def test_gaussian(self):
"""Check the Gaussian likelihood."""
sigma = 1
dim = 5
theta = np.random.random(dim)
logl_expected = -(np.sum(theta ** 2) / (2 * sigma ** 2))
logl_expected -= np.log(2 * np.pi * sigma ** 2) * (dim / 2.0)
logl, phi = likelihoods.Gaussian(sigma=sigma)(theta)
self.assertAlmostEqual(logl, logl_expected, places=12)
self.assertIsInstance(phi, list)
self.assertEqual(len(phi), 0)
# Check matches sum of individal logls
sep_logls = [likelihoods.log_gaussian_pdf(th, sigma=sigma)
for th in theta]
self.assertAlmostEqual(sum(sep_logls), logl)
def test_python_run_func(self):
"""Check functions for running PolyChord via the pypolychord wrapper
(as opposed to with a compiled likelihood) in the form needed for
dynamic nested sampling."""
try:
os.makedirs(TEST_CACHE_DIR)
except FileExistsError:
pass
func = dyPolyChord.pypolychord_utils.RunPyPolyChord(
likelihoods.Gaussian(), priors.Uniform(), 2)
self.assertEqual(set(func.__dict__.keys()),
{'nderived', 'ndim', 'likelihood', 'prior'})
func({'base_dir': TEST_CACHE_DIR, 'file_root': 'temp', 'nlive': 5,
'max_ndead': 5, 'feedback': -1})
shutil.rmtree(TEST_CACHE_DIR)
def setUp(self):
"""Make a directory for saving test results."""
try:
os.makedirs(TEST_CACHE_DIR)
except FileExistsError:
pass
self.ninit = 40
ndim = 4
self.run_func = dyPolyChord.pypolychord_utils.RunPyPolyChord(
likelihoods.Gaussian(sigma=1), priors.Uniform(-10, 10), ndim=ndim)
self.random_seed_msg = (
'First dead point logl {0} != {1}. '
'This indicates that your PolyChord install\'s random number '
'generator is different to the one used to calculate the expected '
'results, so I will skip tests of exact numerical values. If all '
'the other tests pass then everything should work ok.')
self.settings = {
'do_clustering': True,
'posteriors': False,
'equals': False,
'write_dead': True,
'read_resume': False,
'write_resume': False,
'write_stats': True,
'write_prior': False,
'write_live': False,
'num_repeats': 10,
'feedback': -1,
'cluster_posteriors': False,
# Set precision_criterion to a relatively high value (i.e.
# precision) to PolyChord "non-deterministic likelihood" problems.
# These occur due in the low dimension and low and nlive
# cases we use for fast testing as runs sometimes get very close
# to the peak where the likelihood becomes approximately constant,
# and at this point PolyChord gets stuck.
'precision_criterion': 0.1,
'seed': 1,
'max_ndead': -1,
'base_dir': TEST_CACHE_DIR,
'file_root': 'test_run',
'nlive': 100, # used for nlive_const
'nlives': {}
}
def main():
"""Generate PolyChord runs. Also processes the results into a DataFrame and
caches it.
Nested sampling runs are generating for different settings by looping over:
* likelihood_list: different likelihoods;
* nd_nl_nr_list: list of tuples, each containing (number of dimensions,
nlive, nrepeats);
* inds: labels for repeated runs to generate with each setting.
"""
# Settings
# --------
# If true, many runs are made at the same time via concurrent.futures
parallel = True
# Run settings
inds = list(range(1, 101))
# dimensions, nlive, nrepeat settings
# -----------------------------------
nd_nl_nr_list = diagnostic.settings.get_nd_nl_nr_list()
# Likelihood and prior settings
# -----------------------------
likelihood_list = [likelihoods.LogGammaMix(),
likelihoods.Gaussian(sigma=1)]
# PolyChord settings
settings_dict = {
'do_clustering': True,
'posteriors': False,
'equals': False,
'base_dir': 'chains',
'feedback': -1,
'precision_criterion': 0.001,
'nlives': {},
'write_dead': True,
'write_stats': True,
'write_paramnames': False,
'write_prior': False,
'write_live': False,
'write_resume': False,
'read_resume': False,
'max_ndead': -1,
'cluster_posteriors': False,
'boost_posterior': 0.0}
if 'ed' in os.getcwd().split('/'):
# running on laptop - don't use all the processors so I can do other
# stuff without everything getting slow
max_workers = 6
else:
max_workers = None # running on cluster
print('Running with max_workers={}'.format(max_workers))
prior_scale = 30
prior = priors.Uniform(-prior_scale, prior_scale)
# Before running in parallel make sure base_dir exists, as if multiple
# threads try to make one at the same time mkdir throws an error.
if not os.path.exists(settings_dict['base_dir']):
os.makedirs(settings_dict['base_dir'])
if not os.path.exists(settings_dict['base_dir'] + '/clusters'):
os.makedirs(settings_dict['base_dir'] + '/clusters')
for likelihood in likelihood_list:
for ndim, nlive, num_repeats in nd_nl_nr_list:
run_func = dyPolyChord.pypolychord_utils.RunPyPolyChord(
likelihood, prior, ndim)
# make list of settings dictionaries for the different repeats
file_root = dyPolyChord.output_processing.settings_root(
type(likelihood).__name__,
type(prior).__name__, ndim,
prior_scale=prior_scale, nrepeats=num_repeats,
nlive_const=nlive, dynamic_goal=None)
settings_dict['nlive'] = nlive
settings_dict['num_repeats'] = num_repeats
settings_list = []
for extra_root in inds:
settings = copy.deepcopy(settings_dict)
settings['seed'] = extra_root
settings['file_root'] = file_root
settings['file_root'] += '_' + str(extra_root).zfill(3)
settings_list.append(settings)
# Do the nested sampling
# ----------------------
# For standard nested sampling just run PolyChord
desc = '{} ndim={} nlive={} nrep={}'.format(
type(likelihood).__name__, ndim, nlive, num_repeats)
nestcheck.parallel_utils.parallel_apply(
run_func, settings_list,
max_workers=max_workers, parallel=parallel,
tqdm_kwargs={'desc': desc, 'leave': True})
# Cache results DataFrame
# -----------------------
diagnostic.data_loading.get_results_df(
[type(likelihood).__name__.replace('Mix', ' mix')],
[(ndim, nlive, num_repeats)], n_simulate=100,
nrun=inds[-1], summary=True, save=True, load=True,
thread_pvalue=False, bs_stat_dist=False,
include_rmse=True, include_true_values=True, parallel=True)
if ((ndim, nlive, num_repeats) ==
diagnostic.settings.get_default_nd_nl_nr()):
# Cache bs stat and thread values df too
# Use summary=True as caching is done to raw values and this
# protects against unexpected kwarg errors as summary pops some
# kwargs before getting values
diagnostic.data_loading.get_results_df(
[type(likelihood).__name__.replace('Mix', ' mix')],
[(ndim, nlive, num_repeats)], n_simulate=100,
nrun=inds[-1], summary=True, save=True, load=True,
thread_pvalue=True, bs_stat_dist=True,
include_rmse=True, include_true_values=True, parallel=True)