def test_numerical_gradient():
debug_output = False
random = np.random.RandomState(42)
eps = 1e-6
for model in default_models():
ei = EIAcquisitionFunction(model)
for iter in range(10):
high = 1.0 if iter < 5 else 0.02
x = random.uniform(low=0.0, high=high, size=(2, ))
f0, analytical_gradient = ei.compute_acq_with_gradient(x)
analytical_gradient = analytical_gradient.flatten()
if debug_output:
print('x0 = {}, f(x_0) = {}, grad(x_0) = {}'.format(
x, f0, analytical_gradient))
for i in range(2):
h = np.zeros_like(x)
h[i] = eps
fpeps = ei.compute_acq(x + h)[0]
fmeps = ei.compute_acq(x - h)[0]
numerical_derivative = (fpeps - fmeps) / (2 * eps)
if debug_output:
print(
'f(x0+eps) = {}, f(x0-eps) = {}, findiff = {}, deriv = {}'
.format(fpeps[0], fmeps[0], numerical_derivative[0],
analytical_gradient[i]))
np.testing.assert_almost_equal(numerical_derivative.item(),
analytical_gradient[i],
decimal=4)
def test_value_same_as_with_gradient():
# test that compute_acq and compute_acq_with_gradients return the same acquisition values
for model in default_models():
ei = EIAcquisitionFunction(model)
random = np.random.RandomState(42)
X = random.uniform(low=0.0, high=1.0, size=(10, 2))
# assert same as computation with gradients
vec1 = ei.compute_acq(X).flatten()
vec2 = np.array([ei.compute_acq_with_gradient(x)[0] for x in X])
np.testing.assert_almost_equal(vec1, vec2)
def test_optimization_improves():
debug_output = False
# Pick a random point, optimize and the expected improvement should be better:
# But only if the starting point is not too far from the origin
random = np.random.RandomState(42)
for model in default_models():
ei = EIAcquisitionFunction(model)
opt = LBFGSOptimizeAcquisition(model.state, model,
EIAcquisitionFunction)
if debug_output:
print('\n\nGP MCMC' if model.does_mcmc() else 'GP Opt')
fzero = ei.compute_acq(np.zeros((1, 2)))[0]
print('f(0) = {}'.format(fzero))
if debug_output and not model.does_mcmc():
print('Hyperpars: {}'.format(model.get_params()))
# Plot the thing!
plot_ei_mean_std(model, ei, max_grid=0.001)
plot_ei_mean_std(model, ei, max_grid=0.01)
plot_ei_mean_std(model, ei, max_grid=0.1)
plot_ei_mean_std(model, ei, max_grid=1.0)
non_zero_acq_at_least_once = False
for iter in range(10):
#initial_point = random.uniform(low=0.0, high=1.0, size=(2,))
initial_point = random.uniform(low=0.0, high=0.1, size=(2, ))
acq0, df0 = ei.compute_acq_with_gradient(initial_point)
if debug_output:
print('\nInitial point: f(x0) = {}, x0 = {}'.format(
acq0, initial_point))
print('grad0 = {}'.format(df0))
if acq0 != 0:
non_zero_acq_at_least_once = True
optimized = np.array(opt.optimize(tuple(initial_point)))
acq_opt = ei.compute_acq(optimized)[0]
if debug_output:
print('Final point: f(x1) = {}, x1 = {}'.format(
acq_opt, optimized))
assert acq_opt < 0
assert acq_opt < acq0
assert non_zero_acq_at_least_once