from ._model_common import *
from ._diag_gmm import *
from ._full_gmm import *
from . import _full_gmm_ext
from ._full_gmm_normal import *
from ._mle_diag_gmm import *
from ._mle_full_gmm import *
from . import am
from .. import matrix as _matrix
from kaldi.matrix import _matrix_wrapper
from kaldi.matrix.packed import _sp_matrix_wrapper
[docs]class DiagGmm(_diag_gmm.DiagGmm):
"""Gaussian Mixture Model with diagonal covariances.
Args:
nmix (int): Number of Gaussians to mix
dim (int): Dimension of each component
"""
def __init__(self, nmix = 0, dim = 0):
"""Creates a new DiagGmm with specified number of gaussian mixtures
and dimensions.
Args:
nmix (int): number of gaussian to mix
dim (int): dimension
"""
super(DiagGmm, self).__init__()
if nmix < 0 or dim < 0:
raise ValueError("nmix and dimension must be a positive integer.")
if nmix > 0 and dim > 0:
self.resize(nmix, dim)
[docs] def copy(self, src):
"""Copies data from src into this DiagGmm and returns this DiagGmm.
Args:
src (FullGmm or DiagGmm): Source Gmm to copy
Returns:
This DiagGmm after update.
"""
if isinstance(src, FullGmm):
self.copy_from_full(src)
elif isinstance(src, DiagGmm):
self.copy_from_diag(src)
else:
raise ValueError("src must be either FullGmm or DiagGmm")
return self
[docs] def component_posteriors(self, data):
"""Computes the posterior probabilities of all Gaussian components given
a data point.
Args:
data (VectorBase): Data point with the same dimension as each
component.
Returns:
2-element tuple containing
- **loglike** (:class:`float`): Log-likelihood
- **posteriors** (:class:`~kaldi.matrix.Vector`): Vector with the
posterior probabilities
Raises:
ValueError if data is not consistent with gmm dimension.
"""
if data.dim != self.dim():
raise ValueError("data dimension {} does not match gmm dimension {}"
.format(data.dim, self.dim()))
posteriors = _matrix.Vector(self.num_gauss())
loglike = self._component_posteriors(data, posteriors)
return loglike, posteriors
[docs]class FullGmm(_full_gmm.FullGmm):
"""Python wrapper for Kaldi::FullGmm<Float>
Provides a more pythonic access to the C++ methods.
Args:
nmix (int): number of gaussian to mix
dim (int): dimension of each gaussian
Raises:
ValueError if nmix or dimension are not positive integers.
"""
def __init__(self, nmix = 0, dim = 0):
"""Creates a new FullGmm with specified number of gaussian mixtures and
dimensions.
Args:
nmix (int): number of gaussian to mix
dim (int): dimension
"""
super(FullGmm, self).__init__()
if nmix < 0 or dim < 0:
raise ValueError("nmix and dimension must be a positive integer.")
if nmix > 0 and dim > 0:
self.resize(nmix, dim)
[docs] def copy(self, src):
"""Copies data from src into this FullGmm and returns this FullGmm.
Args:
src (FullGmm or DiagGmm): Source Gmm to copy
Returns:
This FullGmm after update.
"""
if isinstance(src, FullGmm):
self.copy_from_full(src)
elif isinstance(src, DiagGmm):
_full_gmm_ext.copy_from_diag(self, src)
else:
raise ValueError("src must be either FullGmm or DiagGmm")
return self
[docs] def component_posteriors(self, data):
"""Computes the posterior probabilities of all Gaussian components given
a data point.
Args:
data (VectorBase): Data point with the same dimension as each
component.
Returns:
2-element tuple containing
- **loglike** (:class:`float`): Log-likelihood
- **posteriors** (:class:`~kaldi.matrix.Vector`): Vector with the
posterior probabilities
Raises:
ValueError if data is not consistent with gmm dimension.
"""
if data.dim != self.dim():
raise ValueError("data dimension {} does not match gmm dimension {}"
.format(data.dim, self.dim()))
posteriors = _matrix.Vector(self.num_gauss())
loglike = self._component_posteriors(data, posteriors)
return loglike, posteriors
[docs] def set_weights(self, weights):
"""Sets gmm mixture weights."""
if not isinstance(weights, _matrix.Vector):
weights = _matrix.Vector(weights)
self._set_weights(weights)
[docs] def set_means(self, means):
"""Sets gmm component means."""
if not isinstance(means, _matrix.Matrix):
means = _matrix.Matrix(means)
self._set_means(means)
[docs] def inv_covars(self):
"""
Returns:
Component inverse covariances
"""
return [_sp_matrix_wrapper(sp) for sp in self._inv_covars()]
[docs] def get_covars(self):
"""
Returns:
Component Covariances
"""
return [_sp_matrix_wrapper(sp) for sp in self._get_covars()]
[docs] def get_covars_and_means(self):
"""
Returns:
Component Covariances
"""
covars, means = self._get_covars_and_means()
covars = [_sp_matrix_wrapper(sp) for sp in covars]
means = _matrix_wrapper(means)
return covars, means
################################################################################
__all__ = [name for name in dir()
if name[0] != '_'
and not name.endswith('Base')]