Crate oxicuda_seq 

oxicuda-seq — Sequence Models & Structured Prediction for OxiCUDA.

Architecture

oxicuda-seq
├── hmm/          — Hidden Markov Models (discrete/Gaussian), forward-backward, Viterbi, Baum-Welch
├── crf/          — Linear-chain CRF: forward-backward in score space, L-BFGS training, Viterbi decoding
├── memm/         — Maximum-Entropy Markov Models
├── ssvm/         — Structured SVM (linear-chain) with cutting-plane optimisation
├── structured/   — Sinkhorn CRF: entropy-regularised optimal-transport normalisation
├── beam/         — Generic beam search with length normalisation and diversity penalty
├── ctc/          — CTC loss (forward-backward) + greedy / prefix-beam decoding
├── decoders/     — Stochastic decoders: top-k, nucleus (top-p), typical sampling
├── alignment/    — Needleman-Wunsch, Smith-Waterman, Gotoh affine-gap, Hirschberg
├── grid_crf/     — Pairwise 2D CRF + mean-field inference
├── kalman/       — Linear/EKF Kalman filter, RTS smoother, EM parameter learning
├── mrf/          — General MRF + Ising, Gibbs sampler, loopy belief propagation
├── perceptron/   — Structured + averaged perceptron sequence taggers (Collins 2002)
├── matching/     — Aho–Corasick multi-pattern matching automaton
├── string/       — Manacher longest-palindrome, suffix automaton (DAWG),
│                   Z-algorithm, suffix array + LCP (SA-IS/Kasai), BWT/FM-index
├── tagging/      — BIO/BIOES conversion + validation, span-based F1 (seqeval)
└── metrics/      — Token/sequence accuracy, edit distance, BLEU, perplexity, log-loss

NOTE on lints: numerical kernels in this crate use straight integer-index loops (for i in 0..n) because every body indexes multiple parallel arrays. Rewriting them in iterator form obscures the math and forces awkward zip chains, so clippy::needless_range_loop is allowed crate-wide.

Re-exports

pub use error::SeqError;

pub use error::SeqResult;

pub use handle::LcgRng;

pub use handle::SeqHandle;

pub use handle::SmVersion;

Modules

align

Wavefront-based sequence alignment.

alignment

Sequence-alignment algorithms.

beam

Generic beam search infrastructure.

crf

Linear-chain Conditional Random Fields (CRF).

ctc

Connectionist Temporal Classification (CTC).

decoders

Stochastic decoders for autoregressive sequence generation.

distance

String-distance algorithms.

error

Error types for oxicuda-seq.

folding

RNA secondary-structure folding algorithms.

grid_crf

2-D pairwise CRF + mean-field variational inference + loopy BP.

handle

Handle and RNG primitives for oxicuda-seq.

hmm

Hidden Markov Models: discrete + Gaussian emissions, forward-backward, Viterbi decoding, and Baum-Welch (EM) parameter learning. Also includes Variational Bayes EM (Dirichlet priors) and Hidden Semi-Markov Models with explicit duration distributions.

kalman

Linear / Extended Kalman filtering, RTS smoothing, EM parameter learning, Unscented Kalman Filter, and Particle Filter.

matching

Multi-pattern and exact string-matching automata.

memm

Maximum-Entropy Markov Models (MEMM): per-state softmax classifiers over the previous label plus features.

metrics

Evaluation metrics for sequence models.

mrf

General Markov Random Fields (graph + Ising), Gibbs sampler, loopy belief propagation.

perceptron

Structured-perceptron sequence taggers (Collins 2002), including the averaged-perceptron variant.

ptx_kernels

GPU PTX kernels for Sequence Models & Structured Prediction.

ssvm

Structured SVM with cutting-plane optimisation for linear-chain prediction.

string

Classical string algorithms.

structured

Differentiable structured-prediction layers.

tagging

Tag-scheme utilities for sequence labelling: BIO / BIOES conversion and validation, entity-span extraction, and span-based (entity-level) F1 (seqeval semantics).