Struct SparseDictFit 

pub struct SparseDictFit {
    pub decoder: Array2<f32>,
    pub indices: Array2<u32>,
    pub codes: Array2<f32>,
    pub explained_variance: f64,
    pub epochs: usize,
    pub converged: bool,
    pub active: usize,
}

Result of a collapsed-linear-lane fit.

The routing is stored fixed-width and sparse: indices[N, s] / codes[N, s]. There is deliberately no dense N×K assignment matrix — reconstructing it would defeat the purpose of the lane.

Fields

decoder: Array2<f32>

Decoder, K×P, unit-norm rows (one atom per row).

indices: Array2<u32>

Active atom indices per row, N×s (column j of row i is the j-th active atom for that row). Rows with fewer than s live atoms pad with repeated indices whose matching code is zero.

codes: Array2<f32>

Sparse codes per row, N×s, aligned with Self::indices.

explained_variance: f64

Final held-in explained variance (1 − RSS/TSS).

epochs: usize

Number of epochs actually run.

converged: bool

Whether the EV-improvement tolerance was reached.

active: usize

Active budget s actually used (min(active, K)).

Implementations

impl SparseDictFit

pub fn reconstruct(&self) -> Array2<f32>

Dense reconstruction N×P of the training rows from the sparse routing.

This does allocate an N×P array (the data size, not N×K); it exists for diagnostics / EV checks, not as part of the trainer’s hot loop.

Trait Implementations

impl Clone for SparseDictFit

fn clone(&self) -> SparseDictFit

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for SparseDictFit

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

Formats the value using the given formatter.