mojolearn.GaussianMixture#

class mojolearn.GaussianMixture(n_components=1, covariance_type='full', tol=0.001, reg_covar=1e-06, max_iter=100, init_params='kmeans', random_state=0, n_init=1, warm_start=False, means_init=None, weights_init=None, precisions_init=None)[source]#

Bases: NumericModeMixin

Gaussian mixture model fitted by expectation-maximization.

Follows scikit-learn’s GaussianMixture: every parameter has the same name, meaning and default. Iterates until the change in the lower bound is below tol or max_iter is reached. X is read as float32 and the fitted arrays are float32.

The default configuration (covariance_type='full', init_params='kmeans' or 'random', n_init=1, no warm start and no initial arrays) runs on mojolearn’s main mixture solver. Any other option runs on a second solver; a model fitted that way can score and predict, but save and sample raise NotImplementedError.

Parameters:

  • n_components (int, default 1) – Number of mixture components.

  • covariance_type ({‘full’, ‘tied’, ‘diag’, ‘spherical’}, default ‘full’) – Form of each component’s covariance, as in scikit-learn.

  • tol (float, default 1e-3) – Convergence threshold on the change in the lower bound.

  • reg_covar (float, default 1e-6) – Added to the diagonal of each covariance.

  • max_iter (int, default 100) – Maximum number of EM iterations.

  • init_params ({‘kmeans’, ‘random’, ‘k-means++’, ‘random_from_data’}, default ‘kmeans’) – How the responsibilities are initialized.

  • random_state (int, default 0) – Seed for initialization and for sample.

  • n_init (int, default 1) – Number of initializations; the best one is kept.

  • warm_start (bool, default False) – Start from the previous fit’s solution when the shape and covariance type match.

  • means_init (array-like of shape (n_components, n_features), default None) – Initial means.

  • weights_init (array-like of shape (n_components,), default None) – Initial weights.

  • precisions_init (array-like, default None) – Initial precisions, shaped by covariance_type as in scikit-learn.

Variables:

  • weights_ (Array of shape (n_components,), float32) – Weight of each component.

  • means_ (Array of shape (n_components, n_features), float32) – Mean of each component.

  • covariances_ (Array, float32) – Covariance of each component, shaped by covariance_type; for ‘full’, (n_components, n_features, n_features). Includes reg_covar on the diagonal, as scikit-learn’s does.

  • precisions_cholesky_ (Array, float32) – Cholesky factors of the precision matrices, shaped like covariances_.

  • log_det_chol_ (Array of shape (n_components,), float32) – Log-determinant of each precision Cholesky factor.

  • n_iter_ (int) – Number of EM iterations run.

  • converged_ (bool) – False when the fit stopped at max_iter before reaching tol. scikit-learn also warns in that case; mojolearn only sets this flag.

  • lower_bound_ (float) – Lower bound on the log likelihood at the end of the fit.

  • n_features_in_ (int) – Number of features seen during fit.

Examples

>>> import numpy as np, mojolearn
>>> X = np.random.default_rng(0).random((1000, 8), dtype=np.float32)
>>> gmm = mojolearn.GaussianMixture(n_components=3).fit(X)
>>> labels = gmm.predict(X)
>>> proba = gmm.predict_proba(X)
>>> X_new, y_new = gmm.sample(100)
__init__(n_components=1, covariance_type='full', tol=0.001, reg_covar=1e-06, max_iter=100, init_params='kmeans', random_state=0, n_init=1, warm_start=False, means_init=None, weights_init=None, precisions_init=None)[source]#

Methods

aic(X)

Return the Akaike information criterion of the model on X.

bic(X)

Return the Bayesian information criterion of the model on X.

fit(X[, y])

Fit the mixture to X by expectation-maximization.

fit_predict(X[, y])

Fit the model, then return predict(X).

get_params([deep])

load(path)

Load a model written by save.

numeric_mode_used()

The numeric mode this estimator runs in.

predict(X)

Return the most likely component for each row of X.

predict_proba(X)

Return the posterior probability of each component for each row.

sample([n_samples])

Draw random rows from the fitted mixture.

save(path)

Save the fitted model to path as an .npz file.

score(X[, y])

Return the mean log likelihood of the rows of X.

score_samples(X)

Return the log likelihood of each row of X under the model.

set_params(**params)

Set constructor parameters.

vendor_used()

The GPU backend this estimator runs on.
fit(X, y=None)[source]#

Fit the mixture to X by expectation-maximization.

Runs until abs(lower_bound - previous) < tol or max_iter iterations.

Parameters:

  • X (array-like of shape (n_samples, n_features)) – Training data, read as float32.

  • y (ignored) – Present for scikit-learn API compatibility.

Returns:

  • self (GaussianMixture) – The fitted estimator.

  • An unknown covariance_type or init_params, or a component that

  • collapses during the fit, raises an error naming the problem; the fit

  • never returns a model with a reset component.

Raises:

  • ValueError – For a non-positive n_components or initial arrays of the wrong shape.

  • TypeError – If n_components is not an int or tol/reg_covar is not a number.

score_samples(X)[source]#

Return the log likelihood of each row of X under the model.

Parameters:

X (array-like of shape (n_samples, n_features))

Returns:

log_likelihood (Array of shape (n_samples,), float32)

predict_proba(X)[source]#

Return the posterior probability of each component for each row.

Parameters:

X (array-like of shape (n_samples, n_features))

Returns:

proba (Array of shape (n_samples, n_components), float32)

predict(X)[source]#

Return the most likely component for each row of X.

A tie goes to the lowest component index.

Parameters:

X (array-like of shape (n_samples, n_features))

Returns:

labels (Array of shape (n_samples,), int32)

fit_predict(X, y=None)[source]#

Fit the model, then return predict(X).

Returns:

labels (Array of shape (n_samples,), int32)

save(path)[source]#

Save the fitted model to path as an .npz file.

The file holds every fitted array plus n_iter_, converged_ and lower_bound_. Load it with load, or with mojolearn.host_model(path) to score on a machine with no GPU.

Parameters:

path (str or path-like) – Destination file.

Raises:

  • RuntimeError – If the model is not fitted.

  • NotImplementedError – If the model was fitted with options outside the default configuration (see the class description).

classmethod load(path)[source]#

Load a model written by save.

Arrays keep their saved dtypes. The result supports score_samples, predict_proba, predict, score, bic and aic.

Parameters:

path (str or path-like) – File written by GaussianMixture.save.

Returns:

model (GaussianMixture)

Raises:

ValueError – If the file is of another format or estimator, or is truncated or inconsistent.

score(X, y=None)[source]#

Return the mean log likelihood of the rows of X.

Parameters:

  • X (array-like of shape (n_samples, n_features))

  • y (ignored) – Present for scikit-learn API compatibility.

Returns:

score (float)

bic(X)[source]#

Return the Bayesian information criterion of the model on X.

Lower is better.

Parameters:

X (array-like of shape (n_samples, n_features))

Returns:

bic (float)

aic(X)[source]#

Return the Akaike information criterion of the model on X.

Lower is better.

Parameters:

X (array-like of shape (n_samples, n_features))

Returns:

aic (float)

sample(n_samples=1)[source]#

Draw random rows from the fitted mixture.

As in scikit-learn, the number of rows per component is a multinomial draw over weights_, and the rows come out grouped by component in ascending order. The draws are keyed by random_state, so the same model and random_state give the same rows on every call and every GPU vendor; they differ from scikit-learn’s draws. X is float32 and y int32, where scikit-learn returns float64 and int64.

Parameters:

n_samples (int, default 1) – Number of rows to draw. Must be at least 1.

Returns:

  • X (Array of shape (n_samples, n_features), float32) – The drawn rows.

  • y (Array of shape (n_samples,), int32) – The component each row was drawn from.

Raises:

  • ValueError – If the model is not fitted or random_state is not an int in [0, 2**64).

  • TypeError – If n_samples is not an int.

  • NotImplementedError – If the model was fitted with options outside the default configuration.