hpseq.py 文件源码

def _beta_update_raw_tfidf(self):
        '''
        Run only once - it does not depend on other parameters.
        '''

        for nodeid in xrange(self.D):
            self.beta[nodeid] = self.W[self.node_vec == nodeid, :
                    ].sum(axis=0)
        for nodeid in xrange(self.D):
            for wordid in xrange(self.beta.shape[1]):
                docs_cnt = np.sum(self.W[self.node_vec == nodeid,
                                  wordid] >= 1)
                docs_cnt += 1  # smooth by adding one
                self.beta[nodeid][wordid] *= 1 + np.log(self.W.shape[0]
                        * 1. / docs_cnt)  # 1+ because we still want to keep words which always occurr, but probably it never happens

        # Laplace smoothing to avoid zeros!

        self.beta += 1
        self._normalize_beta_rowwise()
        return self.beta