def calcELBO_LinearTerms(SS=None,
StartStateCount=None,
TransStateCount=None,
rho=None,
omega=None,
Ebeta=None,
startTheta=None,
transTheta=None,
startAlpha=0,
alpha=0,
kappa=None,
gamma=None,
afterGlobalStep=0,
todict=0,
**kwargs):
""" Calculate ELBO objective terms that are linear in suff stats.
Returns
-------
L : scalar float
L is sum of any term in ELBO that is const/linear wrt suff stats.
"""
Ltop = L_top(rho=rho,
omega=omega,
alpha=alpha,
gamma=gamma,
kappa=kappa,
startAlpha=startAlpha)
LdiffcDir = -c_Dir(transTheta) - c_Dir(startTheta)
if afterGlobalStep:
if todict:
return dict(Lalloc=Ltop + LdiffcDir, Lslack=0)
return Ltop + LdiffcDir
K = rho.size
if Ebeta is None:
Ebeta = rho2beta(rho, returnSize='K+1')
if SS is not None:
StartStateCount = SS.StartStateCount
TransStateCount = SS.TransStateCount
# Augment suff stats to be sure have 0 in final column,
# which represents inactive states.
if StartStateCount.size == K:
StartStateCount = np.hstack([StartStateCount, 0])
if TransStateCount.shape[-1] == K:
TransStateCount = np.hstack([TransStateCount, np.zeros((K, 1))])
LstartSlack = np.inner(StartStateCount + startAlpha * Ebeta - startTheta,
digamma(startTheta) - digamma(startTheta.sum()))
alphaEbetaPlusKappa = alpha * np.tile(Ebeta, (K, 1))
alphaEbetaPlusKappa[:, :K] += kappa * np.eye(K)
digammaSum = digamma(np.sum(transTheta, axis=1))
LtransSlack = np.sum((TransStateCount + alphaEbetaPlusKappa - transTheta) *
(digamma(transTheta) - digammaSum[:, np.newaxis]))
if todict:
return dict(Lalloc=Ltop + LdiffcDir, Lslack=LstartSlack + LtransSlack)
return Ltop + LdiffcDir + LstartSlack + LtransSlack
def L_alloc_no_slack(self):
''' Compute allocation term of objective function, without slack term
Returns
-------
L : scalar float
'''
prior_cDir = L_top(nDoc=self.theta.shape[0],
alpha=self.alpha, gamma=self.gamma,
rho=self.rho, omega=self.omega)
post_cDir = c_Dir(self.theta, self.thetaRem)
return prior_cDir - post_cDir