def __init__(self, mean, covariance, points, lnlikes, quiet=True):
self.mean = mean
self.cov = covariance
#Let's try only interpolating over points that are
#better than, say, 10-ish sigma
dof = len(self.mean)
inds = np.fabs(np.max(lnlikes) - lnlikes) < 100 * dof
print(inds)
print(lnlikes)
self.points = points[inds]
self.lnlikes_true = lnlikes[inds]
self.lnlike_max = np.max(lnlikes[inds])
self.lnlikes = lnlikes[inds] - self.lnlike_max #max is now at 0
self.x = self._transform_data(self.points)
print(max(self.lnlikes), min(self.lnlikes))
_guess = 4.5 #kernel length guess
kernel = kernels.ExpSquaredKernel(metric=_guess, ndim=dof)
lnPmin = np.min(self.lnlikes)
gp = GP(kernel, mean=lnPmin - np.fabs(lnPmin * 3))
gp.compute(self.x)
def neg_ln_likelihood(p):
gp.set_parameter_vector(p)
return -gp.log_likelihood(self.lnlikes)
def grad_neg_ln_likelihood(p):
gp.set_parameter_vector(p)
return -gp.grad_log_likelihood(self.lnlikes)
result = minimize(neg_ln_likelihood,
gp.get_parameter_vector(),
jac=grad_neg_ln_likelihood)
if not quiet:
print(result)
gp.set_parameter_vector(result.x)
self.gp = gp
def train(self, kernel=None):
"""Train a Gaussian Process to interpolate the log-likelihood
of the training samples.
Args:
kernel (george.kernels.Kernel object): kernel to use, or any
acceptable object that can be accepted by the george.GP object
"""
inds = self.training_inds
x = self.chain_rotated_regularized[inds]
lnL = self.lnlikes[inds]
_guess = 4.5
if kernel is None:
kernel = kernels.ExpSquaredKernel(metric=_guess, ndim=len(x[0]))
#Note: the mean is set slightly lower that the minimum lnlike
lnPmin = np.min(self.lnlikes)
gp = GP(kernel, mean=lnPmin - np.fabs(lnPmin * 3))
gp.compute(x)
def neg_ln_likelihood(p):
gp.set_parameter_vector(p)
return -gp.log_likelihood(lnL)
def grad_neg_ln_likelihood(p):
gp.set_parameter_vector(p)
return -gp.grad_log_likelihood(lnL)
result = minimize(neg_ln_likelihood,
gp.get_parameter_vector(),
jac=grad_neg_ln_likelihood)
print(result)
gp.set_parameter_vector(result.x)
self.gp = gp
self.lnL_training = lnL
return
def test_gradient(solver, white_noise, seed=123, N=305, ndim=3, eps=1.32e-3):
np.random.seed(seed)
# Set up the solver.
kernel = 1.0 * kernels.ExpSquaredKernel(0.5, ndim=ndim)
kwargs = dict()
if white_noise is not None:
kwargs = dict(white_noise=white_noise, fit_white_noise=True)
if solver == HODLRSolver:
kwargs["tol"] = 1e-8
gp = GP(kernel, solver=solver, **kwargs)
# Sample some data.
x = np.random.rand(N, ndim)
x = x[np.argsort(x[:, 0])]
y = gp.sample(x)
gp.compute(x, yerr=0.1)
# Compute the initial gradient.
grad0 = gp.grad_log_likelihood(y)
vector = gp.get_parameter_vector()
for i, v in enumerate(vector):
# Compute the centered finite difference approximation to the gradient.
vector[i] = v + eps
gp.set_parameter_vector(vector)
lp = gp.lnlikelihood(y)
vector[i] = v - eps
gp.set_parameter_vector(vector)
lm = gp.lnlikelihood(y)
vector[i] = v
gp.set_parameter_vector(vector)
grad = 0.5 * (lp - lm) / eps
assert np.abs(grad - grad0[i]) < 5 * eps, \
"Gradient computation failed in dimension {0} ({1})\n{2}" \
.format(i, solver.__name__, np.abs(grad - grad0[i]))
