Module generation 

Sequence generation and decoding algorithms.

This module provides decoding strategies for sequence-to-sequence models:

• Greedy decoding: Select the most likely token at each step
• Beam search: Maintain top-k candidates at each step
• Nucleus (top-p) sampling: Sample from the smallest set of tokens with cumulative probability >= p
• Top-k sampling: Sample from the k most likely tokens
• Temperature sampling: Scale logits before sampling

Example

use aprender::nn::generation::{BeamSearch, NucleusSampler, GenerationConfig};

// Beam search decoding
let beam = BeamSearch::new(5);  // beam_size=5
let output = beam.generate(&model, &input, 50);

// Nucleus sampling
let sampler = NucleusSampler::new(0.95);  // top_p=0.95
let output = sampler.sample(&logits);

References

• Holtzman, A., et al. (2020). The Curious Case of Neural Text Degeneration. ICLR. (Nucleus sampling)
• Freitag, M., & Al-Onaizan, Y. (2017). Beam Search Strategies for Neural Machine Translation. ACL Workshop.

Structs

BeamHypothesis

A single beam hypothesis in beam search.

BeamSearch

Beam search decoder for sequence generation.

GenerationConfig

Configuration for text generation.

GreedyDecoder

Greedy decoder - always selects the most likely token.

NucleusSampler

Nucleus (top-p) sampler for diverse text generation.

TeacherForcing

Teacher Forcing scheduler for seq2seq training.

TopKSampler

Top-k sampler for text generation.

Enums

TeacherForcingSchedule

Teacher forcing schedule type.

Functions

apply_repetition_penalty

Apply repetition penalty to logits.

apply_temperature

Apply temperature scaling to logits.