def test_list_samplers():
"""Ensure list_samplers returns a list."""
mcmc = sim.MCMC()
samplers = mcmc.list_samplers()
assert isinstance(samplers, list)
assert len(samplers) > 0
def test_list_samplers_contents():
"""Are the expected values included"""
# Test that the expected values exist in this list,
# but do not enforce these are the only values. This is
# a slightly-different return list to the non-astro version.
#
samplers = sim.MCMC().list_samplers()
for expected in ['mh', 'metropolismh', 'pragbayes', 'fullbayes']:
assert expected in samplers
def test_pragbayes_simarf(setup):
fit = setup['fit']
mcmc = sim.MCMC()
mcmc.set_sampler('PragBayes')
mcmc.set_sampler_opt("simarf", setup['simarf'])
mcmc.set_sampler_opt("p_M", 0.5)
mcmc.set_sampler_opt("nsubiter", 7)
covar_results = fit.est_errors()
cov = covar_results.extra_output
niter = setup['niter']
stats, accept, params = mcmc.get_draws(fit, cov, niter=niter)
assert params.shape == (3, niter + 1)
def test_fullbayes_simarf_fails(setup):
fit = setup['fit']
mcmc = sim.MCMC()
mcmc.set_sampler('FullBAYes')
mcmc.set_sampler_opt("simarf", setup['simarf'])
mcmc.set_sampler_opt("p_M", 0.5)
mcmc.set_sampler_opt("nsubiter", 7)
covar_results = fit.est_errors()
cov = covar_results.extra_output
niter = setup['niter']
with pytest.raises(TypeError) as exc:
mcmc.get_draws(fit, cov, niter=niter)
assert str(exc.value) == 'Simulation ARF must be PCA for FullBayes not SIM1DAdd'
def test_pragbayes_pcaarf(self):
mcmc = sim.MCMC()
self.abs1.nh = 0.092886
self.p1.phoindex = 0.994544
self.p1.norm = 9.26369
mcmc.set_sampler("pragBayes")
mcmc.set_sampler_opt("simarf", self.pcaarf)
mcmc.set_sampler_opt("p_M", 0.5)
mcmc.set_sampler_opt("nsubiter", 5)
covar_results = self.fit.est_errors()
cov = covar_results.extra_output
niter = 10
stats, accept, params = mcmc.get_draws(self.fit, cov, niter=niter)
def test_pragbayes_simarf(self):
datadir = SherpaTestCase.datadir
if datadir is None:
return
mcmc = sim.MCMC()
self.abs1.nh = 0.092886
self.p1.phoindex = 0.994544
self.p1.norm = 9.26369
mcmc.set_sampler("PragBayes")
mcmc.set_sampler_opt("simarf", self.simarf)
mcmc.set_sampler_opt("p_M", 0.5)
mcmc.set_sampler_opt("nsubiter", 7)
covar_results = self.fit.est_errors()
cov = covar_results.extra_output
niter = 10
stats, accept, params = mcmc.get_draws(self.fit, cov, niter=niter)
def test_pragbayes_pcaarf_limits(sampler, setup, caplog, reset_seed):
"""Try and trigger limit issues.
"""
from sherpa.astro.xspec import XSAdditiveModel, XSMultiplicativeModel, \
XSwabs, XSpowerlaw
# Set the seed for the RNG. The seed was adjusted to try and make
# sure the coverage was "good" (i.e. hits parts of
# sherpa/astro/sim/*bayes.py) whilst still passing the test and
# reducing the runtime. This is not a guarantee that this is the
# "fastest" seed, just that it's one of the better ones I've seen.
#
np.random.seed(0x723c)
class HackAbs(XSwabs):
"""Restrict hard limits"""
def __init__(self, name='wabs'):
self.nH = Parameter(name, 'nH', 0.1, 0, 1, 0, 1, '10^22 atoms / cm^2')
XSMultiplicativeModel.__init__(self, name, (self.nH, ))
class HackPowerLaw(XSpowerlaw):
"""Restrict hard limits"""
def __init__(self, name='powerlaw'):
self.PhoIndex = Parameter(name, 'PhoIndex', 1., 0.95, 1.05, 0.95, 1.05)
self.norm = Parameter(name, 'norm', 9.2, 8.8, 9.7, 8.8, 9.7)
XSAdditiveModel.__init__(self, name, (self.PhoIndex, self.norm))
fit = setup['fit']
mcmc = sim.MCMC()
mcmc.set_sampler(sampler)
mcmc.set_sampler_opt("simarf", setup['pcaarf'])
mcmc.set_sampler_opt("p_M", 0.5)
mcmc.set_sampler_opt("nsubiter", 5)
covar_results = fit.est_errors()
cov = covar_results.extra_output
# Restrict the parameter values to try and trigger some
# invalid proposal steps. It's not obvious how the soft,
# hard, and prior function values all interact.
#
myabs = HackAbs()
mypl = HackPowerLaw()
pvals = np.asarray(covar_results.parvals)
pmins = np.asarray(covar_results.parmins)
pmaxs = np.asarray(covar_results.parmaxes)
fit.model = myabs * mypl
fit.model.thawedpars = pvals
fit.model.thawedparmins = pvals + 2 * pmins # pmins are < 0
fit.model.thawedparmaxes = pvals + 2 * pmaxs
# weight values away from the best-fit (does this actually
# help?)
#
for par in fit.model.pars:
mcmc.set_prior(par, inverse2)
niter = setup['niter']
with caplog.at_level(logging.INFO, logger='sherpa'):
# Do nothing with the warning at the moment, which could be
# a RuntimeWarning about the covariance matrix not being
# positive-semidefinite. This is just needed to make sure
# we don't trigger the default warning check.
#
with pytest.warns(Warning):
stats, accept, params = mcmc.get_draws(fit, cov, niter=niter)
# This is a lower bound, in case there's any messages from
# the sampling (either before or after displaying the
# 'Using Priors' message).
#
nrecords = len(caplog.record_tuples)
assert nrecords > 3
i = 0
while caplog.record_tuples[i][2] != 'Using Priors:':
i += 1
assert i < nrecords
assert i < (nrecords - 3)
assert caplog.record_tuples[i + 1][2].startswith('wabs.nH: <function inverse2 at ')
assert caplog.record_tuples[i + 2][2].startswith('powerlaw.PhoIndex: <function inverse2 at ')
assert caplog.record_tuples[i + 3][2].startswith('powerlaw.norm: <function inverse2 at ')
# It is not guaranteed what limits/checks we hit
#
have_hard_limit = False
have_reject = False
for loc, lvl, msg in caplog.record_tuples[i + 4:]:
if msg.startswith('Draw rejected: parameter boundary exception'):
have_reject = True
assert lvl == logging.INFO
elif msg.startswith('hard '):
have_hard_limit = True
assert lvl == logging.WARNING
assert have_hard_limit
assert have_reject