Expand description
hmmlearn — a native Rust port of the Python library hmmlearn.
Hidden Markov Models with categorical, Gaussian, Gaussian-mixture, Poisson, and multinomial emissions, trained by Expectation–Maximization or variational inference.
Fitting proceeds by Expectation–Maximization (maximum-likelihood) or
mean-field variational inference. The concrete models and their builders live
in models, backed by the forward–backward and Viterbi routines in
core, the covariance representations in covariance, and the
supporting numerics (special functions, linear algebra, statistics, and array
utilities) in the remaining modules.
§Quick start
Every model is built with a fluent builder, fit with
fit, and returns a Fitted
handle exposing predict, score, sample, decode, aic/bic, and the
learned parameters.
use hmmlearn::models::GaussianHmm;
use hmmlearn::ndarray::array;
// One feature per column; each row is an observation.
let x = array![[0.0], [0.2], [-0.1], [5.0], [5.2], [4.8]];
let model = GaussianHmm::new(2)
.n_iter(20)
.random_state(42)
.fit(x.view(), None)?;
let states = model.predict(x.view(), None)?;
assert_eq!(states.len(), x.nrows());
let log_likelihood = model.score(x.view(), None)?;
assert!(log_likelihood.is_finite());Multiple independent sequences are concatenated along the rows and delimited
by a lengths slice, exactly as in hmmlearn:
model.fit(x.view(), Some(&[100, 200, 150])).
§Relationship to hmmlearn
This is a faithful, pure-Rust port of
hmmlearn 0.3.3 with no LAPACK/BLAS
dependency. It deviates from upstream in three documented, test-validated
ways, all confined to initialization and sampling: k-means++ replaces
scikit-learn’s KMeans for mean initialization; multivariate-normal sampling
uses a Cholesky factor rather than NumPy’s SVD; and the binomial sampler omits
NumPy’s BTPE branch in favor of the inversion fallback. Fitted parameters and
inference results match upstream.
Re-exports§
Modules§
- cluster
- A small k-means++ used to initialize Gaussian means.
- core
- Inference engine: shared parameters, algorithms, and the generic HMM core.
- covariance
- Covariance parameterization: the type tag, compressed storage, and validation.
- error
- Error type shared across the crate.
- kl
- Kullback–Leibler divergences used by the variational models.
- linalg
- Small dense linear-algebra helpers backed by
nalgebra(pure Rust, no LAPACK). - models
- Concrete HMM models and their builders.
- rng
- Bit-exact reproduction of NumPy’s legacy
RandomState. - special
- Special functions: log-gamma, digamma, and numerically stable log-sum-exp.
- stats
- Probability densities and estimators used by the emission models.
- util
- Array normalization helpers and sequence splitting.