lda.py 文件源码

def perplexity(self):
        log_likelihood = 0.
        for doc_ix in xrange(self.num_docs):
            prob_topic_given_document = np.zeros(self.K)
            for i in xrange(self.K):
                prob_topic_given_document[i] = (self.ndt[doc_ix][i] + self.alpha) / (
                    self.nd[doc_ix] + self.K * self.alpha)

            for word in self.corpus[doc_ix]:
                prob_word_given_topic = np.zeros(self.K)
                word_ix = self.word2ix[word]
                for j in xrange(self.K):
                    prob_word_given_topic[j] = (self.ntw[j][word_ix] + self.beta) / (self.nt[j] + self.K * self.alpha)

                prob_word = 0.
                for j in xrange(self.K):
                    prob_word += prob_topic_given_document[j] * prob_word_given_topic[j]

                log_likelihood += np.log(prob_word)
        perplexity = np.exp2(-log_likelihood / self.num_docs)
        return perplexity