Struct NBestDecoder 

pub struct NBestDecoder {
    pub n: usize,
    pub eos_id: u32,
    pub max_len: usize,
    pub length_penalty: f32,
}

Decoder that expands hypotheses by one step and maintains an N-best list.

Fields

n: usize

Maximum number of hypotheses to track.

eos_id: u32

Token ID that marks end-of-sequence.

max_len: usize

Maximum generation length (inclusive).

length_penalty: f32

Length-penalty exponent α used for normalised scoring.

Implementations

impl NBestDecoder

pub fn new(n: usize, eos_id: u32, max_len: usize) -> Self

Create a new decoder.

pub fn with_length_penalty(self, alpha: f32) -> Self

Set the length-penalty exponent (builder pattern).

pub fn step( &self, hypotheses: &[Hypothesis], logits_per_hyp: &[Vec<f32>], top_k: usize, ) -> Vec<Hypothesis>

Expand a batch of hypotheses by one step.

logits_per_hyp[i] must be the logit vector for hypotheses[i]. Returns the flat list of expanded hypotheses (up to top_k per input).

pub fn init(&self) -> NBestList

Return an empty N-best list with this decoder’s capacity.

pub fn partition(hyps: Vec<Hypothesis>) -> (Vec<Hypothesis>, Vec<Hypothesis>)

Partition hypotheses into (active, complete).