def setUp(self):
f = lambda x: np.sin(2 * np.pi * 10. * x) + np.sin(2 * np.pi * 13. * x)
Ts = 1 / 150.0
self.x = np.atleast_2d(np.arange(0, 1, Ts)).T
self.xt = np.atleast_2d(np.arange(0, 2, Ts)).T
self.y = np.atleast_2d(f(self.x).ravel()).T
self.yt = np.atleast_2d(f(self.xt).ravel()).T
self.hyps = {
'mean':
np.array([]),
'lik': [np.log(0.5)],
'cov':
np.log([
2.54846959e-01, 5.24927544e-01, 2.15436686e-01, 2.86117061e-01,
1.81531436e-01, 1.44032999e-01, 1.29493734e+01, 9.92190150e+00,
1.85152636e+02, 7.49453588e+01, 1.81532400e+00, 2.28497943e+01,
1.03092807e-01, 4.88398945e-02, 1.94985263e+01, 3.19829864e+01,
2.12244665e+00, 8.39227875e-01
])
}
self.hypGP = gp.GaussianProcess(hyp=self.hyps,
inf=gp.inferences.exact,
mean=gp.means.zero,
cov=gp.kernels.spectral_mixture,
lik=gp.likelihoods.gaussian,
xtrain=self.x,
ytrain=self.y,
xtest=self.xt,
ytest=self.yt)
l1Optimizer = 'COBYLA'
l1Options = {'maxiter': 100}
l2Optimizer = 'L-BFGS-B'
l2Options = {'maxiter': 1000}
# Noise std. deviation
sn = 1.0
# Initialize hyperparams
initArgs = {'Q': Q, 'x': x, 'y': y, 'samplingFreq': 1, 'nPeaks': Q, 'sn': sn}
hypGuess = gp.core.initSMParamsFourier(**initArgs)
# Initialize GP object
hypGP = gp.GaussianProcess(hyp=hypGuess,
inf=infFunc,
mean=meanFunc,
cov=covFunc,
lik=likFunc,
xtrain=x,
ytrain=y,
xtest=xt)
# Random starts
for itr in range(nItr):
# Start over
hypGuess = gp.core.initSMParamsFourier(**initArgs)
hypGP.hyp = hypGuess
# Optimize the guessed hyperparams
try:
hypGP.train(l1Optimizer, l1Options)
except Exception as e:
print "Iteration: ", itr, "FAILED"
print "\t", e
continue
import torch
import numpy as np
import kernel as k
import gaussian_process as gp
import bayes_opt as bo
def ackley(x):
exp1 = -0.2*(0.5*x.pow(2)).sum().pow(0.5)
exp2 = 0.5*torch.cos(2*np.pi*x).sum()
y = np.e + 20 - 20*torch.exp(exp1) - torch.exp(exp2)
return y
n_dimensions = 2
bounds = torch.tensor([[-10.0, 10.0]]).expand(n_dimensions, 2)
iters = 50
kernel = k.Matern32Kernel(1.0, 1.0)
model = gp.GaussianProcess(kernel, 0.001)
acq_func = bo.Acquisition(1.0)
optimiser = bo.BayesOptimiser(model, acq_func)
x_min, y_min, x_hist, y_hist = optimiser.minimise(ackley, bounds, iters,
iter_loops=10)
mu_grad = gp.mu_grad(x)
sigma_grad = gp.sigma_grad(x)
grad = mu_grad - self.beta * sigma_grad
return grad
def ackley(x):
x = torch.from_numpy(x)
x = x.view(-1, 1)
exp1 = -0.2 * (0.5 * x.pow(2).sum(1)).pow(0.5)
exp2 = 0.5 * torch.cos(2 * np.pi * x).sum(1)
y = np.e + 20 - 20 * torch.exp(exp1) - torch.exp(exp2)
return y.data.numpy()
bounds = torch.tensor([[-8.0, 8.0]]).expand(1, 2)
iters = 20
kernel = krn.Matern32Kernel(1.0, 1.0, 'cpu') * \
krn.Matern52Kernel(1.0, 1.0, 'cpu')
gp = gaussp.GaussianProcess(kernel, alpha=0.0001)
acq_func = Acquisition(1.0)
optimiser = BayesOptimiser(gp, acq_func)
x_min, y_min, xp, yp = optimiser.minimise(ackley, bounds, iters)
filenames = glob.glob('ackley_frame*.png')
filenames = sorted(filenames)
gif_images = []
for filename in filenames:
gif_images.append(imageio.imread(filename))
imageio.mimsave('ackley.gif', gif_images, duration=1 / 2)