Module decoders 

Stochastic decoders for autoregressive sequence generation.

This module collects sampling-based decoding strategies that are complementary to the greedy/beam-search decoders in crate::beam. Each submodule implements a different truncation criterion on the softmax distribution before drawing a single token:

• top_k — Fan, Lewis & Dauphin (2018), Hierarchical Neural Story Generation: restrict the support to the k most-likely tokens.
• nucleus — Holtzman, Buys, Du, Forbes & Choi (2020), The Curious Case of Neural Text Degeneration: restrict the support to the smallest set whose cumulative probability exceeds p.
• typical — Meister, Pimentel, Wiher & Cotterell (2022), Typical Decoding for Natural Language Generation: restrict the support to tokens whose negative log-probability is closest to the conditional entropy of the distribution.

All decoders are deterministic given a seeded crate::handle::LcgRng and return crate::error::SeqResult.

Re-exports

pub use contrastive::ContrastiveConfig;

pub use contrastive::ContrastiveSearcher;

pub use pointer_network::PointerGrad;

pub use pointer_network::PointerNetwork;

pub use nucleus::*;

pub use top_k::*;

pub use typical::*;

Modules

contrastive

Contrastive search decoding (Su et al. 2022 ACL).

nucleus

Nucleus (top-p) sampling for autoregressive sequence decoding.

pointer_network

Pointer Network.

top_k

Top-k sampling for autoregressive sequence decoding.

typical

Typical (locally typical) decoding for autoregressive sequence generation.