test_grads_aep.py 文件源码

def plot_gplvm_aep_gaussian_stochastic():
    N_train = 2000
    alpha = 0.5
    M = 50
    D = 2
    Q = 3
    y_train = np.random.randn(N_train, Q)
    model = aep.SGPLVM(y_train, D, M, lik='Gaussian')

    # init hypers, inducing points and q(u) params
    params = model.init_hypers(y_train)
    logZ, grad_all = model.objective_function(params, N_train, alpha=alpha)
    mbs = np.logspace(-2, 0, 10)
    reps = 20
    times = np.zeros(len(mbs))
    objs = np.zeros((len(mbs), reps))
    for i, mb in enumerate(mbs):
        no_points = int(N_train * mb)
        start_time = time.time()
        for k in range(reps):
            objs[i, k] = model.objective_function(
                params, no_points, alpha=alpha)[0]
        times[i] = time.time() - start_time

    f, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 6))
    ax1.plot(mbs, times, 'x-')
    ax1.set_xlabel("Minibatch proportion")
    ax1.set_ylabel("Time taken")
    ax1.set_xscale("log", nonposx='clip')

    ax2.plot(mbs, objs, 'kx')
    ax2.axhline(logZ, color='b')
    ax2.set_xlabel("Minibatch proportion")
    ax2.set_ylabel("ELBO estimates")
    ax2.set_xscale("log", nonposx='clip')
    plt.savefig('/tmp/gaussian_stochastic_aep_gplvm.pdf')