.fromCOO(init: CSRMatrixInit): CSRMatrix

Create a CSR matrix from COO (Coordinate) format. Specify non-zero entries as row indices, column indices, and values arrays.

Parameters:

init.rows: number - Number of rows

init.cols: number - Number of columns

init.rowIndices: Int32Array - Row indices of non-zero entries

init.colIndices: Int32Array - Column indices of non-zero entries

init.values: Float64Array - Values of non-zero entries

CSRMatrix Methods

.add(other) — Sparse matrix addition
.sub(other) — Sparse matrix subtraction
.scale(scalar) — Multiply all values by a scalar
.multiply(other) — Element-wise multiplication (Hadamard)
.matvec(vector) — Matrix-vector multiplication
.matmul(other) — Sparse matrix multiplication
.transpose() — Transpose the sparse matrix
.toDense() — Convert to a dense Tensor
.nnz — Number of non-zero entries
.shape — [rows, cols] dimensions

sparse.ts

import { CSRMatrix } from "deepbox/ndarray";// Create sparse matrix from coordinate formatconst sparse = CSRMatrix.fromCOO({  rows: 3,  cols: 3,  rowIndices: new Int32Array([0, 1, 2]),  colIndices: new Int32Array([0, 2, 1]),  values: new Float64Array([1, 2, 3]),});// Represents: [[1, 0, 0], [0, 0, 2], [0, 3, 0]]console.log(sparse.nnz);    // 3 (non-zero entries)const dense = sparse.toDense(); // Full 3×3 tensor// Arithmeticconst scaled = sparse.scale(2);    // All values × 2const transposed = sparse.transpose();

When to Use

Matrices where >90% of entries are zero (e.g., adjacency matrices, document-term matrices)
Graph algorithms with sparse connectivity
Feature matrices with many zero-valued features
Memory-constrained environments where dense matrices would be too large

Autograd (GradTensor)

Decompositions