dgpr_aep_examples.py 文件源码

def run_step_1D():
    np.random.seed(42)

    def step(x):
        y = x.copy()
        y[y < 0.0] = 0.0
        y[y > 0.0] = 1.0
        return y + 0.02 * np.random.randn(x.shape[0], 1)

    print "create dataset ..."
    N = 100
    X = np.random.rand(N, 1) * 3 - 1.5
    Y = step(X) - 0.5
    # plt.plot(X, Y, 'kx', mew=2)

    def plot(m):
        xx = np.linspace(-3, 3, 100)[:, None]
        mean, var = m.predict_f(xx)
        zu = m.sgp_layers[0].zu
        mean_u, var_u = m.predict_f(zu)
        plt.figure()
        plt.plot(X, Y, 'kx', mew=2)
        plt.plot(xx, mean, 'b', lw=2)
        plt.fill_between(
            xx[:, 0],
            mean[:, 0] - 2 * np.sqrt(var[:, 0]),
            mean[:, 0] + 2 * np.sqrt(var[:, 0]),
            color='blue', alpha=0.2)
        plt.errorbar(zu, mean_u, yerr=2 * np.sqrt(var_u), fmt='ro')

        no_samples = 20
        xx = np.linspace(-3, 3, 500)[:, None]
        f_samples = m.sample_f(xx, no_samples)
        for i in range(no_samples):
            plt.plot(xx, f_samples[:, :, i], linewidth=0.5, alpha=0.5)

        plt.xlim(-3, 3)

    # inference
    print "create model and optimize ..."
    M = 20
    hidden_size = [2]
    model = aep.SDGPR(X, Y, M, hidden_size, lik='Gaussian')
    # model.optimise(method='L-BFGS-B', alpha=1, maxiter=1000)
    model.optimise(method='adam', adam_lr=0.05, alpha=1, maxiter=2000)
    plot(model)
    plt.show()