bbvi.py 文件源码

def cv_gradient(self,z):
        """
        The control variate augmented Monte Carlo gradient estimate
        """
        gradient = np.zeros(np.sum(self.approx_param_no))
        z_t = z.T      
        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*(log_p-log_q)

        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)