def __init__(self, pca, eparams):
'''
Provide the emulation products.
:param pca: object storing the principal components, the eigenpsectra
:type pca: PCAGrid
:param eparams: Optimized GP hyperparameters.
:type eparams: 1D np.array
'''
self.pca = pca
self.lambda_xi = eparams[0]
self.h2params = eparams[1:].reshape(self.pca.m, -1) ** 2
# Determine the minimum and maximum bounds of the grid
self.min_params = np.min(self.pca.gparams, axis=0)
self.max_params = np.max(self.pca.gparams, axis=0)
# self.eigenspectra = self.PCAGrid.eigenspectra
self.dv = self.pca.dv
self.wl = self.pca.wl
self.iPhiPhi = (1. / self.lambda_xi) * np.linalg.inv(skinny_kron(self.pca.eigenspectra, self.pca.M))
self.V11 = self.iPhiPhi + Sigma(self.pca.gparams, self.h2params)
self._params = None # Where we want to interpolate
self.V12 = None
self.V22 = None
self.mu = None
self.sig = None
def lnprob(p, fmin=False):
'''
:param p: Gaussian Processes hyper-parameters
:type p: 1D np.array
Calculate the lnprob using Habib's posterior formula for the emulator.
'''
# We don't allow negative parameters.
if np.any(p < 0.):
if fmin:
return 1e99
else:
return -np.inf
lambda_xi = p[0]
hparams = p[1:].reshape((my_pca.m, -1))
# Calculate the prior for parname variables
# We have two separate sums here, since hparams is a 2D array
# hparams[:, 0] are the amplitudes, so we index i+1 here
lnpriors = 0.0
for i in range(0, len(Starfish.parname)):
lnpriors += np.sum(Glnprior(hparams[:, i + 1], *priors[i]))
h2params = hparams**2
#Fold hparams into the new shape
Sig_w = Sigma(my_pca.gparams, h2params)
C = (1. / lambda_xi) * PhiPhi + Sig_w
sign, pref = np.linalg.slogdet(C)
central = my_pca.w_hat.T.dot(np.linalg.solve(C, my_pca.w_hat))
lnp = -0.5 * (pref + central +
my_pca.M * my_pca.m * np.log(2. * np.pi)) + lnpriors
# Negate this when using the fmin algorithm
if fmin:
print("lambda_xi", lambda_xi)
for row in hparams:
print(row)
print()
print(lnp)
return -lnp
else:
return lnp