def test_GP_2D(optimize=True, function=np.linspace):
dim_test = 2
dim_out = 1
n_train_p = 7
X = np.random.uniform(-1, 1, (14, 2))
Z = ((X[:, 1]**2 * X[:, 0]**2) * np.sin((X[:, 1]**2 + X[:, 0]**2)))[:,
None]
#Z=np.sin((X[:, 1] ** 2 + X[:, 0] ** 2))[:,None]
gp = GP(X, Z, kernel=RBF())
gp.fit()
plot = generate_grid(dim_test, 100, [[-1, 1] for i in range(dim_test)])
print(plot.shape)
#pred = gp.predict(plot)
#gp.plot(plot)
# gp.static_compute_marg()
#print("Old marg likelihood :", gp.get_marg(),
#"\n Hyperparameters: ", gp.get_kernel().gethyper())
if optimize:
gp.optimize(n_restarts=30)
pred = gp.predict(plot)
gp.plot(plot)
print("New marg likelihood :", gp.get_marg(), "\n Hyperparameters: ",
gp.get_kernel().gethyper())
def test_GP_1D(optimize=True):
#x = np.arange(-3, 5, 1)[:, None]
#x=np.array([0.1,0.12,0.143,0.3,0.5,0.75,0.67,0.9,0.92,1.1])[:,None]
x = np.random.uniform(-3, 3, 16)[:, None]
def f(X):
#return -(1.4-3*X)*np.sin(18*X)
#return X**0
return np.sin(X)
#return (6 * X - 2) ** 2 * np.sin(12 * X - 4) - X
#return X + np.sin(X)*10
def noise(x, alpha=1):
return f(x) + np.random.randn(*x.shape) * alpha
#y = noise(x, alpha=0)
y = f(x)
print(y)
#gp = GP(x*10000, y*1000, kernel=RBF(sigma_l=0.2, l= 1, noise= 1e-3, gradient=False), normalize_y=True)
gp = GP(x, y, kernel=RBF(gradient=False), normalize_y=False)
gp.fit()
plot = np.linspace(-20, 20, 1000)
#pred_old, var_old = gp.predict(plot[:, None])
#gp.plot(plot[:, None])
gp.log_marginal_likelihood()
print("Old marg likelihood :", gp.get_marg(), "\n Hyperparameters: ",
gp.get_kernel().gethyper())
if optimize:
"""new = gp.grid_search_optimization(constrains=[[1, 30], [1, 30],[0.00001,1]],
n_points=100,
function=np.linspace)"""
gp.optimize(n_restarts=10, optimizer="L-BFGS-B", verbose=False)
#gp.optimize_grid(n_points=50)
#optimized.fit()
#pred, var = gp.predict(plot[:, None])
#plt.plot(plot[:,None],f(plot))
gp.plot(plot[:, None])
#plt.scatter(x,y,marker="x",color="red")
gp.log_marginal_likelihood()
log_gp(gp)