ngclearn.density package

Submodules

ngclearn.density.gmm module

class ngclearn.density.gmm.GMM(k, max_iter=5, assume_diag_cov=False, init_kmeans=True)

Bases: object

Implements a custom/pure-TF Gaussian mixture model (GMM) – or mixture of Gaussians, MoG. Adaptation of parameters is conducted via the Expectation-Maximization (EM) learning algorithm and leverages full covariance matrices in the component multivariate Gaussians.

Note this is a TF wrapper model that houses the sklearn implementation for learning. The sampling process has been rewritten to utilize GPU matrix computation.

Parameters

• k – the number of components/latent variables within this GMM

• max_iter – the maximum number of EM iterations to fit parameters to data (Default = 5)

• assume_diag_cov – if True, assumes a diagonal covariance for each component (Default = False)

• init_kmeans – if True, first learn use the K-Means algorithm to initialize the component Gaussians of this GMM (Default = True)

calc_gaussian_logpdf(X)

Calculates log densities/probabilities of data X under each component given this GMM

Parameters

X – the dataset to calculate the log likelihoods from

calc_prob(X)

Computes probabilities p(z|x) of data samples in X under this GMM

Parameters

X – the dataset to estimate the probabilities from

calc_w_log_prob(X)

Calculates weighted log probabilities of data X under each component given this GMM

Parameters

X – the dataset to calculate the weighted log probabilities from

e_step(X)

estimate_gaussian_parameters(X, resp)

estimate_log_prob(X)

fit(data)

init_from_ScikitLearn(gmm)

Creates a GMM from a pre-trained Scikit-Learn model – conversion sets things up for a row-major form of sampling, i.e., s ~ mu_k + eps * (L_k^T) where k is the sampled component index

Parameters

gmm – the pre-trained GMM (from scikit-learn) to load in

m_step(X, log_resp)

predict(X)

Chooses which component samples in X are likely to belong to given p(z|x)

Parameters

X – the input data to compute p(z|x) from

sample(n_s, mode_i=- 1, samples_modes_evenly=False)

(Efficiently) Draw samples from the current underlying GMM model

Parameters

• n_s – the number of samples to draw from this GMM

• mode_i – if >= 0, will only draw samples from a specific component of this GMM (Default = -1), ignoring the Categorical prior over latent variables/components

• samples_modes_evenly – if True, will ignore the Categorical prior over latent variables/components and draw an approximately equal number of samples from each component

update(X)

Performs a single iterative update of parameters (assuming model initialized)

Parameters

X – the dataset to fit this GMM to

Module contents