Struct ToroidalMask 

pub struct ToroidalMask {
    pub seq_len: usize,
    pub radius: f32,
    pub alpha: f32,
    pub grid_size: usize,
    pub mask_type: MaskType,
}

Toroidal attention mask generator.

Creates masks according to Eq. 17 from the paper:

M_Tonnetz(i, j) = 1                        if d_Tonnetz(i, j) ≤ r
                  exp(-α · d_Tonnetz(i,j)) otherwise

Fields

seq_len: usize

Sequence length

radius: f32

Locality radius (hard cutoff)

alpha: f32

Decay rate for soft falloff

grid_size: usize

Grid size for Tonnetz mapping

mask_type: MaskType

Mask type variant

Implementations

impl ToroidalMask

pub fn new(seq_len: usize, radius: f32, alpha: f32) -> Self

Create a new toroidal mask configuration (hybrid by default).

pub fn with_grid( seq_len: usize, radius: f32, alpha: f32, grid_size: usize, ) -> Self

Create with custom grid size.

pub fn hard_cutoff(seq_len: usize, radius: f32, grid_size: usize) -> Self

Create hard cutoff mask: M(i,j) = 1 if d <= r, else 0

pub fn soft_exponential(seq_len: usize, alpha: f32, grid_size: usize) -> Self

Create soft exponential mask: M(i,j) = exp(-α * d)

pub fn value(&self, i: usize, j: usize) -> f32

Compute mask value for position pair (i, j).

pub fn generate(&self) -> Vec<Vec<f32>>

Generate full mask matrix.

pub fn generate_doubly_stochastic(&self, iterations: usize) -> Vec<Vec<f32>>

Generate and apply Sinkhorn-Knopp to make doubly-stochastic.

Trait Implementations

impl Clone for ToroidalMask

fn clone(&self) -> ToroidalMask

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ToroidalMask

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

Formats the value using the given formatter.