def initialise_parameters(self, data, num_comp):
"""Initialize using D2-weighting
"""
centroids_idx = init_d2_weighting(data, num_comp)
init_mean = np.array([data[k, :] for k in centroids_idx])
init_covar = np.cov(data, rowvar=0)
init_covar = np.repeat(np.array([init_covar]), num_comp, 0)
labels = classify_by_distance(data,
init_mean,
init_covar)
labels = labels.flatten()
init_covar = self._get_covar(data, labels)
init_mixweights = element_weights(labels)
init_params = {}
init_params['mean'] = init_mean
init_params['covar'] = init_covar
init_params['mixweights'] = init_mixweights
return init_params
def initialise_parameters(self, data, num_comp):
'''
Initialize randomly
'''
D = data.shape[1]
data_lims = np.array([[data[:, d].min(), data[:, d].max()]
for d in range(D)])
init_mean = np.array([uniform(*data_lims[d, :], size=num_comp)
for d in range(D)]).T
covar_init = np.repeat([np.diag([1] * D)], num_comp, 0)
labels = classify_by_distance(data,
init_mean,
covar_init)
labels = labels.flatten()
init_covar = self._get_covar(data, labels)
init_mixweights = element_weights(labels)
init_params = {}
init_params['mean'] = init_mean
init_params['covar'] = init_covar
init_params['mixweights'] = init_mixweights
return init_params
def _init_d2_weighting(self, num_comp):
"""Initialize using D2-weighting
"""
centroids_idx = init_d2_weighting(self.data, num_comp)
init_mean = np.array([self.data[k, :] for k in centroids_idx])
init_covar = np.cov(self.data, rowvar=0)
init_covar = np.repeat(np.array([init_covar]), num_comp, 0)
labels = classify_by_distance(self.data, init_mean,
init_covar).flatten()
init_covar = self._get_covar(self.data, labels)
init_mixweights = element_weights(labels)
return (init_mean, labels, init_covar, init_mixweights)
def _init_random(self, num_comp):
"""Initialize randomly
"""
D = self.data.shape[1]
data_lims = np.array([[self.data[:, d].min(), self.data[:, d].max()]
for d in range(D)])
init_mean = np.array([uniform(*data_lims[d, :], size=num_comp)
for d in range(D)]).T
covar_init = np.repeat([np.diag([1] * D)], num_comp, 0)
labels = classify_by_distance(self.data, init_mean,
covar_init).flatten()
init_covar = self._get_covar(self.data, labels)
init_mixweights = element_weights(labels)
return (init_mean, labels, init_covar, init_mixweights)
def initialise_parameters(self, data, init_mean, init_covar, init_mixweights):
'''
If starting solution supplied initialise from it.
'''
if init_mixweights is None:
labels = classify_by_distance(data,
init_mean,
init_covar)
init_mixweights = element_weights(labels)
if init_covar is None:
init_covar = self._get_covar(data, labels)
init_params = {}
init_params['mean'] = init_mean
init_params['covar'] = init_covar
init_params['mixweights'] = init_mixweights
return init_params
def __init__(self, data,
num_comp_init=10,
max_iter=200,
thresh=1e-5,
verbose=False,
init_mean=None,
init_covar=None,
init_mixweights=None,
init_method='d2-weighting',
prior_dirichlet=1e-3,
dof_init=2,
remove_comp_thresh=1e-2,
whiten_data=False,
plot_monitor=False,
use_approx=True):
"""Fit the model to the data using Variational Bayes
"""
(num_obs, num_features) = np.shape(data)
if whiten_data:
data = whiten(data)
self.data = data
self.remove_comp_thresh = remove_comp_thresh
# Choose method to intialize the parameters
if init_method == 'd2-weighting':
init_method = self._init_d2_weighting
elif init_method == 'kmeans':
init_method = self._init_kmeans
elif init_method == 'random':
init_method = self._init_random
if init_mean is None:
# No starting solution was supplied
# Initialize with `init_method`
(init_mean, labels, init_covar, init_mixweights) = \
init_method(num_comp_init)
else:
# Starting solution was supplied
num_comp_init = init_mean.shape[0]
if init_mixweights is None:
labels = classify_by_distance(data, init_mean,
init_covar)
init_mixweights = element_weights(labels)
if init_covar is None:
init_covar = self._get_covar(data, labels)
# Initialize data structures
Prior = _Prior(data, num_comp_init, prior_dirichlet)
ESS = _ESS(data, num_comp_init, init_mean, init_covar,
init_mixweights)
LatentVariables = _LatentVariables(data, ESS, num_comp_init)
Posterior = _Posterior(Prior, num_comp_init, dof_init,
use_approx=use_approx)
LowerBound = _LowerBound(data, num_obs, num_features,
num_comp_init, Prior)
# Initial M-step
Posterior.update_parameters(Prior, ESS, LatentVariables)
# Main loop
iteration = 1
done = False
if plot_monitor:
self._plot_monitor_init()
while not done:
# Update parameters
self._update_step(Prior, Posterior, ESS,
LatentVariables, LowerBound)
# Converged?
if iteration == 1:
converged = False
else:
converged = self._convergence_test(LowerBound, thresh)
done = converged or (iteration >= max_iter)
if plot_monitor:
self._plot_monitor_update(ESS)
if verbose:
print('iteration %d, lower bound: %f' %
(iteration, LowerBound.lower_bound[-1]))
print Posterior.nws_mean
iteration += 1
self.Posterior = Posterior
self.Prior = Prior
self.LatentVariables = LatentVariables
self.ESS = ESS
self.LowerBound = LowerBound
# Call main loop of wxFrame to keep the window from closing
if plot_monitor:
self.frame.end_of_iteration()
self.app.MainLoop()