ldamodel.py 文件源码

def bound(self, corpus, gamma=None, subsample_ratio=1.0):
        """
        Estimate the variational bound of documents from `corpus`:
        E_q[log p(corpus)] - E_q[log q(corpus)]

        `gamma` are the variational parameters on topic weights for each `corpus`
        document (=2d matrix=what comes out of `inference()`).
        If not supplied, will be inferred from the model.

        """
        score = 0.0
        _lambda = self.state.get_lambda()
        Elogbeta = dirichlet_expectation(_lambda)

        for d, doc in enumerate(corpus):  # stream the input doc-by-doc, in case it's too large to fit in RAM
            if d % self.chunksize == 0:
                logger.debug("bound: at document #%i" % d)
            if gamma is None:
                gammad, _ = self.inference([doc])
            else:
                gammad = gamma[d]
            Elogthetad = dirichlet_expectation(gammad)

            # E[log p(doc | theta, beta)]
            score += numpy.sum(cnt * logsumexp(Elogthetad + Elogbeta[:, id]) for id, cnt in doc)

            # E[log p(theta | alpha) - log q(theta | gamma)]; assumes alpha is a vector
            score += numpy.sum((self.alpha - gammad) * Elogthetad)
            score += numpy.sum(gammaln(gammad) - gammaln(self.alpha))
            score += gammaln(numpy.sum(self.alpha)) - gammaln(numpy.sum(gammad))

        # compensate likelihood for when `corpus` above is only a sample of the whole corpus
        score *= subsample_ratio

        # E[log p(beta | eta) - log q (beta | lambda)]; assumes eta is a scalar
        score += numpy.sum((self.eta - _lambda) * Elogbeta)
        score += numpy.sum(gammaln(_lambda) - gammaln(self.eta))
        score += numpy.sum(gammaln(self.eta * self.num_terms) - gammaln(numpy.sum(_lambda, 1)))
        return score