__author__ = "aymgal, johannesulf"
import os
import shutil
import numpy as np
import sys
import copy
from lenstronomy.Sampling.Samplers.base_nested_sampler import NestedSampler
import lenstronomy.Util.sampling_util as utils
__all__ = ["DyPolyChordSampler"]
[docs]class DyPolyChordSampler(NestedSampler):
"""Wrapper for dynamical nested sampling algorithm DyPolyChord by E. Higson, M.
Hobson, W. Handley, A. Lasenby.
papers : arXiv:1704.03459, arXiv:1804.06406
doc : https://dypolychord.readthedocs.io
"""
[docs] def __init__(
self,
likelihood_module,
prior_type="uniform",
prior_means=None,
prior_sigmas=None,
width_scale=1,
sigma_scale=1,
output_dir=None,
output_basename="-",
resume_dyn_run=False,
polychord_settings=None,
remove_output_dir=False,
use_mpi=False,
): # , num_mpi_procs=1):
"""
:param likelihood_module: likelihood_module like in likelihood.py (should be callable)
:param prior_type: 'uniform' of 'gaussian', for converting the unit hypercube to param cube
:param prior_means: if prior_type is 'gaussian', mean for each param
:param prior_sigmas: if prior_type is 'gaussian', std dev for each param
:param width_scale: scale the widths of the parameters space by this factor
:param sigma_scale: if prior_type is 'gaussian', scale the gaussian sigma by this factor
:param output_dir: name of the folder that will contain output files
:param output_basename: prefix for output files
:param resume_dyn_run: if True, previous resume files will not be deleted so that previous run can be resumed
:param polychord_settings: settings dictionary to send to pypolychord. Check dypolychord documentation for details.
:param remove_output_dir: remove the output_dir folder after completion
:param use_mpi: Use MPI computing if `True`
"""
self._check_install()
super(DyPolyChordSampler, self).__init__(
likelihood_module,
prior_type,
prior_means,
prior_sigmas,
width_scale,
sigma_scale,
)
# if use_mpi:
# mpi_str = 'mpirun -np {}'.format(num_mpi_procs)
# else:
# mpi_str = None
if polychord_settings == None:
polychord_settings = {}
self._use_mpi = use_mpi
self._output_dir = output_dir
self._is_master = True
if self._use_mpi:
from mpi4py import MPI
self._comm = MPI.COMM_WORLD
if self._comm.Get_rank() != 0:
self._is_master = False
else:
self._comm = None
if not resume_dyn_run:
if self._is_master:
if os.path.exists(self._output_dir):
shutil.rmtree(self._output_dir, ignore_errors=True)
os.mkdir(self._output_dir)
self._output_basename = output_basename
self._settings = copy.deepcopy(polychord_settings)
self._settings["file_root"] = self._output_basename
self._settings["base_dir"] = self._output_dir
if self._all_installed:
# create the dyPolyChord callable object
self._sampler = self._RunPyPolyChord(
self.log_likelihood, self.prior, self.n_dims
)
else:
self._sampler = None
self._rm_output = remove_output_dir
self._has_warned = False
[docs] def run(self, dynamic_goal, kwargs_run):
"""Run the DyPolyChord dynamical nested sampler.
see https://dypolychord.readthedocs.io for content of kwargs_run
:param dynamic_goal: 0 for evidence computation, 1 for posterior computation
:param kwargs_run: kwargs directly passed to dyPolyChord.run_dypolychord
:return: samples, means, logZ, logZ_err, logL, ns_run
"""
print("prior type :", self.prior_type)
print("parameter names :", self.param_names)
if self._all_installed:
# TODO : put a default dynamic_goal ?
# dynamic_goal = 0 for evidence-only, 1 for posterior-only
self._dyPolyChord.run_dypolychord(
self._sampler,
dynamic_goal,
settings_dict_in=self._settings,
comm=self._comm,
**kwargs_run
)
if self._is_master:
ns_run = self._ns_process_run(
self._settings["file_root"], self._settings["base_dir"]
)
else:
# in case DyPolyChord or NestCheck was not compiled properly, for unit tests
ns_run = {
"theta": np.zeros((1, self.n_dims)),
"logl": np.zeros(1),
"output": {
"logZ": np.zeros(1),
"logZerr": np.zeros(1),
"param_means": np.zeros(self.n_dims),
},
}
self._write_equal_weights(ns_run["theta"], ns_run["logl"])
if self._is_master:
samples, logL = self._get_equal_weight_samples()
# logL = ns_run['logl']
# samples_w = ns_run['theta']
logZ = ns_run["output"]["logZ"]
logZ_err = ns_run["output"]["logZerr"]
means = ns_run["output"]["param_means"]
print("The log evidence estimate using the first run is {}".format(logZ))
print("The estimated mean of the first parameter is {}".format(means[0]))
if self._rm_output:
shutil.rmtree(self._output_dir, ignore_errors=True)
return samples, means, logZ, logZ_err, logL, ns_run
else:
sys.exit(0)
[docs] def log_likelihood(self, args):
"""Compute the log-likelihood given list of parameters.
:param args: parameter values
:return: log-likelihood (from the likelihood module)
"""
return super().log_likelihood(args), []
def _get_equal_weight_samples(self):
"""Inspired by pymultinest's Analyzer, because DyPolyChord has more or less the
same output conventions as MultiNest."""
file_name = "{}_equal_weights.txt".format(self._output_basename)
file_path = os.path.join(self._output_dir, file_name)
data = np.loadtxt(file_path, ndmin=2)
logL = -0.5 * data[:, 0]
samples = data[:, 1:]
return samples, logL
def _write_equal_weights(self, samples, logL):
# write fake output file for unit tests
file_name = "{}_equal_weights.txt".format(self._output_basename)
file_path = os.path.join(self._output_dir, file_name)
data = np.zeros((samples.shape[0], 1 + samples.shape[1]), dtype=float)
data[:, 0] = -2.0 * logL
data[:, 1:] = samples
np.savetxt(file_path, data, fmt="% .14E")
def _check_install(self):
try:
import dyPolyChord
from dyPolyChord import pypolychord_utils
except ImportError:
dyPolyChord = None
pypolychord_utils = None
print(
"Warning : dyPolyChord not properly installed. \
You can get it from : https://github.com/ejhigson/dyPolyChord"
)
dypolychord_installed = False
else:
dypolychord_installed = True
self._dyPolyChord = dyPolyChord
self._RunPyPolyChord = pypolychord_utils.RunPyPolyChord
try:
from nestcheck import data_processing
except ImportError:
print(
"Warning : nestcheck not properly installed (results might be unexpected). \
You can get it from : https://github.com/ejhigson/nestcheck"
)
nestcheck_installed = False
else:
nestcheck_installed = True
self._ns_process_run = data_processing.process_polychord_run
self._all_installed = dypolychord_installed and nestcheck_installed