Struct SparseTensor 

pub struct SparseTensor<T: TensorElement> { /* private fields */ }

Sparse tensor in COO (Coordinate) format

COO format stores sparse tensors efficiently by only keeping track of non-zero elements and their coordinates. This is particularly useful for tensors with high sparsity ratios.

Format

• indices: N x D matrix where N is the number of non-zero elements and D is the number of dimensions
• values: N-length vector containing the non-zero values
• shape: The shape of the full dense tensor

Example

use torsh_tensor::sparse::SparseTensor;

// Create a 3x3 sparse matrix with values at (0,0)=1.0, (1,2)=2.0, (2,1)=3.0
let indices = vec![vec![0, 0], vec![1, 2], vec![2, 1]];
let values = vec![1.0, 2.0, 3.0];
let shape = vec![3, 3];

let sparse = SparseTensor::from_coo(indices, values, shape).expect("sparse tensor creation should succeed");

Implementations

impl<T: TensorElement> SparseTensor<T>

pub fn from_coo( indices: Vec<Vec<usize>>, values: Vec<T>, shape: Vec<usize>, ) -> Result<Self>

Create a new sparse tensor from COO format

Arguments

• indices - Coordinates of non-zero elements (each inner vector is one coordinate)
• values - Non-zero values corresponding to the indices
• shape - Shape of the full dense tensor

Returns

A new sparse tensor in COO format

Errors

Returns error if indices and values have mismatched lengths or if coordinates are out of bounds

pub fn from_dense(dense: &Tensor<T>, tolerance: T) -> Result<Self>

where T: Copy + PartialOrd + Zero + Signed,

Create a sparse tensor from a dense tensor by extracting non-zero elements

Arguments

• dense - The dense tensor to convert
• tolerance - Values with absolute value below this threshold are considered zero

Returns

A new sparse tensor containing only the non-zero elements

pub fn to_dense(&self) -> Result<Tensor<T>>

where T: Copy + Zero,

Convert this sparse tensor to a dense tensor

Returns

A dense tensor with zeros filled in for missing elements

pub fn nnz(&self) -> usize

Get the number of non-zero elements

pub fn shape(&self) -> &[usize]

Get the shape of the sparse tensor

pub fn device(&self) -> DeviceType

Get the device where the tensor resides

pub fn indices(&self) -> &[Vec<usize>]

Get the indices (coordinates) of non-zero elements

pub fn values(&self) -> &[T]

Get the non-zero values

pub fn sparsity(&self) -> f64

Calculate the sparsity ratio (fraction of zero elements)

pub fn memory_usage(&self) -> usize

Get the memory footprint of the sparse representation in bytes

pub fn memory_efficiency(&self) -> f64

Compare memory usage with equivalent dense representation

pub fn add(&self, other: &Self) -> Result<Self>

where T: Copy + Add<Output = T> + Zero + PartialEq,

Element-wise addition with another sparse tensor

Arguments

• other - The other sparse tensor to add

Returns

A new sparse tensor containing the sum

pub fn mul(&self, other: &Self) -> Result<Self>

where T: Copy + Mul<Output = T> + Zero + PartialEq,

Element-wise multiplication with another sparse tensor

For sparse tensors, multiplication only produces non-zero results where both tensors have non-zero elements at the same location.

pub fn mul_scalar(&self, scalar: T) -> Result<Self>

where T: Copy + Mul<Output = T> + Zero + PartialEq,

Scalar multiplication

pub fn matmul(&self, other: &Self) -> Result<Self>

where T: Copy + Add<Output = T> + Mul<Output = T> + Zero + PartialEq,

Matrix multiplication for 2D sparse tensors

Arguments

• other - The other 2D sparse tensor to multiply with

Returns

A new sparse tensor containing the matrix product

pub fn transpose(&self) -> Result<Self>

where T: Copy,

Transpose a 2D sparse tensor

pub fn map<F>(&self, f: F) -> Result<Self>

where F: Fn(T) -> T, T: Copy + Zero + PartialEq,

Apply a function to all non-zero values

pub fn is_valid(&self) -> bool

Check if the sparse tensor is structurally valid

pub fn coalesce(&mut self) -> Result<()>

where T: Copy + AddAssign + Zero + PartialEq,

Remove duplicate indices by summing their values

impl<T: TensorElement> SparseTensor<T>

Conversion utilities for sparse tensors

pub fn eye(size: usize) -> Result<Self>

where T: Copy + One,

Create a sparse identity matrix

pub fn from_triplets( rows: Vec<usize>, cols: Vec<usize>, vals: Vec<T>, shape: Vec<usize>, ) -> Result<Self>

Create a sparse tensor from triplets (row, col, value) for 2D case

Trait Implementations

impl<T: Clone + TensorElement> Clone for SparseTensor<T>

fn clone(&self) -> SparseTensor<T>

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl<T: Debug + TensorElement> Debug for SparseTensor<T>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.