Struct LengthPenalty 

pub struct LengthPenalty { /* private fields */ }

Computes GNMT-style length and coverage penalties for beam-search hypothesis scoring.

Implementations

impl LengthPenalty

pub fn new(config: LengthPenaltyConfig) -> SeqResult<Self>

Create a new LengthPenalty. Returns Err if alpha < 0 or beta < 0.

pub fn lp(&self, length: usize) -> f64

GNMT length penalty: ((5 + length) / (5 + 1))^alpha.

At length=1: returns 1.0. Monotonically increasing for alpha > 0.

pub fn cp(&self, coverage_probs: &[f64], n_source: usize, seq_len: usize) -> f64

Coverage penalty: Σ_i log(min(Σ_t p_{t,i}, 1.0)).

coverage_probs has layout [seq_len × n_source] (for each target step t, attention over n_source source tokens).

Returns 0.0 when all source tokens are fully covered (sum ≥ 1.0). Returns a negative value when coverage is partial.

pub fn score( &self, log_prob: f64, length: usize, coverage_probs: &[f64], n_source: usize, ) -> SeqResult<f64>

Combined beam-search score.

score = log_prob / lp(length) - beta * |cp|

The magnitude of cp is used so that beta ≥ 0 always penalises under-coverage (cp is ≤ 0 when coverage < 1).

pub fn rank(&self, log_probs: &[f64], lengths: &[usize]) -> Vec<usize>

Rank hypotheses by descending combined score (no coverage penalty applied in this simplified batch form — coverage is assumed uniform).

Returns indices sorted best-first.

Trait Implementations

impl Clone for LengthPenalty

fn clone(&self) -> LengthPenalty

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LengthPenalty

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

Formats the value using the given formatter.