svgd.py 文件源码

def update(self, x0, lnprob, n_iter = 1000, stepsize = 1e-3, bandwidth = -1, alpha = 0.9, debug = False):
        # Check input
        if x0 is None or lnprob is None:
            raise ValueError('x0 or lnprob cannot be None!')

        theta = np.copy(x0) 

        # adagrad with momentum
        fudge_factor = 1e-6
        historical_grad = 0
        for iter in range(n_iter):
            if debug and (iter+1) % 1000 == 0:
                print 'iter ' + str(iter+1) 

            lnpgrad = lnprob(theta)
            # calculating the kernel matrix
            kxy, dxkxy = self.svgd_kernel(theta, h = -1)  
            grad_theta = (np.matmul(kxy, lnpgrad) + dxkxy) / x0.shape[0]  

            # adagrad 
            if iter == 0:
                historical_grad = historical_grad + grad_theta ** 2
            else:
                historical_grad = alpha * historical_grad + (1 - alpha) * (grad_theta ** 2)
            adj_grad = np.divide(grad_theta, fudge_factor+np.sqrt(historical_grad))
            theta = theta + stepsize * adj_grad 

        return theta