def EMiterate(self):
"""
Do one iteration
:return: logL
"""
# Get params for current iteration
data = self.s.getFeatures(self.label)
nbSamples, nbFeatures = data.shape
log_pXn_mat = np.zeros((nbSamples, self.K))
for i in range(self.K):
tmp = GMM.logN(data, self.mean[i], self.cov[i])
log_pXn_mat[:, i] = tmp + np.log(self.w[i])
pMax = np.max(log_pXn_mat, axis=1)
log_pXn = pMax + np.log(np.sum(np.exp(log_pXn_mat.T - pMax), axis=0).T)
logL = np.sum(log_pXn)
log_pNk = np.zeros((nbSamples, self.K))
for i in range(self.K):
log_pNk[:, i] = log_pXn_mat[:, i] - log_pXn
pNk = np.e**log_pNk
for i in range(self.K):
self.mean[i] = np.sum(pNk[:, i] * data.T, axis=1) / np.sum(pNk[:, i])
self.cov[i] = np.dot(pNk[:, i] * (data - self.mean[i]).T, data - self.mean[i]) / np.sum(pNk[:, i])
self.w[i] = np.sum(pNk[:, i]) / nbSamples
return logL
def pdf(self, X):
"""
Evaluate the pdf of data to belong to the mixture
:param data: (npArray) data to evaluate
:return: proba (npArray)
"""
proba = None
for i in range(self.K):
w = self.w[i]
probaToAdd = GMM.gaussianPdf(X, self.mean[i], self.cov[i]) * w
if proba is None:
proba = probaToAdd
else:
proba += probaToAdd
return proba
