def objective_function(self, reps, collapse_z=False):
assert len(reps) == self.nterms
fobj = 0.
m = np.zeros(self.nfea)
S = np.zeros((self.nfea, self.nfea))
for estimator, rep in zip(self.estimators, reps):
# E[log P(u|U)]
fobj += estimator.expected_log_prob(rep)
# H(q)
fobj += rep.entropy()
# sufficient statistics
m += np.dot(estimator.A.T, rep.expected_value())
S += misc.mult([estimator.A.T, rep.covariance(), estimator.A])
Lambda_N = np.linalg.inv(self.Sigma_N)
fobj += -0.5 * self.nfea * np.log(2 * np.pi) - 0.5 * misc.logdet(
self.Sigma_N)
diff = self.x - m
fobj += -0.5 * np.dot(np.dot(diff, Lambda_N), diff)
fobj += -0.5 * np.sum(S * Lambda_N)
return fobj
def objective_function(self, reps, collapse_z=False):
assert len(reps) == self.nterms
fobj = 0.
m = np.zeros(self.nfea)
S = np.zeros((self.nfea, self.nfea))
for estimator, rep in zip(self.estimators, reps):
# E[log P(u|U)]
fobj += estimator.expected_log_prob(rep)
# H(q)
fobj += rep.entropy()
# sufficient statistics
m += np.dot(estimator.A.T, rep.expected_value())
S += misc.mult([estimator.A.T, rep.covariance(), estimator.A])
Lambda_N = np.linalg.inv(self.Sigma_N)
fobj += -0.5 * self.nfea * np.log(2*np.pi) - 0.5 * misc.logdet(self.Sigma_N)
diff = self.x - m
fobj += -0.5 * np.dot(np.dot(diff, Lambda_N), diff)
fobj += -0.5 * np.sum(S * Lambda_N)
return fobj
def value(self):
return misc.mult([child.value() for child in self.children])