def log_predictive_prob(self, new_corpus, num_samples):
D, V, T = self.D, self.V, self.T
Nvt_plus_beta_n = self.Nvt_plus_beta_n
Nt_plus_beta = self.Nt_plus_beta
Dt_plus_alpha_m = self.Dt_plus_alpha_m
D_plus_alpha = self.D_plus_alpha
Nvt_new, Nt_new, Dt_new, z_new = [], [], [], []
for r in xrange(num_samples):
Nvt_new.append(zeros((T, V), dtype=int))
Nt_new.append(zeros(T, dtype=int))
Dt_new.append(zeros(T, dtype=int))
z_new.append(zeros(len(new_corpus), dtype=int))
log_p = 0
for d, doc in enumerate(iterview(new_corpus)):
tmp = zeros(num_samples, dtype=float)
for r in xrange(num_samples):
for prev_d in xrange(0, d):
prev_doc = corpus.documents[prev_d]
t = z_new[r][prev_d]
Nvt_new[r][t, :] -= prev_doc.Nv
Nt_new[r][t] -= len(prev_doc)
Dt_new[r][t] -= 1
t = log_sample(gammaln(Nt_new[r] + Nt_plus_beta) - gammaln(Nvt_new[r] + Nvt_plus_beta_n).sum(axis=1) + gammaln(tile(prev_doc.Nv, (T, 1)) + Nvt_new[r] + Nvt_plus_beta_n).sum(axis=1) - gammaln(len(prev_doc) * ones(T) + Nt_new[r] + Nt_plus_beta) + log(Dt_new[r] + Dt_plus_alpha_m))
Nvt_new[r][t, :] += prev_doc.Nv
Nt_new[r][t] += len(prev_doc)
Dt_new[r][t] += 1
z_new[r][prev_d] = t
log_dist = gammaln(Nt_new[r] + Nt_plus_beta) - gammaln(Nvt_new[r] + Nvt_plus_beta_n).sum(axis=1) + gammaln(tile(doc.Nv, (T, 1)) + Nvt_new[r] + Nvt_plus_beta_n).sum(axis=1) - gammaln(len(doc) * ones(T) + Nt_new[r] + Nt_plus_beta) + log(Dt_new[r] + Dt_plus_alpha_m) - log(d + D_plus_alpha)
tmp[r] = log_sum_exp(log_dist)
t = log_sample(log_dist)
Nvt_new[r][t, :] += doc.Nv
Nt_new[r][t] += len(doc)
Dt_new[r][t] += 1
z_new[r][d] = t
log_p += log_sum_exp(tmp) - log(num_samples)
return log_p
def gibbs_iteration(self, init=False):
"""
Uses Gibbs sampling to draw a single sample from the posterior
distribution over document--component assignments (i.e.,
document groups) given this instance's corpus (i.e., document
tokens). By default (i.e., if keyword argument 'init' is set
to the value 'False') all document--component assignments (and
corresponding counts) are assumed to have been initialized
previously; otherwise, they are initialized.
Keyword arguments:
init -- whether to initialize document--component assignments
"""
corpus = self.corpus
T = self.T
alpha_m = self.alpha_m
Nvt_plus_beta_n = self.Nvt_plus_beta_n
Nt_plus_beta = self.Nt_plus_beta
Dt = self.Dt
z = self.z
for d, (doc, t) in enumerate(iterview(zip(corpus, z))):
if not init:
Nvt_plus_beta_n[t, :] -= doc.Nv
Nt_plus_beta[t] -= len(doc)
Dt[t] -= 1
t = log_sample(
gammaln(Nt_plus_beta)
- gammaln(Nvt_plus_beta_n).sum(axis=1)
+ gammaln(tile(doc.Nv, (T, 1)) + Nvt_plus_beta_n).sum(axis=1)
- gammaln(len(doc) * ones(T) + Nt_plus_beta)
+ log(Dt + alpha_m)
)
Nvt_plus_beta_n[t, :] += doc.Nv
Nt_plus_beta[t] += len(doc)
Dt[t] += 1
z[d] = t
def gibbs_iteration(self, init=False):
"""
Uses Gibbs sampling to draw a single sample from the posterior
distribution over document--component assignments (i.e.,
document groups) given this instance's corpus (i.e., document
tokens). By default (i.e., if keyword argument 'init' is set
to the value 'False') all document--component assignments (and
corresponding counts) are assumed to have been initialized
previously; otherwise, they are initialized.
Keyword arguments:
init -- whether to initialize document--component assignments
"""
corpus = self.corpus
T = self.T
Nvt_plus_beta_n = self.Nvt_plus_beta_n
Nt_plus_beta = self.Nt_plus_beta
Dt_plus_alpha_m = self.Dt_plus_alpha_m
z = self.z
for d, (doc, t) in enumerate(iterview(zip(corpus, z))):
if not init:
Nvt_plus_beta_n[t, :] -= doc.Nv
Nt_plus_beta[t] -= len(doc)
Dt_plus_alpha_m[t] -= 1
t = log_sample(
gammaln(Nt_plus_beta) - gammaln(Nvt_plus_beta_n).sum(axis=1) +
gammaln(tile(doc.Nv, (T, 1)) + Nvt_plus_beta_n).sum(axis=1) -
gammaln(len(doc) * ones(T) + Nt_plus_beta) +
log(Dt_plus_alpha_m))
Nvt_plus_beta_n[t, :] += doc.Nv
Nt_plus_beta[t] += len(doc)
Dt_plus_alpha_m[t] += 1
z[d] = t
def log_predictive_prob(self, new_corpus, num_samples):
D, V, T = self.D, self.V, self.T
Nvt_plus_beta_n = self.Nvt_plus_beta_n
Nt_plus_beta = self.Nt_plus_beta
Dt_plus_alpha_m = self.Dt_plus_alpha_m
D_plus_alpha = self.D_plus_alpha
Nvt_new, Nt_new, Dt_new, z_new = [], [], [], []
for r in xrange(num_samples):
Nvt_new.append(zeros((T, V), dtype=int))
Nt_new.append(zeros(T, dtype=int))
Dt_new.append(zeros(T, dtype=int))
z_new.append(zeros(len(new_corpus), dtype=int))
log_p = 0
for d, doc in enumerate(iterview(new_corpus)):
tmp = zeros(num_samples, dtype=float)
for r in xrange(num_samples):
for prev_d in xrange(0, d):
prev_doc = corpus.documents[prev_d]
t = z_new[r][prev_d]
Nvt_new[r][t, :] -= prev_doc.Nv
Nt_new[r][t] -= len(prev_doc)
Dt_new[r][t] -= 1
t = log_sample(
gammaln(Nt_new[r] + Nt_plus_beta) -
gammaln(Nvt_new[r] + Nvt_plus_beta_n).sum(axis=1) +
gammaln(
tile(prev_doc.Nv, (T, 1)) + Nvt_new[r] +
Nvt_plus_beta_n).sum(axis=1) - gammaln(
len(prev_doc) * ones(T) + Nt_new[r] +
Nt_plus_beta) +
log(Dt_new[r] + Dt_plus_alpha_m))
Nvt_new[r][t, :] += prev_doc.Nv
Nt_new[r][t] += len(prev_doc)
Dt_new[r][t] += 1
z_new[r][prev_d] = t
pass # YOUR CODE GOES HERE
Nvt_new[r][t, :] += doc.Nv
Nt_new[r][t] += len(doc)
Dt_new[r][t] += 1
z_new[r][d] = t
log_p += log_sum_exp(tmp) - log(num_samples)
return log_p