hdpmodel.py 文件源码

def update_chunk(self, chunk, update=True, opt_o=True):
        # Find the unique words in this chunk...
        unique_words = dict()
        word_list = []
        for doc in chunk:
            for word_id, _ in doc:
                if word_id not in unique_words:
                    unique_words[word_id] = len(unique_words)
                    word_list.append(word_id)

        Wt = len(word_list) # length of words in these documents

        # ...and do the lazy updates on the necessary columns of lambda
        rw = np.array([self.m_r[t] for t in self.m_timestamp[word_list]])
        self.m_lambda[:, word_list] *= np.exp(self.m_r[-1] - rw)
        self.m_Elogbeta[:, word_list] = \
            sp.psi(self.m_eta + self.m_lambda[:, word_list]) - \
            sp.psi(self.m_W * self.m_eta + self.m_lambda_sum[:, np.newaxis])

        ss = SuffStats(self.m_T, Wt, len(chunk))

        Elogsticks_1st = expect_log_sticks(self.m_var_sticks) # global sticks

        # run variational inference on some new docs
        score = 0.0
        count = 0
        for doc in chunk:
            if len(doc) > 0:
                doc_word_ids, doc_word_counts = zip(*doc)
                doc_score = self.doc_e_step(doc, ss, Elogsticks_1st,
                    word_list, unique_words, doc_word_ids,
                    doc_word_counts, self.m_var_converge)
                count += sum(doc_word_counts)
                score += doc_score

        if update:
            self.update_lambda(ss, word_list, opt_o)

        return (score, count)