bbvi.py 文件源码

def cv_gradient(self, z):
        """
        The control variate augmented Monte Carlo gradient estimate
        RAO-BLACKWELLIZED!
        """        
        z_t = np.transpose(z)
        log_q = self.normal_log_q(z_t)
        log_p = self.log_p(z_t)
        grad_log_q = self.grad_log_q(z)
        gradient = grad_log_q*np.repeat((log_p - log_q).T,2,axis=0)

        alpha0 = alpha_recursion(np.zeros(np.sum(self.approx_param_no)), grad_log_q, gradient, np.sum(self.approx_param_no))      

        vectorized = gradient - ((alpha0/np.var(grad_log_q,axis=1))*grad_log_q.T).T

        return np.mean(vectorized,axis=1)