#
# Copyright (C) 2011, 2015, 2016 Smithsonian Astrophysical Observatory
#
#
# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License along
# with this program; if not, write to the Free Software Foundation, Inc.,
# 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
#
from six.moves import zip as izip
import numpy
import numpy.random
from sherpa.estmethods import Covariance, Confidence
from sherpa.utils.err import EstErr
from sherpa.utils import parallel_map, NoNewAttributesAfterInit
import logging
warning = logging.getLogger("sherpa").warning
__all__ = ['multivariate_t', 'multivariate_cauchy',
'normal_sample', 'uniform_sample', 't_sample',
'ParameterScaleVector', 'ParameterScaleMatrix',
'UniformParameterSampleFromScaleVector',
'NormalParameterSampleFromScaleVector',
'NormalParameterSampleFromScaleMatrix',
'StudentTParameterSampleFromScaleMatrix',
'NormalSampleFromScaleMatrix', 'NormalSampleFromScaleVector',
'UniformSampleFromScaleVector', 'StudentTSampleFromScaleMatrix',
]
[docs]def multivariate_t(mean, cov, df, size=None):
"""
multivariate_t(mean, cov, df[, size])
Draw random deviates from a multivariate Student's T distribution Such a
distribution is specified by its mean covariance matrix, and degrees of
freedom. These parameters are analogous to the mean (average or "center"),
variance (standard deviation, or "width," squared), and the degrees of
freedom of the one-dimensional t distribution.
Parameters
----------
mean : 1-D array_like, length N
Mean of the N-dimensional distribution
cov : 2-D array_like, shape (N, N)
Covariate matrix of the distribution. Must be symmetric and
positive semi-definite for "physically meaningful" results.
df : int
Degrees of freedom of the distribution
size : tuple of ints, optional
Given a shape of, for example, ``(m,n,k)``, ``m*n*k`` samples are
generated, and packed in an `m`-by-`n`-by-`k` arrangement. Because
each sample is `N`-dimensional, the output shape is ``(m,n,k,N)``.
If no shape is specified, a single (`N`-D) sample is returned.
Returns
-------
out : ndarray
The drawn samples, of shape *size*, if that was provided. If not,
the shape is ``(N,)``.
In other words, each entry ``out[i,j,...,:]`` is an N-dimensional
value drawn from the distribution.
Is this right? This needs to be checked! A reference to the literature
the better
"""
df = float(df)
mean = numpy.asarray(mean)
normal = numpy.random.multivariate_normal(
numpy.zeros_like(mean), cov, size)
# x = numpy.sqrt(numpy.random.chisquare(df)/df)
# numpy.divide(normal, x, normal)
x = numpy.sqrt(numpy.random.chisquare(df, size) / df)
numpy.divide(normal, x[numpy.newaxis].T, normal)
numpy.add(mean, normal, normal)
x = normal
return x
[docs]def multivariate_cauchy(mean, cov, size=None):
"""
This needs to be checked too! A reference to the literature the better
"""
return multivariate_t(mean, cov, 1, size=None)
[docs]class ParameterScale(NoNewAttributesAfterInit):
sigma = 1
[docs] def get_scales(self, fit, myscale=None):
raise NotImplementedError
[docs]class ParameterScaleVector(ParameterScale):
[docs] def get_scales(self, fit, myscales=None):
scales = []
thawedpars = [par for par in fit.model.pars if not par.frozen]
if None == myscales:
oldestmethod = fit.estmethod
covar = Covariance()
covar.config['sigma'] = self.sigma
fit.estmethod = Covariance()
try:
r = fit.est_errors()
finally:
fit.estmethod = oldestmethod
for par, val, lo, hi in izip(thawedpars, r.parvals, r.parmins, r.parmaxes):
scale = None
if lo is not None and hi is not None:
scale = numpy.abs(lo)
else:
warning("Covariance failed for '%s', trying Confidence..." %
par.fullname)
conf = Confidence()
conf.config['sigma'] = self.sigma
fit.estmethod = conf
try:
t = fit.est_errors(parlist=(par,))
if t.parmins[0] is not None and t.parmaxes[0] is not None:
scale = numpy.abs(t.parmins[0])
else:
if t.parmins[0] is None and t.parmaxes[0] is not None:
scale = numpy.abs(t.parmaxes[0])
else:
warning('1 sigma bounds for parameter ' +
par.fullname +
' could not be found, using soft limit minimum')
if 0.0 == numpy.abs(par.min):
scale = 1.0e-16
else:
scale = numpy.abs(par.min)
finally:
fit.estmethod = oldestmethod
scales.append(scale)
else:
if not numpy.iterable(myscales):
raise TypeError(
"scales option must be iterable of length %d " % len(thawedpars))
scales = list(map(abs, myscales))
scales = numpy.asarray(scales).transpose()
return scales
[docs]class ParameterScaleMatrix(ParameterScale):
[docs] def get_scales(self, fit, myscales=None):
def get_size_of_covar(pars):
thawedpars = [par for par in pars if not par.frozen]
npar = len(thawedpars)
msg = 'scales must be a numpy array of size (%d,%d)' % (npar, npar)
return npar, msg
if myscales is None:
oldestmethod = fit.estmethod
fit.estmethod = Covariance()
try:
r = fit.est_errors()
finally:
fit.estmethod = oldestmethod
cov = r.extra_output
else:
if isinstance(myscales, (numpy.ndarray)):
npar, msg = get_size_of_covar(fit.model.pars)
if (npar, npar) == myscales.shape:
cov = myscales
else:
raise EstErr(msg)
else:
npar, msg = get_size_of_covar(fit.model.pars)
raise EstErr(msg)
cov = myscales
if cov is None:
raise EstErr('nocov')
cov = numpy.asarray(cov)
# Investigate spectral decomposition to avoid requirement that the cov be
# semi-positive definite. Nevermind, NumPy already uses SVD to generate
# deviates from a multivariate normal. An alternative is to use Cholesky
# decomposition, but it assumes that the matrix is semi-positive
# definite.
if numpy.min(numpy.linalg.eigvalsh(cov)) <= 0:
raise TypeError("The covariance matrix is not positive definite")
return cov
[docs]class ParameterSampleFromScaleVector(NoNewAttributesAfterInit):
def __init__(self):
self.scale = ParameterScaleVector()
NoNewAttributesAfterInit.__init__(self)
[docs] def get_sample(self):
raise NotImplementedError
[docs]class ParameterSampleFromScaleMatrix(NoNewAttributesAfterInit):
def __init__(self):
self.scale = ParameterScaleMatrix()
NoNewAttributesAfterInit.__init__(self)
[docs] def get_sample(self):
raise NotImplementedError
[docs]class NormalParameterSampleFromScaleVector(ParameterSampleFromScaleVector):
[docs] def get_sample(self, fit, myscales=None, num=1):
vals = numpy.array(fit.model.thawedpars)
scales = self.scale.get_scales(fit, myscales)
samples = [numpy.random.normal(
val, scale, int(num)) for val, scale in izip(vals, scales)]
return numpy.asarray(samples).T
[docs]class NormalParameterSampleFromScaleMatrix(ParameterSampleFromScaleMatrix):
[docs] def get_sample(self, fit, mycov=None, num=1):
vals = numpy.array(fit.model.thawedpars)
cov = self.scale.get_scales(fit, mycov)
return numpy.random.multivariate_normal(vals, cov, int(num))
[docs]class StudentTParameterSampleFromScaleMatrix(ParameterSampleFromScaleMatrix):
[docs] def get_sample(self, fit, dof, num=1):
vals = numpy.array(fit.model.thawedpars)
cov = self.scale.get_scales(fit)
return multivariate_t(vals, cov, dof, int(num))
class Evaluate(object):
"""
Callable class for _sample_stat multiprocessing call
This class used to be a nested function, which can't be pickled and results in
python3 failing to execute the code.
"""
def __init__(self, fit):
self.fit = fit
def __call__(self, sample):
self.fit.model.thawedpars = sample
return self.fit.calc_stat()
def _sample_stat(fit, samples, numcores=None):
oldvals = fit.model.thawedpars
try:
fit.model.startup()
stats = numpy.asarray(parallel_map(Evaluate(fit), samples, numcores))
finally:
fit.model.teardown()
fit.model.thawedpars = oldvals
return numpy.concatenate([stats[:, numpy.newaxis], samples], axis=1)
[docs]class NormalSampleFromScaleMatrix(NormalParameterSampleFromScaleMatrix):
[docs] def get_sample(self, fit, num=1, numcores=None):
samples = NormalParameterSampleFromScaleMatrix.get_sample(
self, fit, num=num)
return _sample_stat(fit, samples, numcores)
[docs]class NormalSampleFromScaleVector(NormalParameterSampleFromScaleVector):
[docs] def get_sample(self, fit, num=1, numcores=None):
samples = NormalParameterSampleFromScaleVector.get_sample(
self, fit, num=num)
return _sample_stat(fit, samples, numcores)
[docs]class StudentTSampleFromScaleMatrix(StudentTParameterSampleFromScaleMatrix):
[docs] def get_sample(self, fit, num=1, dof=2, numcores=None):
samples = StudentTParameterSampleFromScaleMatrix.get_sample(
self, fit, dof, num)
return _sample_stat(fit, samples, numcores)
[docs]def normal_sample(fit, num=1, sigma=1, correlate=True, numcores=None):
"""Sample the fit statistic by taking the parameter values
from a normal distribution.
For each iteration (sample), change the thawed parameters by
drawing values from a uni- or multi-variate normal (Gaussian)
distribution, and calculate the fit statistic.
Parameters
----------
fit :
The fit results.
num : int, optional
The number of samples to use (default is `1`).
sigma : number, optional
The width of the normal distribution (the default
is `1`).
correlate : bool, optional
Should a multi-variate normal be used, with parameters
set by the covariance matrix (`True`) or should a
uni-variate normal be used (`False`)?
numcores : optional
The number of CPU cores to use. The default is to use all
the cores on the machine.
Returns
-------
samples :
A NumPy array table with the first column representing the
statistic and later columns the parameters used.
See Also
--------
t_sample : Sample from the Student's t-distribution.
uniform_sample : Sample from a uniform distribution.
Notes
-----
All thawed model parameters are sampled from the Gaussian
distribution, where the mean is set as the best-fit parameter
value and the variance is determined by the diagonal elements
of the covariance matrix. The multi-variate Gaussian is
assumed by default for correlated parameters, using the
off-diagonal elements of the covariance matrix.
"""
sampler = NormalSampleFromScaleVector()
sampler.scale.sigma = sigma
if correlate:
sampler = NormalSampleFromScaleMatrix()
return sampler.get_sample(fit, num, numcores)
[docs]def t_sample(fit, num=1, dof=2, numcores=None):
"""Sample the fit statistic by taking the parameter values from
a Student's t-distribution.
For each iteration (sample), change the thawed parameters
by drawing values from a Student's t-distribution, and
calculate the fit statistic.
Parameters
----------
fit :
The fit results.
num : int, optional
The number of samples to use (default is `1`).
dof : optional
The number of degrees of freedom to use (the default
is to use the number from the current fit).
numcores : optional
The number of CPU cores to use. The default is to use all
the cores on the machine.
Returns
-------
samples :
A NumPy array table with the first column representing the
statistic and later columns the parameters used.
See Also
--------
normal_sample : Sample from the normal distribution.
uniform_sample : Sample from a uniform distribution.
"""
sampler = StudentTSampleFromScaleMatrix()
return sampler.get_sample(fit, num, dof, numcores)