# 15 initial values for the 7 hod and 8 cosmo params
p0 = np.full(nparams, 0.1)
k1 = kernels.ExpSquaredKernel(p0, ndim=len(p0))
k2 = kernels.Matern32Kernel(p0, ndim=len(p0))
k3 = kernels.ConstantKernel(0.1, ndim=len(p0))
#k4 = kernels.WhiteKernel(0.1, ndim=len(p0))
k5 = kernels.ConstantKernel(0.1, ndim=len(p0))
kernel = k2 + k5
#kernel = np.var(y)*k1
ppt = pp[j]
gp = george.GP(kernel, mean=np.mean(y), solver=george.BasicSolver)
#gp = george.GP(kernel, solver=george.BasicSolver)
gp.compute(rr, yerr)
#gp.kernel.vector = ppt
gp.set_parameter_vector(ppt)
gp.compute(rr, yerr)
gps.append(gp)
if fixed_hod:
#HH_test = range(hod, hod+1)
CC_test = range(0, 7)
# TODO: add more tests, for now just did first 10 hod
def lnlike_gp(params, wl, flux, fluxerror):
a, tau = np.exp(params[:2])
gp = george.GP(a * kernels.Matern32Kernel(tau))
gp.compute(wl, fluxerror)
return gp.lnlikelihood(flux - model(params[2:], wl))
def run_vs_3paramFit():
args = parse_args()
ext = args.output_file_extension
from pygp.canvas import Canvas
xMinFit = 300
# Getting data points
xRaw, yRaw, xerrRaw, yerrRaw = getDataPoints(args.input_file,
'dijetgamma_g85_2j65',
'Zprime_mjj_var')
xSig, ySig, xerrSig, yerrSig = getDataPoints(args.signal_file,
'dijetgamma_g85_2j65',
'Zprime_mjj_var')
#Data processing, cutting out the not desired range
x, y, xerr, yerr = dataCut(xMinFit, 1500, 0, xRaw, yRaw, xerrRaw, yerrRaw)
xFit, yFit, xerrFit, yerrFit = dataCut(xMinFit, 1500, 0, xRaw, yRaw,
xerrRaw, yerrRaw)
# make an evently spaced x
t = np.linspace(np.min(x), np.max(x), 500)
#calculating the log-likihood and minimizing for the gp
lnProb = logLike_minuit(x, y, xerr)
#
# idecay = np.random.random() * 0.64
# ilength = np.random.random() * 5e5
# ipower = np.random.random() * 1.0
# isub = np.random.random() * 1.0
# 'amp': 5701461179.0,
# 'p0': 0.23,
# 'p1': 0.46,
# 'p2': 0.89
#
#dan crap initial guess->returns infiinity
print("prob:", lnProb(5701461179.0, 0.64, 5e5, 1.0, 1.0, 0.23, 0.46, 0.89))
print(
"prob:",
lnProb(7894738685.23, 91.507809530688036, 152892.47486888882,
0.77936430405302681, 393.57106174440003, 0.46722747265429021,
0.92514297129196166, 1.7803224928740065))
#def __call__(self, Amp, decay, length, power, sub, p0, p1, p2):
min_likelihood, best_fit = fit_gp_minuit(10, lnProb)
fit_pars = [best_fit[x] for x in FIT3_PARS]
#making the GP
kargs = {x: y for x, y in best_fit.items() if x not in FIT3_PARS}
kernel_new = get_kernel(**kargs)
print(kernel_new.get_parameter_names())
#making the kernel
gp_new = george.GP(kernel_new, mean=Mean(fit_pars), fit_mean=True)
gp_new.compute(x, yerr)
mu, cov = gp_new.predict(y, t)
mu_x, cov_x = gp_new.predict(y, x)
# calculating the fit function value
#GP compute minimizes the log likelihood of the best_fit function
best = [best_fit[x] for x in FIT3_PARS]
print("best param meghan GP minmization:", best)
meanFromGPFit = Mean(best)
fit_meanM = meanFromGPFit.get_value(
x, xerr) #so this is currently shit. can't get the xErr thing working
print("fit_meanM:", fit_meanM)
#fit_mean_smooth = gp_new.mean.get_value(t)
#----3 param fit function in a different way
lnProb = logLike_3ff(xFit, yFit, xerrFit)
minimumLLH, best_fit_params = fit_3ff(100, lnProb)
fit_mean = model_3param(xFit, best_fit_params, xerrFit)
##----4 param fit function
#lnProb = logLike_4ff(xFit,yFit,xerrFit)
#minimumLLH, best_fit_params = fit_4ff(100, lnProb)
#fit_mean = model_4param(xFit, best_fit_params, xerrFit)
#calculating significance
signif = significance(mu_x, y, cov_x, yerr)
initialCutPos = np.argmax(x > xMinFit)
#sigFit = (fit_mean - y[initialCutPos:]) / np.sqrt(np.diag(cov_x[initialCutPos:]) + yerr[initialCutPos:]**2)
sigFit = significance(fit_mean, y[initialCutPos:], cov_x[initialCutPos:],
yerr[initialCutPos:])
#sigit = (mu_x)
std = np.sqrt(np.diag(cov))
ext = args.output_file_extension
title = "test"
with Canvas(f'%s{ext}' % title) as can:
can.ax.errorbar(x, y, yerr=yerr, fmt='.')
can.ax.set_yscale('log')
# can.ax.set_ylim(1, can.ax.get_ylim()[1])
can.ax.plot(t, mu, '-r')
can.ax.plot(xFit, fit_mean, '-b')
#this only works with the xErr part of Mean commented out
#can.ax.plot(x, fit_meanM, '.g')
# can.ax.plot(t, fit_mean_smooth, '--b')
#can.ax.fill_between(t, mu - std, mu + std,
#facecolor=(0, 1, 0, 0.5),
#zorder=5, label='err = 1')
#can.ratio.stem(x, signif, markerfmt='.', basefmt=' ')
can.ratio.stem(xFit, sigFit, markerfmt='.', basefmt=' ')
can.ratio.axhline(0, linewidth=1, alpha=0.5)
can.ratio2.stem(x, signif, markerfmt='.', basefmt=' ')
can.save(title)
def create_testcase07():
import george
bjd0 = bjd_obs - Tref
err = bjd0 * 0 + instrument['RV_precision']
gp_pams = np.zeros(4)
gp_pams[0] = np.log(activity['Hamp_RV1']) * 2
gp_pams[1] = np.log(activity['Pdec']) * 2
gp_pams[2] = 1. / (2 * activity['Oamp']**2)
gp_pams[3] = np.log(activity['Prot'])
kernel = np.exp(gp_pams[0]) * \
george.kernels.ExpSquaredKernel(metric=np.exp(gp_pams[1])) * \
george.kernels.ExpSine2Kernel(gamma=gp_pams[2], log_period=gp_pams[3])
gp = george.GP(kernel)
gp.compute(bjd0, err)
prediction1 = gp.sample(bjd0)
gp_pams[0] = np.log(activity['Hamp_RV2']) * 2
kernel = np.exp(gp_pams[0]) * \
george.kernels.ExpSquaredKernel(metric=np.exp(gp_pams[1])) * \
george.kernels.ExpSine2Kernel(gamma=gp_pams[2], log_period=gp_pams[3])
gp = george.GP(kernel)
gp.compute(bjd0, err)
prediction2 = gp.sample(bjd0)
y_pla = kp.kepler_RV_T0P(bjd0, planet_b['f'], planet_b['P'], planet_b['K'],
planet_b['e'], planet_b['o'])
mod_pl1 = np.random.normal(y_pla + prediction1 + instrument['RV_offset1'],
instrument['RV_precision'])
mod_pl2 = np.random.normal(y_pla + prediction2 + instrument['RV_offset2'],
instrument['RV_precision'])
Tcent_b = np.random.normal(
np.arange(0, 1) * planet_b['P'] + kp.kepler_Tcent_T0P(
planet_b['P'], planet_b['f'], planet_b['e'], planet_b['o']) + Tref,
instrument['T0_precision'])
fileout = open('TestCase07_RV_dataset1.dat', 'w')
for b, r in zip(bjd_obs, mod_pl1):
fileout.write('{0:f} {1:.2f} {2:.2f} {3:d} {4:d} {5:d}\n'.format(
b, r, instrument['RV_precision'], 0, 0, -1))
fileout.close()
fileout = open('TestCase07_RV_dataset2.dat', 'w')
for b, r in zip(bjd_obs, mod_pl2):
fileout.write('{0:f} {1:.2f} {2:.2f} {3:d} {4:d} {5:d}\n'.format(
b, r, instrument['RV_precision'], 0, 0, -1))
fileout.close()
fileout = open('TestCase07_Tcent_b.dat', 'w')
for i_Tc, v_Tc in enumerate(Tcent_b):
fileout.write('{0:d} {1:.4f} {2:.4f} {3:d}\n'.format(
i_Tc, v_Tc, instrument['T0_precision'], 0))
fileout.close()
idx_params = np.array(eparams.split(',')).astype('int')
X = X[idx_params,:]
# Sace other inputs:
ld_law = args.ldlaw
pmean = np.double(args.pmean)
psd = np.double(args.psd)
n_live_points = int(args.nlive)
# Cook the george kernel:
import george
kernel = np.var(f)*george.kernels.ExpSquaredKernel(np.ones(X.shape[0]),ndim=X.shape[0],axes=range(X.shape[0]))
# Cook jitter term
jitter = george.modeling.ConstantModel(np.log((200.*1e-6)**2.))
# Wrap GP object to compute likelihood
gp = george.GP(kernel, mean=0.0,fit_mean=False,white_noise=jitter,fit_white_noise=True)
#print gp.get_parameter_names(),gp.get_parameter_vector()
#print dir(gp)
#sys.exit()
gp.compute(X.T)
# Define transit-related functions:
def reverse_ld_coeffs(ld_law, q1, q2):
if ld_law == 'quadratic':
coeff1 = 2.*np.sqrt(q1)*q2
coeff2 = np.sqrt(q1)*(1.-2.*q2)
elif ld_law=='squareroot':
coeff1 = np.sqrt(q1)*(1.-2.*q2)
coeff2 = 2.*np.sqrt(q1)*q2
elif ld_law=='logarithmic':
coeff1 = 1.-np.sqrt(q1)*q2
def train(self, X, y, do_optimize=True, **kwargs):
X_norm, _, _ = normalization.zero_one_normalization(
X[:, :-1], self.lower, self.upper)
s_ = self.basis_func(X[:, -1])[:, None]
self.X = np.concatenate((X_norm, s_), axis=1)
if self.normalize_output:
# Normalize output to have zero mean and unit standard deviation
self.y, self.y_mean, self.y_std = normalization.zero_mean_unit_var_normalization(
y)
else:
self.y = y
# Use the mean of the data as mean for the GP
mean = np.mean(self.y, axis=0)
self.gp = george.GP(self.kernel, mean=mean)
if do_optimize:
# We have one walker for each hyperparameter configuration
sampler = emcee.EnsembleSampler(self.n_hypers,
len(self.kernel) + 1,
self.loglikelihood)
# Do a burn-in in the first iteration
if not self.burned:
# Initialize the walkers by sampling from the prior
if self.prior is None:
self.p0 = np.random.rand(self.n_hypers,
len(self.kernel.pars) + 1)
else:
self.p0 = self.prior.sample_from_prior(self.n_hypers)
# Run MCMC sampling
self.p0, _, _ = sampler.run_mcmc(self.p0,
self.burnin_steps,
rstate0=self.rng)
self.burned = True
# Start sampling
pos, _, _ = sampler.run_mcmc(self.p0,
self.chain_length,
rstate0=self.rng)
# Save the current position, it will be the start point in
# the next iteration
self.p0 = pos
# Take the last samples from each walker
self.hypers = sampler.chain[:, -1]
else:
if self.hypers is None:
self.hypers = self.gp.kernel[:].tolist()
self.hypers.append(self.noise)
self.hypers = [self.hypers]
self.models = []
for sample in self.hypers:
# Instantiate a GP for each hyperparameter configuration
kernel = deepcopy(self.kernel)
kernel.set_parameter_vector(sample[:-1])
noise = np.exp(sample[-1])
model = FabolasGP(kernel,
basis_function=self.basis_func,
normalize_output=self.normalize_output,
noise=noise,
lower=self.lower,
upper=self.upper,
rng=self.rng)
model.train(X, y, do_optimize=False)
self.models.append(model)
self.is_trained = True
k2=np.std(y) / 100,
w2=0.,
e2=0.4,
offset1=0.,
offset2=0.)
kwargs = dict(**truth)
kwargs["bounds"] = dict(P1=(7.5, 8.5),
k1=(0, 0.1),
w1=(-2 * np.pi, 2 * np.pi),
e1=(0, 0.9),
tau2=(-50, 50),
k2=(0, 0.2),
w2=(-2 * np.pi, 2 * np.pi),
e2=(0, 0.9))
mean_model = Model(**kwargs)
gp = george.GP(kernel, mean=mean_model, fit_mean=True)
gp.compute(t, yerr)
def lnprob2(p):
gp.set_parameter_vector(p)
return gp.log_likelihood(y, quiet=True) + gp.log_prior()
#==============================================================================
# MCMC
#==============================================================================
import emcee
initial = gp.get_parameter_vector()
names = gp.get_parameter_names()
0.1
])
k1 = ExpSquaredKernel(p0, ndim=len(p0))
k2 = Matern32Kernel(p0, ndim=len(p0))
k3 = ConstantKernel(0.1, ndim=len(p0))
k4 = WhiteKernel(0.1, ndim=len(p0))
k5 = ConstantKernel(0.1, ndim=len(p0))
kernel = k1 * k5 + k2 + k3 + k4
ppt = pp[j]
if j == 0:
if HODLR == True:
gp0 = george.GP(kernel, mean=np.mean(y), solver=george.HODLRSolver)
else:
gp0 = george.GP(kernel, mean=np.mean(y), solver=george.BasicSolver)
gp0.compute(rr, yerr)
gp0.kernel.vector = ppt
gp0.compute(rr, yerr)
if j == 1:
if HODLR == True:
gp1 = george.GP(kernel, mean=np.mean(y), solver=george.HODLRSolver)
else:
gp1 = george.GP(kernel, mean=np.mean(y), solver=george.BasicSolver)
gp1.compute(rr, yerr)
gp1.kernel.vector = ppt
def neo_update_kernel(theta, params):
gp = george.GP(mean=0.0, fit_mean=False, white_noise=jitt)
pass
def defaultGP(theta, y, order=None, white_noise=-12, fitAmp=False):
"""
Basic utility function that initializes a simple GP with an ExpSquaredKernel.
This kernel works well in many applications as it effectively enforces a
prior on the smoothness of the function and is infinitely differentiable.
Parameters
----------
theta : array
Design points
y : array
Data to condition GP on, e.g. the lnlike + lnprior at each design point,
theta.
order : int, optional
Order of PolynomialKernel to add to ExpSquaredKernel. Defaults to None,
that is, no PolynomialKernel is added and the GP only uses the
ExpSquaredKernel
white_noise : float, optional
From george docs: "A description of the logarithm of the white noise
variance added to the diagonal of the covariance matrix". Defaults to
ln(white_noise) = -12. Note: if order is not None, you might need to
set the white_noise to a larger value for the computation to be
numerically stable, but this, as always, depends on the application.
fitAmp : bool, optional
Whether or not to include an amplitude term. Defaults to False.
Returns
-------
gp : george.GP
Gaussian process with initialized kernel and factorized covariance
matrix.
"""
# Tidy up the shapes and determine dimensionality
theta = np.asarray(theta).squeeze()
y = np.asarray(y).squeeze()
if theta.ndim <= 1:
ndim = 1
else:
ndim = theta.shape[-1]
# Guess initial metric, or scale length of the covariances (must be > 0)
initialMetric = np.fabs(np.random.randn(ndim))
# Create kernel: We'll model coveriances in loglikelihood space using a
# ndim-dimensional Squared Expoential Kernel
kernel = george.kernels.ExpSquaredKernel(metric=initialMetric,
ndim=ndim)
# Include an amplitude term?
if fitAmp:
kernel = np.var(y) * kernel
# Add a linear regression kernel of order order?
# Use a meh guess for the amplitude and for the scale length (log(gamma^2))
if order is not None:
kernel = kernel + (np.var(y)/10.0) * george.kernels.LinearKernel(log_gamma2=initialMetric[0],
order=order,
bounds=None,
ndim=ndim)
# Create GP and compute the kernel, aka factor the covariance matrix
gp = george.GP(kernel=kernel, fit_mean=True, mean=np.median(y),
white_noise=white_noise, fit_white_noise=False)
gp.compute(theta)
return gp
# Load the dataset.
data = sm.datasets.get_rdataset("co2").data
t = np.array(data.time)
y = np.array(data.co2)
# Initialize the kernel.
k1 = 66.0**2 * kernels.ExpSquaredKernel(67.0**2)
k2 = 2.4**2 * kernels.ExpSquaredKernel(90**2) \
* kernels.ExpSine2Kernel(2.0 / 1.3**2, 1.0)
k3 = 0.66**2 * kernels.RationalQuadraticKernel(0.78, 1.2**2)
k4 = 0.18**2 * kernels.ExpSquaredKernel(1.6**2) + kernels.WhiteKernel(0.19)
kernel = k1 + k2 + k3 + k4
# Set up the Gaussian process.
gp = george.GP(kernel, mean=np.mean(y))
s = time.time()
gp.compute(t)
gp.lnlikelihood(y)
print(time.time() - s)
# Define the probabilistic model.
def lnprob(p):
# Trivial prior: uniform in the log.
if np.any((-10 > p) + (p > 10)):
return -np.inf
lnprior = 0.0
# Update the kernel and compute the lnlikelihood.
# l = np.where((fc >= fmin + j * df) & (fc < fmin + (j+1) * df))[0]
# np.random.shuffle(l)
# l = l[:nperbin]
# sel[i,l] = True
sel = np.ones(flux.shape, bool)
#####################
# Rolls Royce model #
#####################
t0 = clock()
a0 = 0.71
r0 = 17.0
kernel = a0**2 * george.kernels.Matern52Kernel(r0**2)
gp = george.GP(kernel, mean=1.0)
lwav_shift = np.copy(lwav_corr)
wav_shift = np.copy(wav_corr)
ndat = sel.sum()
nseg = 1
n = ndat / nseg
while (ndat / nseg) > nmax:
nseg += 1
n = ndat / nseg
print nseg, n
def lnprob2(p):
npar = len(p)
def fit_gaussian_process(data, fix_scale=True, length_scale=20.):
"""Fit photometric observations with a gaussian regression
Args:
data (Table): Data table from sndata (format_sncosmo=True)
fix_scale (bool): Whether to fix the scale while fitting
length_scale (float): The initial length scale to use for the fit.
Returns:
A Gaussian process conditioned on the object's light-curve.
"""
if isinstance(data, Table):
data = data.to_pandas()
# Format data. Not included in original Avocado code
fluxes = data['flux']
flux_errors = data['fluxerr']
wavelengths = data['band'].map(get_effective_wavelength)
times = data['time']
# Use the highest signal-to-noise observation to estimate the scale.
# Include an error floor so that for very high
# signal-to-noise observations we pick the maximum flux value.
signal_to_noises = (np.abs(fluxes) / np.sqrt(flux_errors**2 +
(1e-2 * np.max(fluxes))**2))
scale = np.abs(fluxes[signal_to_noises.idxmax()])
# Construct gaussian process
kernel = get_kernel(scale, fix_scale, length_scale)
gp = george.GP(kernel)
# Compute the covariance matrix
x_data = np.vstack([times, wavelengths]).T
gp.compute(x_data, flux_errors)
# Define log likelihood calculations
def neg_ln_like(p):
gp.set_parameter_vector(p)
return -gp.log_likelihood(fluxes)
def grad_neg_ln_like(p):
gp.set_parameter_vector(p)
return -gp.grad_log_likelihood(fluxes)
bounds = [(0, np.log(1000**2))]
if not fix_scale:
bounds = [(-30, 30)] + bounds
fit_result = minimize(
neg_ln_like,
gp.get_parameter_vector(),
jac=grad_neg_ln_like,
bounds=bounds,
)
if not fit_result.success:
warn("GP fit failed! Using guessed GP parameters. ")
gp.set_parameter_vector(fit_result.x)
# Wrap the output function to return error instead of variance
# Not included in original Avocado code
def out_func(*args, **kwargs):
p, pv = partial(gp.predict, fluxes)(*args, **kwargs)
return p, np.sqrt(pv)
return out_func
def run(self, dataSlice, slicePoint=None):
"""
Args:
dataSlice (ndarray): Values passed to metric by the slicer,
which the metric will use to calculate metric values
at each slicePoint.
slicePoint (Dict): Dictionary of slicePoint metadata passed
to each metric.
Returns:
float: Interpolated static-probes Figure-of-Merit.
"""
# Chop off any outliers
good_pix = np.where(dataSlice[self.col] > 0)[0]
# Calculate area and med depth from
area = hp.nside2pixarea(self.nside, degrees=True) * np.size(good_pix)
median_depth = np.median(dataSlice[self.col][good_pix])
# FoM is calculated at the following values
parameters = dict(
area=np.array([
7623.22, 14786.3, 9931.47, 8585.43, 17681.8, 15126.9, 9747.99,
8335.08, 9533.42, 18331.3, 12867.8, 17418.9, 19783.1, 12538.8,
15260., 16540.7, 19636.8, 11112.7, 10385.5, 16140.2, 18920.1,
17976.2, 11352., 9214.77, 16910.7, 11995.6, 16199.8, 14395.1,
8133.86, 13510.5, 19122.3, 15684.5, 12014.8, 14059.7, 10919.3,
13212.7
]),
depth=np.array([
25.3975, 26.5907, 25.6702, 26.3726, 26.6691, 24.9882, 25.0814,
26.4247, 26.5088, 25.5596, 25.3288, 24.8035, 24.8792, 25.609,
26.2385, 25.0351, 26.7692, 26.5693, 25.8799, 26.3009, 25.5086,
25.4219, 25.8305, 26.2953, 26.0183, 25.26, 25.7903, 25.1846,
26.7264, 26.0507, 25.6996, 25.2256, 24.9383, 26.1144, 25.9464,
26.1878
]),
shear_m=np.array([
0.00891915, 0.0104498, 0.0145972, 0.0191916, 0.00450246,
0.00567828, 0.00294841, 0.00530922, 0.0118632, 0.0151849,
0.00410151, 0.0170622, 0.0197331, 0.0106615, 0.0124445,
0.00994507, 0.0136251, 0.0143491, 0.0164314, 0.016962,
0.0186608, 0.00945903, 0.0113246, 0.0155225, 0.00800846,
0.00732104, 0.00649453, 0.00243976, 0.0125932, 0.0182587,
0.00335859, 0.00682287, 0.0177269, 0.0035219, 0.00773304,
0.0134886
]),
sigma_z=np.array([
0.0849973, 0.0986032, 0.0875521, 0.0968222, 0.0225239,
0.0718278, 0.0733675, 0.0385274, 0.0425549, 0.0605867,
0.0178555, 0.0853407, 0.0124119, 0.0531027, 0.0304032,
0.0503145, 0.0132213, 0.0941765, 0.0416444, 0.0668198,
0.063227, 0.0291332, 0.0481633, 0.0595606, 0.0818742,
0.0472518, 0.0270185, 0.0767401, 0.0219945, 0.0902663,
0.0779705, 0.0337666, 0.0362358, 0.0692429, 0.0558841,
0.0150457
]),
sig_delta_z=np.array([
0.0032537, 0.00135316, 0.00168787, 0.00215043, 0.00406031,
0.00222358, 0.00334993, 0.00255186, 0.00266499, 0.00159226,
0.00183664, 0.00384965, 0.00427765, 0.00314377, 0.00456113,
0.00347868, 0.00487938, 0.00418152, 0.00469911, 0.00367598,
0.0028009, 0.00234161, 0.00194964, 0.00200982, 0.00122739,
0.00310886, 0.00275168, 0.00492736, 0.00437241, 0.00113931,
0.00104864, 0.00292328, 0.00452082, 0.00394114, 0.00150756,
0.003613
]),
sig_sigma_z=np.array([
0.00331909, 0.00529541, 0.00478151, 0.00437497, 0.00443062,
0.00486333, 0.00467423, 0.0036723, 0.00426963, 0.00515357,
0.0054553, 0.00310132, 0.00305971, 0.00406327, 0.00594293,
0.00348709, 0.00562526, 0.00396025, 0.00540537, 0.00500447,
0.00318595, 0.00460592, 0.00412137, 0.00336418, 0.00524988,
0.00390092, 0.00498349, 0.0056667, 0.0036384, 0.00455861,
0.00554822, 0.00381061, 0.0057615, 0.00357705, 0.00590572,
0.00422393
]),
FOM=np.array([
11.708, 33.778, 19.914, 19.499, 41.173, 17.942, 12.836, 26.318,
25.766, 28.453, 28.832, 14.614, 23.8, 21.51, 27.262, 20.539,
39.698, 19.342, 17.103, 25.889, 25.444, 32.048, 24.611, 23.747,
32.193, 18.862, 34.583, 14.54, 23.31, 25.812, 39.212, 25.078,
14.339, 24.12, 24.648, 29.649
]))
# Standardizing data
df_unscaled = pd.DataFrame(parameters)
X_params = [
'area', 'depth', 'shear_m', 'sigma_z', 'sig_delta_z', 'sig_sigma_z'
]
scalerX = StandardScaler()
scalerY = StandardScaler()
X = df_unscaled.drop('FOM', axis=1)
X = pd.DataFrame(scalerX.fit_transform(df_unscaled.drop('FOM',
axis=1)),
columns=X_params)
Y = pd.DataFrame(scalerY.fit_transform(
np.array(df_unscaled['FOM']).reshape(-1, 1)),
columns=['FOM'])
# Building Gaussian Process based emulator
kernel = kernels.ExpSquaredKernel(metric=[1, 1, 1, 1, 1, 1], ndim=6)
gp = george.GP(kernel, mean=Y['FOM'].mean())
gp.compute(X)
def neg_ln_lik(p):
gp.set_parameter_vector(p)
return -gp.log_likelihood(Y['FOM'])
def grad_neg_ln_like(p):
gp.set_parameter_vector(p)
return -gp.grad_log_likelihood(Y['FOM'])
result = minimize(neg_ln_lik,
gp.get_parameter_vector(),
jac=grad_neg_ln_like)
gp.set_parameter_vector(result.x)
# Survey parameters to predict FoM at
to_pred = np.array([[
area, median_depth, self.shear_m, self.sigma_z, self.sig_delta_z,
self.sig_sigma_z
]])
to_pred = scalerX.transform(to_pred)
predSFOM = gp.predict(Y['FOM'], to_pred, return_cov=False)
predFOM = scalerY.inverse_transform(list(predSFOM))
return predFOM[0]
#tic = time.time()
#args = (m,mNoisy,w,sigma,t)
#f = lnprob(initial,*args)
#print '1 exec took %0.2f' % (time.time() - tic)
# best fit noise model
#p = initial
p = res.x
m.set_all(*p[2:])
fit = m.source
mBroad = m.broaden(sigma)
tic = time.time()
a, tau = np.exp(p[:2])
k = a * kernels.Matern32Kernel(tau, ndim=2)
gp = george.GP(k)
#gp.compute(t, 1/np.sqrt(w))
gp.compute(t, mBroad.flatten() / mBroad.sum() / np.sqrt(w))
mu = gp.predict(mNoisy.flatten() - mBroad.flatten(), t, mean_only=True)
print 'one exec took %0.2fs' % (time.time() - tic)
# figures
plt.figure()
plt.imshow(mu.reshape(fit.shape))
#plt.subplot(121); plt.imshow(mTrue); plt.title('true')
#plt.subplot(122); plt.imshow(fit); plt.title('fit')
plt.figure()
vmin, vmax = 0, mNoisy.max()
plt.subplot(122)
plt.imshow(mNoisy, vmin=0, vmax=vmax)
fig, axes = pl.subplots(nrow, ncol, sharex=True, sharey=True, figsize=(12, 8))
fig.subplots_adjust(left=0.05,
right=0.95,
top=0.95,
bottom=0.05,
wspace=0.025,
hspace=0.025)
for i, ax in enumerate(axes.flatten()):
ax.set_xticklabels([])
ax.set_yticklabels([])
# The time array
time = np.linspace(0, 1, tlen)
# The intrinsic fluxes (stellar variability + a little white noise)
gp_stellar = george.GP(astr**2 * george.kernels.Matern32Kernel(tstr**2))
truths = gp_stellar.sample(time, size=nflx)
truths += awht * np.random.randn(nflx, tlen)
# The systematic components
gp_sys = george.GP(asys**2 * george.kernels.Matern32Kernel(tsys**2))
systematics = gp_sys.sample(time, size=nsys)
fluxes = np.zeros_like(truths)
for i in range(nflx):
fluxes[i] = truths[i] + np.dot(np.random.randn(nsys), systematics)
# Subtract the median
fluxes -= np.nanmedian(fluxes, axis=1).reshape(-1, 1)
# Assign constant errors (for now)
errors = awht * np.ones_like(fluxes)
s_obs = np.zeros((nsim, npix))
istart = np.random.uniform(0, dx, nsim).astype(int)
for i in np.arange(nsim):
x_obs[i, :] = x_degraded[istart[i]:istart[i] + npix * dx:dx]
y_obs[i, :] = y_degraded[istart[i]:istart[i] + npix * dx:dx]
sigma = 0.01 * np.sqrt(y_obs[i, :])
s_obs[i, :] = sigma
noise = np.random.normal(0, 1, npix) * sigma
y_obs[i, :] += noise
print 'observed spectra'
pl.plot(x_obs.T, y_obs.T, '.')
pl.draw()
raw_input('continue?')
k = 1.0 * george.kernels.ExpSquaredKernel(1.0)
gp = george.GP(k)
x = x_obs[4, :].flatten()
y = y_obs[4, :].flatten()
s = s_obs[4, :].flatten()
ss = np.argsort(x)
x = x[ss]
y = y[ss]
s = s[ss]
gp.compute(x, yerr=s)
gp.optimize(x, y, yerr=s)
print np.sqrt(gp.kernel.pars)
gp.compute(x, yerr=s)
mu, cov = gp.predict(y, x_true)
err = np.sqrt(np.diag(cov))
print 'conditioned on one observation only'
pl.plot(x_true, mu, 'r-')
plt.plot(freqs[1], split_vals_plot[1], label='Real', color='blue')
#plt.plot(freqs[1],mean_model.get_value_plot(None)[1], label = 'First Guess',color = 'red', linestyle = '-.')
plt.plot(freqs[2], split_vals_plot[2], label='Real', color='green')
#plt.plot(freqs[2],mean_model.get_value_plot(None)[2], label = 'First Guess', color = 'green', linestyle = '-.')
plt.legend(loc='best')
plt.title('Actual Rotation Profile splitting values')
plt.show()
#gp = george.GP(0.001 * kernels.ExpSquaredKernel(metric=([1,0],[0,1]), metric_bounds = [(0,7),(None,None),(-5,-2)], ndim = 2), mean=mean_model)
#gp = george.GP(1 * kernels.ExpSquaredKernel(metric=np.array([[1,0],[0,100]]), metric_bounds=[(-5, 5), (None, None), (-5, -2.3)], ndim=2), mean=mean_model)
gp = george.GP(1 * kernels.ExpSquaredKernel(metric=1e-3 * np.eye(2),
ndim=2,
metric_bounds=[(-7, 5),
(None, None),
(-5, -1.5)]),
mean=mean_model)
#gp = george.GP(1 * kernels.ExpSquaredKernel(metric=1e-5 * np.eye(2), ndim=2, metric_bounds=[(-10, 10), (None, None), (-6, -2.3)]))
gp.set_parameter('kernel:k2:metric:L_0_1', 0)
gp.freeze_parameter('kernel:k2:metric:L_0_1')
#this is the part where im confused
#gp.compute(flatx)
gp.compute(xvals, e_split_vals)
def log_prior(p):
xsinx = lambda x: -x*np.sin(x)
# Plot and show the function.
t = np.linspace(0, 10, 500) # Return 500 evenly spaced numbers over range [0, 10]
plt.plot(t, xsinx(t))
plt.ylabel("y")
plt.xlabel("x")
plt.show()
#
# Gaussian Process (with zero knowledge).
#
# Set up a gaussian process. The main ingredient of a gaussian process is the
# kernel. It describes how correlated each point is with every other.
kernel = ExpSquaredKernel(1.0)
gp = george.GP(kernel)
# Here we draw plot to show what kind of function we have when the gaussian
# process have no knowledge of our objective value of our function. We draw
# 10 samples below:
fig, ax = plt.subplots()
for pred in gp.sample(t, 10):
ax.plot(t, pred)
ax.set_ylim(-9, 7)
ax.set_xlabel("x")
ax.set_ylabel("y")
plt.show()
# If we choose a different kernel, the functions we sample will end up looking
# different.
matern_kernel = Matern32Kernel(1.0)
print("")
x = np.array(x)
y_fof2 = np.array(y_fof2)
mean_fof2 = np.mean(y_fof2)
y_fof2 -= mean_fof2
y_mufd = np.array(y_mufd)
mean_mufd = np.mean(y_mufd)
y_mufd -= mean_mufd
y_hmf2 = np.array(y_hmf2)
mean_hmf2 = np.mean(y_hmf2)
y_hmf2 -= mean_hmf2
sigma = np.array(sigma)
gp = george.GP(kernel)
gp.compute(x, sigma + 1e-3)
tm = np.array(range(last_tm - (86400 * 2), last_tm + (86400 * 7) + 1, 300))
pred_fof2, sd_fof2 = gp.predict(y_fof2, tm / 86400., return_var=True)
pred_fof2 += mean_fof2
sd_fof2 = np.sqrt(sd_fof2)
pred_mufd, sd_mufd = gp.predict(y_mufd, tm / 86400., return_var=True)
pred_mufd += mean_mufd
sd_mufd = np.sqrt(sd_mufd)
pred_hmf2, sd_hmf2 = gp.predict(y_hmf2, tm / 86400., return_var=True)
pred_hmf2 += mean_hmf2
sd_hmf2 = np.sqrt(sd_hmf2)
for i in range(len(tm)):
def create_testcase06():
import george
bjd0 = photometry['phot_bjd'] - Tref
err = bjd0 * 0 + photometry['phot_precision']
""" Conversion of the physical parameter to the internally defined parameter
to be passed to george
"""
gp_pams = np.zeros(4)
gp_pams[0] = np.log(activity['Hamp_PH']) * 2
gp_pams[1] = np.log(activity['Pdec']) * 2
gp_pams[2] = 1. / (2 * activity['Oamp']**2)
gp_pams[3] = np.log(activity['Prot'])
kernel = np.exp(gp_pams[0]) * \
george.kernels.ExpSquaredKernel(metric=np.exp(gp_pams[1])) * \
george.kernels.ExpSine2Kernel(gamma=gp_pams[2], log_period=gp_pams[3])
gp = george.GP(kernel)
gp.compute(bjd0, err)
prediction = gp.sample(bjd0)
obs_photometry = np.random.normal(prediction, photometry['phot_precision'])
fileout = open('TestCase06_photometry.dat', 'w')
for b, p in zip(photometry['phot_bjd'], obs_photometry):
fileout.write('{0:14f} {1:14f} {2:14f} {3:5d} {4:5d} {5:5d} \n'.format(
b, p, photometry['phot_precision'], 0, 0, -1))
fileout.close()
bjd0 = bjd_obs - Tref
err = bjd0 * 0 + instrument['RV_precision']
gp_pams[0] = np.log(activity['Hamp_RV1']) * 2
kernel = np.exp(gp_pams[0]) * \
george.kernels.ExpSquaredKernel(metric=np.exp(gp_pams[1])) * \
george.kernels.ExpSine2Kernel(gamma=gp_pams[2], log_period=gp_pams[3])
gp = george.GP(kernel)
gp.compute(bjd0, err)
prediction = gp.sample(bjd0)
y_pla = kp.kepler_RV_T0P(bjd0, planet_b['f'], planet_b['P'], planet_b['K'],
planet_b['e'],
planet_b['o']) + instrument['RV_offset1']
mod_pl = np.random.normal(y_pla + prediction, instrument['RV_precision'])
Tcent_b = np.random.normal(
np.arange(0, 1) * planet_b['P'] + kp.kepler_Tcent_T0P(
planet_b['P'], planet_b['f'], planet_b['e'], planet_b['o']) + Tref,
instrument['T0_precision'])
fileout = open('TestCase06_Tcent_b.dat', 'w')
for i_Tc, v_Tc in enumerate(Tcent_b):
fileout.write('{0:d} {1:.4f} {2:.4f} {3:d}\n'.format(
i_Tc, v_Tc, instrument['T0_precision'], 0))
fileout.close()
fileout = open('TestCase06_RV.dat', 'w')
for b, r in zip(bjd_obs, mod_pl):
fileout.write('{0:f} {1:.2f} {2:.2f} {3:d} {4:d} {5:d}\n'.format(
b, r, instrument['RV_precision'], 0, 0, -1))
fileout.close()
def gp_train(kernel, lctrain):
gp = george.GP(kernel)
gp.compute(lctrain.x, lctrain.yerr**2)
return gp
#y_ = 1 + d
import george
from george import kernels
t = t_[use_for_pred]
y = y_[use_for_pred]
var_y = np.var(y)
k1 = var_y * kernels.ExpSquaredKernel(metric=1)
k2 = var_y * kernels.ExpSquaredKernel(metric=1) \
* kernels.ExpSine2Kernel(gamma=100, log_period=0.0)
kernel = k1 + k2
#kernel = k1
gp = george.GP(kernel, white_noise=np.log(var_y), fit_white_noise=True)
gp.compute(t)
pn = 'kernel:k2:k2:log_period'
if pn in gp.get_parameter_names():
gp.freeze_parameter(pn)
#gp.freeze_parameter('kernel:k2:k2:log_period')
import scipy.optimize as op
def nll(p):
gp.set_parameter_vector(p)
ll = gp.log_likelihood(y, quiet=True)
return -ll if np.isfinite(ll) else 1e25
# For each bin, make GP predictions of the bin centroid.
parameter_names = list(results[f"models/{model_name}/gp_model"].keys())[2:]
gps = {}
for parameter_name in parameter_names:
X = results[f"models/{model_name}/gp_model/{parameter_name}/X"][()]
Y = results[f"models/{model_name}/gp_model/{parameter_name}/Y"][()]
attrs = results[f"models/{model_name}/gp_model/{parameter_name}"].attrs
metric = np.var(X, axis=0)
kernel = george.kernels.ExpSquaredKernel(metric=metric, ndim=metric.size)
gp = george.GP(kernel,
mean=np.mean(Y), fit_mean=True,
white_noise=np.log(np.std(Y)), fit_white_noise=True)
for p in gp.parameter_names:
gp.set_parameter(p, attrs[p])
gp.compute(X)
gps[parameter_name] = (gp, Y)
# Now make predictions at each centroid.
mg = np.array(np.meshgrid(*centroids)).reshape((len(lns), -1)).T
gp_predictions = np.empty((mg.shape[0], 4, 2))
gp_predictions[:] = np.nan
for i, parameter_name in enumerate(parameter_names):
d_harps1=0.,
d_harps2=0.)
kwargs = dict(**truth)
kwargs["bounds"] = dict(P1=(12.0, 13.0),
k1=(0, 1.),
w1=(-2 * np.pi, 2 * np.pi),
e1=(0.7, 0.95),
P2=(35, 200),
k2=(0, 2.),
w2=(-2 * np.pi, 2 * np.pi),
e2=(0., 0.5))
mean_model = Model(**kwargs)
gp = george.GP(kernel,
mean=mean_model,
fit_mean=True,
white_noise=np.log(0.5**2),
fit_white_noise=True)
# gp.freeze_parameter('kernel:k2:k1:log_constant')
gp.compute(x, yerr)
lnp1 = gp.log_likelihood(y)
def lnprob2(p):
gp.set_parameter_vector(p)
return gp.log_likelihood(y, quiet=True) + gp.log_prior()
#==============================================================================
# MCMC
#==============================================================================
# Other inputs:
n_live_points = int(args.nlive)
jitter = george.modeling.ConstantModel(np.log((200. * 1e-6)**2.))
# create dictionaries to store kernels and gp classes for individual nights
kernels = {}
gps = {}
for key in nights:
kernels[key] = np.var(nights[key][1]) * george.kernels.ExpSquaredKernel(
np.ones(nights[key][2].shape[0]),
ndim=(nights[key][2].shape[0]),
axes=range(nights[key][2].shape[0]))
# Wrap GP object to compute likelihood
gps[key] = george.GP(kernels[key],
mean=0.0,
fit_mean=False,
white_noise=jitter,
fit_white_noise=True)
gps[key].compute(nights[key][2].T)
# Now define MultiNest priors and log-likelihood:
def prior(cube):
# Prior on "median flux" is uniform:
cube[0] = utils.transform_uniform(cube[0], -2., 2.)
# Pior on the log-jitter term (note this is the log VARIANCE, not sigma); from 0.01 to 100 ppm:
cube[1] = utils.transform_uniform(cube[1], np.log((0.01e-3)**2),
np.log((100e-3)**2))
pcounter = 2
# Prior on coefficients of comparison stars:
if compfilename is not None:
def train(self, X, y, do_optimize=True, **kwargs):
"""
Performs MCMC sampling to sample hyperparameter configurations from the
likelihood and trains for each sample a GP on X and y
Parameters
----------
X: np.ndarray (N, D)
Input data points. The dimensionality of X is (N, D),
with N as the number of points and D is the number of features.
y: np.ndarray (N,)
The corresponding target values.
do_optimize: boolean
If set to true we perform MCMC sampling otherwise we just use the
hyperparameter specified in the kernel.
"""
if self.normalize_input:
# Normalize input to be in [0, 1]
self.X, self.lower, self.upper = normalization.zero_one_normalization(X, self.lower, self.upper)
else:
self.X = X
if self.normalize_output:
# Normalize output to have zero mean and unit standard deviation
self.y, self.y_mean, self.y_std = normalization.zero_mean_unit_var_normalization(y)
if self.y_std == 0:
raise ValueError("Cannot normalize output. All targets have the same value")
else:
self.y = y
# Use the mean of the data as mean for the GP
self.mean = np.mean(self.y, axis=0)
self.gp = george.GP(self.kernel, mean=self.mean)
if do_optimize:
# We have one walker for each hyperparameter configuration
sampler = emcee.EnsembleSampler(self.n_hypers,
len(self.kernel.pars) + 1,
self.loglikelihood)
sampler.random_state = self.rng.get_state()
# Do a burn-in in the first iteration
if not self.burned:
# Initialize the walkers by sampling from the prior
if self.prior is None:
self.p0 = self.rng.rand(self.n_hypers, len(self.kernel.pars) + 1)
else:
self.p0 = self.prior.sample_from_prior(self.n_hypers)
# Run MCMC sampling
self.p0, _, _ = sampler.run_mcmc(self.p0,
self.burnin_steps,
rstate0=self.rng)
self.burned = True
# Start sampling
pos, _, _ = sampler.run_mcmc(self.p0,
self.chain_length,
rstate0=self.rng)
# Save the current position, it will be the start point in
# the next iteration
self.p0 = pos
# Take the last samples from each walker
self.hypers = sampler.chain[:, -1]
else:
self.hypers = self.gp.kernel[:].tolist()
self.hypers.append(self.noise)
self.hypers = [self.hypers]
self.models = []
for sample in self.hypers:
# Instantiate a GP for each hyperparameter configuration
kernel = deepcopy(self.kernel)
kernel.pars = np.exp(sample[:-1])
noise = np.exp(sample[-1])
model = GaussianProcess(kernel,
normalize_output=self.normalize_output,
normalize_input=self.normalize_input,
noise=noise,
lower=self.lower,
upper=self.upper,
rng=self.rng)
model.train(X, y, do_optimize=False)
self.models.append(model)
self.is_trained = True
def PLD_then_GP(q, id, kernel, kinit, kbounds, maxfun, datadir):
'''
'''
global data
if data is None:
data = GetData(id, data_type='bkg', datadir=datadir)
qd = data[q]
if qd['time'] == []:
return None
# t/y for all chunks
t_all = np.array([], dtype=float)
y_all = np.array([], dtype=float)
# Solve the (linear) PLD problem for this quarter
fpix = np.array([x for y in qd['fpix'] for x in y])
fsum = np.sum(fpix, axis=1)
init = np.array([np.median(fsum)] * fpix.shape[1])
def pm(y, *x):
return np.sum(fpix * np.outer(1. / fsum, x), axis=1)
x, _ = curve_fit(pm, None, fsum, p0=init)
# Here's our detrended data
y = []
t = []
e = []
for time, fpix, perr in zip(qd['time'], qd['fpix'], qd['perr']):
# The pixel model
fsum = np.sum(fpix, axis=1)
pixmod = np.sum(fpix * np.outer(1. / fsum, x), axis=1)
# The errors
X = np.ones_like(time) + pixmod / fsum
B = X.reshape(len(time), 1) * perr - x * perr / fsum.reshape(
len(time), 1)
yerr = np.sum(B**2, axis=1)**0.5
# Append to our arrays
y.append(fsum - pixmod)
t.append(time)
e.append(yerr)
# Now we solve for the best GP params
res = fmin_l_bfgs_b(NegLnLikeGP,
kinit,
approx_grad=False,
args=(t, y, e, kernel),
bounds=kbounds,
m=10,
factr=1.e1,
pgtol=1e-05,
maxfun=maxfun)
# Finally, detrend the data
kernel.pars = res[0]
gp = george.GP(kernel)
for ti, yi, ei in zip(t, y, e):
gp.compute(ti, ei)
mu, _ = gp.predict(yi, ti)
y_all = np.append(y_all, yi - mu)
t_all = np.append(t_all, ti)
return t_all, y_all
def run_vs_SearchPhase_UA2Fit():
args = parse_args()
ext = args.output_file_extension
from pygp.canvas import Canvas
# Getting data points
xRaw, yRaw, xerrRaw, yerrRaw = getDataPoints(args.input_file,
'dijetgamma_g85_2j65',
'Zprime_mjj_var')
#Data processing, cutting out the not desired range
x, y, xerr, yerr = dataCut(0, 1500, 0, xRaw, yRaw, xerrRaw, yerrRaw)
xMinFit = 303
xMaxFit = 1500
xFit, yFit, xerrFit, yerrFit = dataCut(xMinFit, xMaxFit, 0, xRaw, yRaw,
xerrRaw, yerrRaw)
# make an evently spaced x
t = np.linspace(np.min(x), np.max(x), 500)
#calculating the log-likihood and minimizing for the gp
lnProb = logLike_minuit(x, y, xerr)
min_likelihood, best_fit = fit_gp_minuit(20, lnProb)
fit_pars = [best_fit[x] for x in FIT3_PARS]
#making the GP
kargs = {x: y for x, y in best_fit.items() if x not in FIT3_PARS}
kernel_new = get_kernel(**kargs)
print(kernel_new.get_parameter_names())
#making the kernel
gp_new = george.GP(kernel_new, mean=Mean(fit_pars), fit_mean=True)
gp_new.compute(x, yerr)
mu, cov = gp_new.predict(y, t)
mu_x, cov_x = gp_new.predict(y, xFit)
# calculating the fit function value
#GP compute minimizes the log likelihood of the best_fit function
best = [best_fit[x] for x in FIT3_PARS]
print("best param meghan GP minmization:", best)
meanFromGPFit = Mean(best)
fit_meanM = meanFromGPFit.get_value(
x, xerr) #so this is currently shit. can't get the xErr thing working
print("fit_meanM:", fit_meanM)
#fit results from search phase
best_fit_params = (1.0779, 2.04035, 58.5769, -157.945)
fit_mean = model_UA2(xFit, best_fit_params, xerrFit)
##----4 param fit function
#lnProb = logLike_4ff(xFit,yFit,xerrFit)
#minimumLLH, best_fit_params = fit_4ff(100, lnProb)
#fit_mean = model_4param(xFit, best_fit_params, xerrFit)
#calculating significance
signif = significance(mu_x, y, cov_x, yerr)
initialCutPos = np.argmax(x > xMinFit)
finalCutPos = np.argmin(x < xMaxFit)
#sigFit = (fit_mean - y[initialCutPos:]) / np.sqrt(np.diag(cov_x[initialCutPos:]) + yerr[initialCutPos:]**2)
sigFit = significance(fit_mean, y[initialCutPos:], cov_x[initialCutPos:],
yerr[initialCutPos:])
#sigit = (mu_x)
#std = np.sqrt(np.diag(cov))
ext = args.output_file_extension
title = "compareGPvsSearchPhaseUA2"
with Canvas(f'%s{ext}' % title) as can:
can.ax.errorbar(x, y, yerr=yerr, fmt='.')
can.ax.set_yscale('log')
# can.ax.set_ylim(1, can.ax.get_ylim()[1])
can.ax.plot(t, mu, '-r')
can.ax.plot(xFit, fit_mean, '-b')
#this only works with the xErr part of Mean commented out
#can.ax.plot(x, fit_meanM, '.g')
# can.ax.plot(t, fit_mean_smooth, '--b')
#can.ax.fill_between(t, mu - std, mu + std,
#facecolor=(0, 1, 0, 0.5),
#zorder=5, label='err = 1')
#can.ratio.stem(x, signif, markerfmt='.', basefmt=' ')
can.ratio.stem(xFit, sigFit, markerfmt='.', basefmt=' ')
can.ratio.axhline(0, linewidth=1, alpha=0.5)
can.save(title)
def predict_Q_data(input_dir_2, filename4="training.txt"):
nu, lam_squared, stokesQ, stokesU = read_data(input_dir_2, filename4)
nu_R = nu
stokesQ_R = stokesQ
lam_R = lam_squared
# Squared exponential kernel
k1 = 0.3**2 * kernels.ExpSquaredKernel(0.02**2)
# periodic covariance kernel with exponential component toallow decay away from periodicity
k2 = 0.6**2 * kernels.ExpSquaredKernel(0.5**2) * kernels.ExpSine2Kernel(
gamma=2 / 2.5**2, log_period=0.0)
###vary gamma value to widen the funnel
# rational quadratic kernel for medium term irregularities.
k3 = 0.3**2 * kernels.RationalQuadraticKernel(log_alpha=np.log(0.1),
metric=0.1**2)
# noise kernel: includes correlated noise & uncorrelated noise
k4 = 0.18**2 * kernels.ExpSquaredKernel(1.6**2)
kernel = k1 + k2 #+k3 + k4
gp = george.GP(kernel,
mean=np.mean(stokesQ),
fit_mean=True,
white_noise=np.log(0.02**2),
fit_white_noise=True)
#gp = george.GP(kernel)
gp.compute(lam_R)
# range of times for prediction:
#x = np.linspace(0.18, 0.21998, 360)
x = x_values(input_dir_2)
# calculate expectation and variance at each point:
mu_initial, cov = gp.predict(stokesQ_R, x)
#mu, cov = gp.predict(stokesQ_2Gz, x, return_var = True)
std_initial = np.sqrt(np.diag(cov))
print("Initial prediction of missing Q values: \n {}".format(mu_initial))
# Define the objective function (negative log-likelihood in this case).
def nll(p):
gp.set_parameter_vector(p)
ll = gp.log_likelihood(stokesQ_R, quiet=True)
return -ll if np.isfinite(ll) else 1e25
# And the gradient of the objective function.
def grad_nll(p):
gp.set_parameter_vector(p)
return -gp.grad_log_likelihood(stokesQ_R, quiet=True)
gp.compute(lam_R)
p0 = gp.get_parameter_vector()
results = op.minimize(nll, p0, jac=grad_nll, method="L-BFGS-B")
# run optimization:
#results = op.minimize(nll, p0, jac=grad_nll)
gp.set_parameter_vector(results.x)
x = x_values(input_dir_2)
mu_optimized, cov = gp.predict(stokesQ_R, x)
std_optimized = np.sqrt(np.diag(cov))
print("Final prediction of missing Q values: \n {}".format(mu_optimized))
return nu_R, lam_R, stokesQ_R, mu_initial, std_initial, mu_optimized, std_optimized
