def E_step(self):
"""
Find the likelihoods for each datum for each component and
each translation
The likelihood is also computed
"""
for cur_component, vals in enumerate(zip(self.means,
self.covs,
self.norm_constants)):
mean, cov, norm_constant = vals
self.affinities[cur_component] = np.sum(- .5 *(self.trans_data_mat - mean)**2 * cov,2) + norm_constant
self.max_affinities[:] = np.max(
np.max(self.affinities,
0),
1)
subtract_max_affinities(self.affinities,self.max_affinities,
self.num_mix,
self.num_data,
self.trans_amount)
self.affinities = np.exp(self.affinities)
self.log_likelihood = np.sum(np.log(np.sum(
np.sum(
self.affinities * self.trans_weights,
2).T * self.mix_weights,1)) + self.max_affinities)
affinity_sums = np.sum(self.affinities.sum(0),1)
normalize_affinities(self.affinities,affinity_sums,
self.num_mix,
self.num_data,
self.trans_amount)
self.mix_weights = np.sum(np.sum(self.affinities,1),1)
self.mix_weights /= np.sum(self.mix_weights)
affinities[cur_component] = np.sum((trans_data_mat - mean)**2 * cov,2) + norm_constant
#
#
#
from _translation_invariant_Gauss_EM import subtract_max_affinities, normalize_affinities
affinities *=2
normalize_affinities(affinities,np.sum(affinities.sum(0),1),
num_mix, num_data, trans_amount)
subtract_max_affinities(affinities,np.sum(affinities.sum(0),1),
num_mix, num_data, trans_amount)
import translation_invariant_Gauss_EM
reload(translation_invariant_Gauss_EM)
TranslationInvariantGaussEM = translation_invariant_Gauss_EM.TranslationInvariantGaussEM
tig = TranslationInvariantGaussEM(1,data_mat,12)
tig.run_EM(.0001)
datamat = np.zeros((100,10))
component_length = 9
for i in xrange(100):
if i % 2 == 0:
datamat[i,:component_length] = 100 * np.ones(component_length)