def __word_proposal(self, doc, token, old_topic):
if isinstance(doc, LDADoc) and isinstance(token, Token):
new_topic = self.__propose(token.id)
if new_topic != old_topic:
proposal_old = self.__word_proposal_distribution(token.id, old_topic)
proposal_new = self.__word_proposal_distribution(token.id, new_topic)
proportion_old = self.__proportional_function(doc, token, old_topic)
proportion_new = self.__proportional_function(doc, token, new_topic)
transition_prob = float((proportion_new * proposal_old) / (proportion_old * proposal_new))
rejection = rand()
mask = -(rejection < transition_prob)
return (new_topic & mask) | (old_topic & ~mask)
return new_topic
elif isinstance(doc, SLDADoc) and isinstance(token, Sentence):
sent = token
new_topic = old_topic
for word_id in sent.tokens:
new_topic = self.__propose(word_id)
if new_topic != old_topic:
proportion_old = self.__proportional_function(doc, sent, old_topic)
proportion_new = self.__proportional_function(doc, sent, new_topic)
proposal_old = self.__word_proposal_distribution(word_id, old_topic)
proposal_new = self.__word_proposal_distribution(word_id, new_topic)
transition_prob = float((proportion_new * proposal_old) / (proportion_old * proposal_new))
rejection = rand()
mask = -(rejection < transition_prob)
new_topic = (new_topic & mask) | (old_topic & ~mask)
return new_topic
def __doc_proposal(self, doc, token):
if isinstance(doc, LDADoc) and isinstance(token, Token):
old_topic = token.topic
dart = rand() * (doc.size() + self.__model.alpha_sum())
if dart < doc.size():
token_index = int(dart)
new_topic = doc.token(token_index).topic
else:
new_topic = rand_k(self.__model.num_topics())
if new_topic != old_topic:
proposal_old = self.__doc_proposal_distribution(doc, old_topic)
proposal_new = self.__doc_proposal_distribution(doc, new_topic)
proportion_old = self.__proportional_function(
doc, token, old_topic)
proportion_new = self.__proportional_function(
doc, token, new_topic)
transition_prob = float((proportion_new * proposal_old) /
(proportion_old * proposal_new))
rejection = rand()
mask = -(rejection < transition_prob)
return (new_topic & mask) | (old_topic & ~mask)
return new_topic
elif isinstance(doc, SLDADoc) and isinstance(token, Sentence):
sent = token
old_topic = sent.topic
dart = rand() * (doc.size() + self.__model.alpha_sum())
if dart < doc.size():
token_index = int(dart)
new_topic = doc.sent(token_index).topic
else:
new_topic = rand_k(self.__model.num_topics())
if new_topic != old_topic:
proportion_old = self.__proportional_function(
doc, sent, old_topic)
proportion_new = self.__proportional_function(
doc, sent, new_topic)
proposal_old = self.__doc_proposal_distribution(doc, old_topic)
proposal_new = self.__doc_proposal_distribution(doc, new_topic)
transition_prob = float((proportion_new * proposal_old) /
(proportion_old * proposal_new))
rejection = rand()
mask = -(rejection < transition_prob)
return (new_topic & mask) | (old_topic & ~mask)
return new_topic
def __sample_sentence(self, doc, sent):
old_topic = sent.topic
num_topics = self.__model.num_topics()
accum_prob = np.zeros(num_topics)
prob = np.zeros(num_topics)
sum_ = 0
for t in range(num_topics):
dt_alpha = doc.topic_sum(t) + self.__model.alpha()
t_sum_beta_sum = self.__model.topic_sum(t) + self.__model.beta_sum()
if t == old_topic:
if dt_alpha > 1:
dt_alpha -= 1
if t_sum_beta_sum > 1:
t_sum_beta_sum -= 1
prob[t] = dt_alpha
for i in range(len(sent.tokens)):
w = sent.tokens[i]
wt_beta = self.__model.word_topic_value(w, t) + self.__model.beta()
if t == old_topic and wt_beta > 1:
wt_beta -= 1
# Note: if the length of the sentence is too long, the probability will be
# too small and the accuracy will be lost if there are too many multiply items
prob[t] *= wt_beta / t_sum_beta_sum
sum_ += prob[t]
accum_prob[t] = prob[t] if t == 0 else accum_prob[t - 1] + prob[t]
dart = rand() * sum
if dart <= accum_prob[0]:
return 0
for t in range(1, num_topics):
if accum_prob[t - 1] < dart <= accum_prob[t]:
return t
return num_topics - 1
def __sample_token(self, doc, token):
old_topic = token.topic
num_topics = self.__model.num_topics()
accum_prob = np.zeros(num_topics)
prob = np.zeros(num_topics)
sum_ = 0
for i in range(num_topics):
dt_alpha = doc.topic_sum(i) + self.__model.alpha()
wt_beta = self.__model.word_topic_value(token.id, i) + self.__model.beta()
t_sum_beta_sum = self.__model.topic_sum(i) + self.__model.beta_sum()
if i == old_topic and wt_beta > 1:
if dt_alpha > 1:
dt_alpha -= 1
wt_beta -= 1
t_sum_beta_sum -= 1
prob[i] = dt_alpha * wt_beta / t_sum_beta_sum
sum_ += prob[i]
accum_prob[i] = prob[i] if i == 0 else accum_prob[i - 1] + prob[i]
dart = rand() * sum_
if dart <= accum_prob[0]:
return 0
for i in range(1, num_topics):
if accum_prob[i - 1] < dart <= accum_prob[i]:
return i
return num_topics - 1
def __propose(self, word_id):
dart = rand() * (self.__prob_sum[word_id] + self.__beta_prior_sum)
if dart < self.__prob_sum[word_id]:
idx = self.__alias_tables[word_id].generate()
topic = self.__topic_indexes[word_id][idx]
else:
topic = self.__beta_alias.generate()
return topic
def generate(self):
"""Generate samples from given distribution.
"""
dart1 = rand_k(self.size())
dart2 = int(rand())
return dart1 if dart2 > self.__prob[dart1] else self.__alias[dart1]