def testSanity(self):
"""
Sanity check online Gibbs init scheme against deltaLDA
"""
# Don't even try unless deltaLDA module present
if(not hasDelta):
return
randseed = 194582
# Use 1 'online' Gibbs sample to build initial gamma
gibbs_docs = [[1,1,2],
[1,1,1,1,2],
[3,3,3,4],
[3,3,3,3,4,4],
[0,0,0,0,0],
[0,0,0,0]]
numsamp = 0
(phi,theta,sample) = deltaLDA(gibbs_docs,self.alpha,self.beta,
numsamp,randseed)
gamma_init = []
for (d,di) in zip(self.docs_w,range(len(self.docs_w))):
gamma = zeros((self.T,len(d)))
for (w,i) in zip(d,range(len(d))):
gamma[:,i] = theta[di,:] * phi[:,w]
# normalize
gamma[:,i] = gamma[:,i] / gamma[:,i].sum()
# save
gamma_init.append(gamma)
# Run cvbLDA with this gamma
(gphi,gtheta,gamma) = cvbLDA(self.docs_w,self.docs_c,
self.alpha,self.beta,
gamma_init=gamma_init,
maxiter=self.maxiter,
convtol=self.convtol)
# Run cvbLDA no init gamma, same randseed
(phi,theta,gamma) = cvbLDA(self.docs_w,self.docs_c,
self.alpha,self.beta,
randseed=randseed,
maxiter=self.maxiter,
convtol=self.convtol)
self.assert_(self.matAgree(phi,gphi))
self.assert_(self.matProb(phi))
self.assert_(self.matProb(gphi))
self.assert_(self.matAgree(theta,gtheta))
self.assert_(self.matProb(theta))
self.assert_(self.matProb(gtheta))
def testStandard(self):
""" Test standard LDA with base data/params """
(phi,theta,gamma) = cvbLDA(self.docs_w,self.docs_c,
self.alpha,self.beta,
maxiter=self.maxiter,
convtol=self.convtol)
# theta should clust docs [0,1], [2,3], [4,5]
maxtheta = argmax(theta,axis=1)
self.assert_(maxtheta[0] == maxtheta[1])
self.assert_(maxtheta[2] == maxtheta[3])
self.assert_(maxtheta[4] == maxtheta[5])
# theta valid prob matrix
self.assert_(self.matProb(theta))
# corresponding phi should emph [1,2], [3,4], [0]
maxphi = argmax(phi,axis=1)
self.assert_(maxphi[maxtheta[0]] == 1)
self.assert_(maxphi[maxtheta[2]] == 3)
self.assert_(maxphi[maxtheta[4]] == 0)
# phi valid prob matrix
self.assert_(self.matProb(phi))
beta = .1 * np.ones((nb_topics, vocab_size))
nb_iter_max = 100
tol = .001
#If the user has specified some parameters, we fix them
else:
alpha = alpha * np.ones((1, nb_topics))
beta = beta * np.ones((nb_topics, vocab_size))
start_time = time.time()
(phi, theta, gamma) = cvbLDA(words,
counts,
alpha,
beta,
maxiter=nb_iter_max,
verbose=0,
convtol=tol)
print('\nCollapsed variational inference LDA exec time: ' +
str(time.time() - start_time) + 's')
#print('Theta, p(z|d)')
#print(str(theta))
#print('Phi, p(w|z)')
#print(str(phi))
topic = []
# Stopping conditions for inference:
# -stop after maxiter iterations
# -stop once sum of absolute changes in all gamma variational
# parameters for a single iteration falls below convtol
# (whichever occurs FIRST)
#
# If these parameters are not supplied, stop after 100 iterations
#
(maxiter, convtol) = (10, .01)
# Do CVB inference for LDA
#
(phi, theta, gamma) = cvbLDA(docs_w,
docs_c,
alpha,
beta,
maxiter=maxiter,
verbose=1,
convtol=convtol)
# theta is the matrix of document-topic probabilities
# (estimated from expected counts under variational posterior)
#
# theta = D x T
# theta[di,zj] = P(z=zj | d=di)
#
print ''
print 'Theta - P(z|d)'
print str(theta)
print ''
[5],
[4]]
# Stopping conditions for inference:
# -stop after maxiter iterations
# -stop once sum of absolute changes in all gamma variational
# parameters for a single iteration falls below convtol
# (whichever occurs FIRST)
#
# If these parameters are not supplied, stop after 100 iterations
#
(maxiter,convtol) = (10,.01)
# Do CVB inference for LDA
#
(phi,theta,gamma) = cvbLDA(docs_w,docs_c,alpha,beta,
maxiter=maxiter,verbose=1,convtol=convtol)
# theta is the matrix of document-topic probabilities
# (estimated from expected counts under variational posterior)
#
# theta = D x T
# theta[di,zj] = P(z=zj | d=di)
#
print ''
print 'Theta - P(z|d)'
print str(theta)
print ''
# phi is the matrix of topic-word probabilities
# (estimated from expected counts under variational posterior)
#