posterior.py 文件源码

def prior_and_posterior_for_category(self, category_to_primitives):
        category_to_prior = {}
        category_to_posterior = {}
        # We want to take the logs of the lengh prior, but some elements 
        # will be zero; entries in the result being -inf is okay, but we want
        # to avoid a potentially alarming warning message
        M = np.zeros(self.length_prior.shape)
        M[self.length_prior > 0.] = np.log(self.length_prior[self.length_prior > 0.])
        M[self.length_prior <= 0.] = -np.inf
        lengths = np.arange(len(M))

        for category, primitives in category_to_primitives.iteritems():
            base_prior = 0.
            for primitive in primitives:
                base_prior += self.base_prior_by_primitive[primitive]
            # Effectively we calculate the probability of never
            # choosing this category in a rule with 0 primitives, 1
            # primitive, etc, then add those up, scaling by the
            # probability of having 0 primitives, 1 primitive, etc.;
            # for speed and numerical behavior we do this in a
            # vectorized way on a log scale:
            inclusion_prior = -1.0*np.expm1(
                logsumexp(M + (np.log1p(-1.0*base_prior) * lengths))
            )
            category_to_prior[category] = inclusion_prior

            states_in_category = 0
            for state in self.state_ensemble_primitives:
                if state.intersection(primitives):
                    states_in_category += 1
            category_to_posterior[category] = (
                states_in_category / float(len(self.state_ensemble))
            )
        return category_to_prior, category_to_posterior