def sample(self, i):
s = self.data[i]
ss = self.dd[i]
probs = []
for t in range(self.K):
prior = math.log(self.counts[t])
ll = self.mvNormalLL(s, ss, 1, self.mu[t], self.precision[t], self.logdet[t])
probs.append(prior + ll)
if self.K < self.con.pruningfactor:
prior = self.con.logalpha
ll = self.integrated[i]
probs.append(prior + ll)
normed = logNormalize(probs)
return sampleIndex(normed)
def sample(self, i):
probs = []
for t in range(self.K):
prior = math.log(self.counts[t])
ll = self.posteriorPredictive(set([i]), self.data[i], self.dd[i], t)
indices = self.getIndices(t)
suffprior = np.zeros(len(indices))
for j, ind in enumerate(indices):
suffprior[j] = self.prior[i, ind]
avgprior = np.log(suffprior.sum() / len(indices))
probs.append(prior + avgprior + ll)
if self.K < self.con.pruningfactor:
prior = self.con.logalpha
ll = self.integrated[i]
probs.append(prior + ll)
normed = logNormalize(probs)
return sampleIndex(normed)
def sample(self, i, oldt, oldf, followset, s, ss):
llcache = np.zeros(self.K)
n = self.n
if self.con.seq:
n = i + 1
probs = self.probs[:n]
probs.fill(0.)
newll = 0
for j in xrange(n):
if i == j:
prior = self.con.logalpha
else:
prior = -self.featsetvals[self.features[i,j]]
t = self.assignments[j]
if t == -1:
ll = newll
if ll == 0:
ll = newll = self.integrateOverParameters(len(followset), s, ss)
else:
ll = llcache[t]
if ll == 0:
if t == oldt and self.counts[t] == len(followset):
ll = self.integrateOverParameters(len(followset), s, ss)
else:
ll = self.mvNormalLL(s, ss, len(followset), self.mu[t], self.precision[t], self.logdet[t])
llcache[t] = ll
probs[j] = prior + ll
normed = np.exp(probs - logsumexp(probs))
#np.testing.assert_allclose(normed, normed2)
ind = sampleIndex(normed)
table = self.assignments[ind]
return table, ind
def sample(self, i, oldt, oldf, followset, s, ss):
if oldt == 26:
pass
# import pdb; pdb.set_trace()
llcache = np.zeros(self.K)
n = self.n
if self.con.seq:
n = i + 1
#probs = self.probs[:n]
#probs.fill(0.)
probs = []
#inds = []
newll = 0
priors = []
dist_priors = []
for j in xrange(n):
if i == j:
prior = self.con.logalpha
else:
prior = -self.featsetvals[self.features[i,j]]
priors.append(prior)
distprior = -self.mdist[i,j]
dist_priors.append(distprior)
#if prior < self.con.priorth:
# continue
#inds.append(j)
t = self.assignments[j]
if t == -1:
ll = newll
if ll == 0:
ll = newll = self.integrateOverParameters(len(followset), s, ss)
else:
ll = llcache[t]
if ll == 0:
ll = self.posteriorPredictive(len(followset), s, ss, t)
llcache[t] = ll
#probs[j] = prior + ll
if oldt == 26:
pass
prob = self.con.a * (prior + distprior) + ll
probs.append(prob)
#probs.append(prior + ll)
sorted_prior = sorted(zip(priors, range(n)), reverse=True)
sorted_dist_prior = sorted(zip(dist_priors, range(n)), reverse=True)
probs = np.array(probs)
normed = np.exp(probs - logsumexp(probs))
sorted_normed = sorted(zip(normed, range(n)), reverse=True)
if oldt == 26:
for score, ind_ in sorted_normed[:10]:
print self.vocab[ind_], self.assignments[ind_]
print
#np.testing.assert_allclose(normed, normed2)
ind = sampleIndex(normed)
if oldt == 26:
pass
#ind = inds[ind]
table = self.assignments[ind]
return table, ind