Struct SparseRunningStatistics 

pub struct SparseRunningStatistics<T>
where
    T: Float,
{ /* private fields */ }

Running statistics that accepts sparse column input but stores sufficient statistics in dense vectors.

This is more efficient than RunningStatistics<Ix1> when the input data is sparse and has many rows, as it avoids materializing dense matrices during reads.

Implementations

impl<T> SparseRunningStatistics<T>

where T: Float + AddAssign + Sum + Zero,

pub fn new(nrows: usize) -> Self

Create a new SparseRunningStatistics object

Arguments

• nrows - Number of rows (features)

pub fn add_sparse_column(&mut self, row_indices: &[usize], values: &[T])

Add a sparse column (row indices + values) to the running statistics. The column advances ncols_processed by one.

pub fn nrows(&self) -> usize

pub fn ncols_processed(&self) -> usize

pub fn to_vecs(&self) -> (Vec<T>, Vec<T>, Vec<T>, Vec<T>)

Convert to owned vectors (npos, sum, mean, std)

pub fn add_csc(&mut self, csc: &CscMatrix<T>)

Add columns from a CscMatrix

Arguments

• csc - Sparse matrix in CSC format

pub fn add_dense_column(&mut self, values: &[T])

Add a dense column directly. Skips zero / non-finite entries so npos stays equivalent to what add_sparse_column would produce — useful when an upstream coarsening step yields a dense [D, n] intermediate we don’t want to re-sparsify.

Loop is structured for autovectorization: hoists the running-sum and running-square accumulations through zip_eq over independent &mut slices (no aliasing), and uses a branchless 1/0 mask for the npos increment so the compiler can emit cmov / vmaskmovps instead of a control-flow branch per element.

pub fn add_dense_column_scaled(&mut self, values: &[T], scale: T)

Self::add_dense_column with every value scaled by scale on the way in. Folding the multiply into the accumulation loop touches each element once, where scale-into-a-buffer-then-add would write and re-read the whole column.

pub fn add_dense_columns(&mut self, dense: &DMatrix<T>)

where T: Scalar,

Add every column of a dense [D, n] matrix in column-major order. Calls add_dense_column per column so the inner loop stays vectorizable; the per-column overhead is negligible (one += 1 for ncols_processed) compared to the per-element accumulation work.

pub fn merge(&mut self, other: &Self)

Combine another SparseRunningStatistics into this one. Used to reduce per-thread accumulators back to a single result without holding a global lock during the streaming pass.

impl<T> SparseRunningStatistics<T>

where T: Float + AddAssign + Sum + Zero + Display + ToPrimitive,

pub fn save(&self, filename: &str, names: &[Box<str>], sep: &str) -> Result<()>

Save the statistics to a file (parquet if filename ends with .parquet, otherwise a separator-delimited text file).

pub fn to_f32_vecs(&self) -> (Vec<f32>, Vec<f32>, Vec<f32>, Vec<f32>)

Get statistics as vectors (nnz, tot, mu, sig) converted to f32

pub fn to_string_vec( &self, names: &[Box<str>], sep: &str, ) -> Result<Vec<Box<str>>>

Convert statistics to string vectors for output

Trait Implementations

impl<T> Clone for SparseRunningStatistics<T>

where T: Float + Clone,

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<T> RunningStatOps<T> for SparseRunningStatistics<T>

where T: Float + AddAssign + Sum + Zero,

fn count_positives(&self) -> Vec<T>

Count of positive (non-zero) values per row

fn sum(&self) -> Vec<T>

Sum per row

fn mean(&self) -> Vec<T>

Mean per row Uses ncols_processed as the denominator (implicit zeros count)

fn variance(&self) -> Vec<T>

Variance per row

fn std(&self) -> Vec<T>

Standard deviation per row

type Output = Vec<T>

fn clear(&mut self)