def log_prob(self):
'''
Return the log_prob of the Training Set = log_likelihood + log_prior
Running log_prob will cache the log_likelihood, log_prior (for training set) and log_likelihood_test_set if they are not cached
'''
if not self.__ll_cached:
self.__ll = 0
# intermediate variables storing log_likelihood and log_prior seperately
# ll of the training set
self.__log_likelihood = sum(
message.log_prob() for message_list in self._messages
for message in message_list
if message.get_end_time() <= self._time_to_split)
# ll of the test set
self.__log_likelihood_test_set = sum(
message.log_prob() for message_list in self._messages
for message in message_list
if message.get_start_time() > self._time_to_split)
#
if self._use_prior:
self.__log_prior = 0
self.__log_prior += log_gamma(self._time_decay,
*self._time_decay_prior).sum()
if self._non_diagonal:
# off-diagonal entries
self.__log_prior += log_gamma(
self._influence[eye(self._A) == 0],
*self._influence_prior).sum()
else:
self.__log_prior += log_gamma(
self._influence, *self._influence_prior).sum()
self.__log_prior += log_gamma(
self._word_concentration,
*self._word_concentration_prior).sum()
self.__log_prior += log_gamma(
self._word_pseudocounts,
*self._word_pseudocount_prior).sum()
else:
# no prior
self.__log_prior = 0
self.__ll = self.__log_likelihood + self.__log_prior
self.__ll_cached = True
return self.__ll
def log_prior(self):
if self._prior_cached:
return self._log_prior
else:
assert all(self._hyperparams_tau>0)
# prior for tau
self._log_prior = sum(log_gamma(self._tau, *self._hyperparams_tau))
self._prior_cached = True
return self._log_prior