def bootstrap_evaluation(x, y, dropout, learning_rate, epochs, n_heads, ax):
sess, x_placeholder, dropout_placeholder, mask_placeholder =\
bootstrap_training(x, y, dropout, learning_rate, epochs, n_heads)
prediction_op = sess.graph.get_collection("prediction")
uncertainty_op = sess.graph.get_collection("uncertainties")
heads_op = sess.graph.get_collection("heads")
additional_range = 0.2 * np.max(x)
x_eval = np.linspace(
np.min(x) - additional_range,
np.max(x) + additional_range, 100).reshape([-1, 1])
feed_dict = {
x_placeholder: x_eval,
dropout_placeholder: 0,
mask_placeholder: np.ones(shape=(len(x_eval), n_heads, 1))
}
y_eval, uncertainties_eval, heads_eval = sess.run(
[prediction_op, uncertainty_op, heads_op], feed_dict)
heads_eval = np.array(heads_eval).reshape(len(x_eval), n_heads)
y_eval = y_eval[0].flatten()
uncertainties_eval = uncertainties_eval[0].flatten()
for i in range(n_heads):
ax.plot(x_eval, heads_eval[:, i], alpha=0.3)
plotting.plot_mean_vs_truth(x, y, x_eval, y_eval, uncertainties_eval, ax)
def dropout_evaluation(x, y, dropout, learning_rate, epochs, n_passes, ax):
# Hardcoded training dropout
sess, x_placeholder, dropout_placeholder = \
dropout_training(x, y, 0.2, learning_rate, epochs)
prediction_op = sess.graph.get_collection("prediction")
additional_range = 0.1 * np.max(x)
x_eval = np.linspace(np.min(x) - additional_range, np.max(x) + additional_range, 100).reshape([-1, 1])
feed_dict = {x_placeholder: x_eval,
dropout_placeholder: dropout}
predictions = []
for _ in range(n_passes):
predictions.append(sess.run(prediction_op, feed_dict)[0])
y_eval = np.mean(predictions, axis=0).flatten()
uncertainty_eval = np.var(predictions, axis=0).flatten()
plotting.plot_mean_vs_truth(x, y, x_eval, y_eval, uncertainty_eval, ax)
x_eval = np.linspace(np.min(x) - additional_range, np.max(x) + additional_range, 100).reshape([-1, 1])
x = x.reshape((-1, 1))
y = y.reshape((-1, 1))
fig, axs = plt.subplots(2, 1, figsize=(30, 10))
kernel = 0.5 * RBF(length_scale=0.01)
gp = GaussianProcessRegressor(kernel=kernel, alpha=0.03, n_restarts_optimizer=10)
y_prior = gp.sample_y(x_eval, 5)
axs[0].plot(x_eval, y_prior)
gp.fit(x, y)
y_eval, sigma = gp.predict(x_eval, return_std=True)
y_eval = y_eval.flatten()
plotting.plot_mean_vs_truth(x, y, x_eval, y_eval, sigma, axs[1])
axs[1].set_title("Posterior (kernel: %s)\n Log-Likelihood: %.3f"
% (gp.kernel_, gp.log_marginal_likelihood(gp.kernel_.theta)))
plt.show()
fig.savefig("GP_Sinus.pdf")
plt.close()
x, y = sample_generators.generate_osband_nonlinear_samples()
additional_range = 0.2 * np.max(x)
x_eval = np.linspace(np.min(x) - additional_range, np.max(x) + additional_range, 100).reshape([-1, 1])
x = x.reshape((-1, 1))
y = y.reshape((-1, 1))
fig, axs = plt.subplots(2, 1, figsize=(30, 10))
kernel = 1 * RBF(length_scale=1)
gp = GaussianProcessRegressor(kernel=kernel, alpha=0.5, n_restarts_optimizer=10)