"""Main driver of the fitting routine."""
__all__ = ['Fitter', 'dynesty_log_like', 'dynesty_loglike_bma', 'pt_dynesty',
'pt_multinest']
import copy
import logging
import os
import pickle
import time
import warnings
logger = logging.getLogger(__name__)
from multiprocessing import (Pool, set_start_method)
from tqdm import tqdm
import extinction
import astropy.units as u
import pandas as pd
import numpy as np
import scipy.stats as st
from numpy.random import choice
from astropy.constants import sigma_sb
from isochrones.interp import DFInterpolator
from termcolor import colored
from .config import (filesdir, gridsdir, priorsdir, filter_names, colors,
iso_mask, iso_bands, grid_wave_coverage)
from .error import *
from .isochrone import estimate
from .phot_utils import *
from .sed_library import *
from .utils import *
try:
import dynesty
from dynesty.utils import resample_equal
bma_flag = True
except ModuleNotFoundError:
wrn = 'Dynesty package not found.\n'
wrn += 'Install dynesty with `pip install dynesty`'
warnings.warn(wrn)
bma_flag = False
try:
import pymultinest
except ModuleNotFoundError:
warnings.warn(
'(py)MultiNest installation (or libmultinest.dylib) not detected.'
)
set_start_method('fork', force=True)
[docs]
class Fitter:
"""The Fitter class handles the fitting routines and parameter estimation.
Examples
--------
The fitter isn't instantiaded with any arguments, rather you instantiate a
Fitter object, then you set up the configurations and finally you
initialize the object by running the initialize method.
>>> f = Fitter()
>>> f.star = s # s must be a valid Star object.
>>> f.initialize()
Attributes
----------
star : Star
The ``Star`` object to fit.
setup : list
Sampler configuration. The first element selects the engine
(``'dynesty'``); the remaining elements are
``[nlive, dlogz, bound, sample, threads, dynamic, walks]``.
out_folder : str
Folder where results and plots are written. Defaults to the star name.
verbose : bool
Program verbosity. Default ``True``.
norm : bool
If ``True``, fit a single normalisation constant instead of radius and
distance; the radius is recovered afterwards from the Gaia parallax.
grid : str
Atmosphere grid for a single-grid fit (e.g. ``'phoenix'``,
``'btsettl'``, ``'ck04'``, ``'kurucz'``, ``'coelho'``, ``'bosz'``,
``'sphinx'``, ``'tlusty'``).
bma : bool
If ``True``, run Bayesian Model Averaging across several grids. This
loads every model interpolator and fits all of them, so it is slower.
models : list
The grids to average over when ``bma`` is ``True``.
sequential : bool
If ``True``, fit the BMA grids one after another. See also
``n_grid_jobs`` to fit grids concurrently. Default ``True``.
n_grid_jobs : int
Number of BMA grids fit concurrently, one core each. Default ``1``
(sequential). Each job loads its own grid, so size this to the
machine's core count and memory.
n_samples : int or None
Number of samples drawn from the BMA-averaged posterior. ``None`` uses
the combined chain length.
walks : int
Number of ``rwalk`` MCMC steps per dynesty proposal. Default ``25``;
also settable as ``setup[7]``.
av_law : str
Extinction law. Default ``'fitzpatrick'``.
estimate_logg : bool
If ``True``, treat log g as a derived quantity via the isochrone step.
Default ``False``.
estimate_age : bool
If ``True``, run the MIST isochrone fit for age, mass and EEP after the
SED fit. Default ``True``.
isochrone_dlogz : float
Evidence tolerance for the MIST isochrone age/mass fit (the dominant
cost of a BMA run). Lower is more precise but slower. Default ``0.01``.
prior_setup : dict or None
Manual prior specification consumed by ``create_priors_from_setup``.
experimental : bool
Enable experimental code paths. Default ``False``.
"""
colors = colors
def __init__(self):
# Default values for attributes
self._interpolators = []
self._grids = []
self.out_folder = None
self.verbose = True
self.star = None
self.setup = ['dynesty']
self.norm = False
self.grid = 'phoenix'
self.estimate_logg = False
self.estimate_age = True
self.av_law = 'fitzpatrick'
self.n_samples = None
self.bma = False
self.prior_setup = None
self.sequential = True
self.experimental = False
# Number of MCMC steps per proposal for dynesty's rwalk sampler.
self.walks = 25
# Number of BMA model grids to fit concurrently, each on a single
# core. 1 (default) keeps the original sequential behaviour. Set this
# to len(models) to run every grid at once. CAUTION: this spawns
# ``n_grid_jobs`` processes, so choose a value your machine's core
# count and memory can handle (each grid loads its own model grid).
self.n_grid_jobs = 1
# Evidence tolerance for the MIST isochrone age/mass nested sampling
# (the dominant cost of a BMA run). Lower is more precise but slower;
# the isochrone fit itself is parallelised across ``threads``.
self.isochrone_dlogz = 0.01
@property
def star(self):
"""Star to fit for."""
return self._star
@star.setter
def star(self, star):
self._star = star
@property
def setup(self):
"""Set up options."""
return self._setup
@setup.setter
def setup(self, setup):
err_msg = 'The setup has to contain at least the fitting engine'
err_msg += f', multinest or dynesty.\nThe setup was {setup}'
if len(setup) < 1:
InputError(err_msg).__raise__()
self._setup = setup
self._engine = setup[0]
defaults = False
if len(setup) == 1:
defaults = True
if self._engine == 'multinest':
warnings.warn(
"The 'multinest' engine is deprecated and will be removed "
"entirely in astroARIADNE v2.0. It only supports single-grid "
"fits (not BMA); use engine='dynesty' instead.",
DeprecationWarning, stacklevel=2,
)
if defaults:
self._nlive = 500
self._dlogz = 0.5
else:
self._nlive = setup[1]
self._dlogz = setup[2]
if self._engine == 'dynesty':
if defaults:
self._nlive = 500
self._dlogz = 0.5
self._bound = 'multi'
self._sample = 'rwalk'
self._threads = 1
self._dynamic = False
else:
self._nlive = setup[1]
self._dlogz = setup[2]
self._bound = setup[3]
self._sample = setup[4]
self._threads = setup[5]
self._dynamic = setup[6]
# Optional 8th element: number of rwalk MCMC steps per
# proposal. Omitted -> keep the current value (default 25),
# so existing 7-element setups are unaffected.
if len(setup) > 7:
self.walks = setup[7]
@property
def norm(self):
"""Bool to decide if a normalization constant will be fitted.
Set as True to not fit for radius and distance and fit for a
normalization constant. After the fit a radius is calculated using
the Gaia parallax.
"""
return self._norm
@norm.setter
def norm(self, norm):
if type(norm) is not bool:
InputError('norm must be True or False.').__raise__()
self._norm = norm
@property
def grid(self):
"""Model grid selected."""
return self._grid
@grid.setter
def grid(self, grid):
assert type(grid) == str
self._grid = grid
if grid.lower() == 'phoenix':
with open(gridsdir + '/Phoenixv2_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'btsettl':
with open(gridsdir + '/BTSettl_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'btnextgen':
with open(gridsdir + '/BTNextGen_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'btcond':
with open(gridsdir + '/BTCond_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'ck04':
with open(gridsdir + '/CK04_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'kurucz':
with open(gridsdir + '/Kurucz_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'coelho':
with open(gridsdir + '/Coelho_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'bosz':
with open(gridsdir + '/BOSZ_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'sphinx':
with open(gridsdir + '/SPHINX_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
if grid.lower() == 'tlusty':
with open(gridsdir + '/TLUSTY_DF.pkl', 'rb') as intp:
self._interpolator = DFInterpolator(pd.read_pickle(intp))
@property
def bma(self):
"""Bayesian Model Averaging (BMA).
Set to True if BMA is wanted. This loads every model grid interpolator
and fits an SED to all of them, so the runtime will be slower!
"""
return self._bma
@bma.setter
def bma(self, bma):
self._bma = bma if bma_flag else False
@property
def models(self):
"""Models to be used in BMA."""
return self._bma_models
@models.setter
def models(self, mods):
self._bma_models = mods
@property
def sequential(self):
"""Set to True to make BMA sequentially instead of parallel."""
return self._sequential
@sequential.setter
def sequential(self, sequential):
self._sequential = sequential
@property
def n_samples(self):
"""Set number of samples for BMA."""
return self._nsamp
@n_samples.setter
def n_samples(self, nsamp):
self._nsamp = nsamp
@property
def verbose(self):
"""Program verbosity. Default is True."""
return self._verbose
@verbose.setter
def verbose(self, verbose):
if type(verbose) is not bool:
InputError('Verbose must be True or False.').__raise__()
self._verbose = verbose
@property
def out_folder(self):
"""Output folder.
If none is provided the default will be the starname.
"""
return self._out_folder
@out_folder.setter
def out_folder(self, out_folder):
if type(out_folder) is not str and out_folder is not None:
err_msg = 'Output folder must be an address or None.'
InputError(err_msg).__raise__()
self._out_folder = out_folder
@property
def av_law(self):
"""Select extinction law."""
return self._av_law
@av_law.setter
def av_law(self, law):
laws = [
'cardelli',
'odonnell',
'calzetti',
'fitzpatrick'
]
law = law.lower()
if law not in laws:
err_msg = f'Extinction law {law} not recognized.'
err_msg += 'Available extinction laws are: `cardelli`, `odonnell`'
err_msg += ', `calzetti`, and `fitzpatrick`'
InputError(err_msg).__raise__()
law_f = None
if law == laws[0]:
law_f = extinction.ccm89
if law == laws[1]:
law_f = extinction.odonnell94
if law == laws[2]:
law_f = extinction.calzetti00
if law == laws[3]:
law_f = extinction.fitzpatrick99
self._av_law = law_f
def _apply_grid_coverage(self):
"""Drop bands outside the current grid's native wavelength coverage.
Some grids (e.g. Coelho 2014) only cover the optical; their redder
photometry was extrapolated when the grid was built and is unreliable.
For a standalone fit we therefore exclude those bands. The user's Star
is left untouched: a shallow copy with a restricted ``used_filters`` /
``filter_mask`` is substituted, so every downstream consumer
(dimension count, priors, likelihood, output) sees the reduced set.
"""
cov = grid_wave_coverage.get(self._grid.lower())
if cov is None:
return
used = self.star.used_filters == 1
outside = used & (self.star.wave > cov)
if not outside.any():
return
dropped = self.star.filter_names[outside]
s = copy.copy(self.star)
s.used_filters = self.star.used_filters.copy()
s.used_filters[outside] = 0
s.filter_mask = np.where(s.used_filters == 1)[0]
self.star = s
logger.warning(
"Grid '%s' only covers <= %.2f um; excluding %d out-of-coverage "
"band(s) from the fit: %s", self._grid, cov, len(dropped),
', '.join(dropped))
[docs]
def initialize(self):
"""Initialize the fitter.
To be run only after every input is added.
This function calculates the number of dimensions, runs the prior
creation, creates output directory, initializes coordinators and sets
up global variables.
"""
global prior_dict, coordinator, fixed, order, star, use_norm, av_law
self.start = time.time()
err_msg = 'No star is detected. Please create an instance of Star.'
if self.star is None:
er = InputError(err_msg)
er.log(self.out + '/output.log')
er.__raise__()
# For a standalone fit with a grid of limited wavelength coverage,
# exclude out-of-coverage bands (e.g. Coelho's extrapolated IR). Not
# applied in BMA, where every grid must share the same band set.
if not self._bma:
self._apply_grid_coverage()
star = self.star
if not self._bma:
global interpolator
interpolator = self._interpolator
use_norm = self.norm
# Extinction law
av_law = self._av_law
# Declare order of parameters.
if not self.norm:
order = np.array(
[
'teff', 'logg', 'z',
'dist', 'rad', 'Av'
]
)
else:
order = np.array(['teff', 'logg', 'z', 'norm', 'Av'])
# Create output directory
if self.out_folder is None:
self.out_folder = self.star.starname + '/'
create_dir(self.out_folder)
self.star.save_mags(self.out_folder + '/')
# Parameter coordination.
# Order for the parameters are:
# teff, logg, z, dist, rad, Av, noise
# or
# teff, logg, z, norm, Av, noise
npars = 6 if not self.norm else 5
npars += self.star.used_filters.sum()
npars = int(npars)
self.coordinator = np.zeros(npars) # 1 for fixed params
self.fixed = np.zeros(npars)
coordinator = self.coordinator
fixed = self.fixed
# Setup priors.
self.default_priors = self._default_priors()
self.create_priors_from_setup()
prior_dict = self.priors
# Get dimensions.
self.ndim = self.get_ndim()
# warnings
if len(self._setup) == 1:
logger.warning('Using default setup values')
# BMA settings
# if BMA is used, load all interpolators requested.
if self._bma:
if self.n_samples is None:
self.n_samples = 'max'
if self.star.offline:
if self.star.temp is None:
self.star.temp = 4001
off_msg = 'Offline mode assumes that the stellar'
off_msg += ' temperature is greater than 4000 K'
off_msg += '. If you believe this is not the case then please '
off_msg += 'add a temperature to the Star constructor. '
off_msg += f'The input temperature is {self.star.temp}'
print(colored(off_msg, 'yellow'))
for mod in self._bma_models:
# We'll assume that if ARIADNE is running in offline mode
# Then the star will have > 4000 K
if (mod.lower() in ['btcond', 'btnextgen'] and
self.star.temp > 4000) or \
(mod.lower() in ['ck04', 'kurucz'] and
self.star.temp < 4000) or \
(mod.lower() == 'coelho' and self.star.temp < 3500) or \
(mod.lower() == 'bosz' and
(self.star.temp < 2800 or self.star.temp > 16000)) or \
(mod.lower() == 'sphinx' and self.star.temp > 4000) or \
(mod.lower() == 'tlusty' and self.star.temp < 15000):
continue
df = self.load_interpolator(mod.lower())
self._interpolators.append(df)
self._grids.append(mod)
thr = self._threads if self._sequential else len(
self._interpolators)
else:
thr = self._threads
en = 'Bayesian Model Averaging' if self._bma else self._engine
display_routine(en, self._nlive, self._dlogz, self.ndim, self._bound,
self._sample, thr, self._dynamic)
self.show_priors()
[docs]
def get_ndim(self):
"""Calculate number of dimensions."""
ndim = 6 if not self.norm else 5
ndim += self.star.used_filters.sum()
ndim -= self.coordinator.sum()
return int(ndim)
def _default_priors(self):
global order
defaults = dict()
# Initialize prior source tracking
if not hasattr(self, 'prior_sources'):
self.prior_sources = {}
# Resolve spectroscopic parameters: prefer spectroscopic_params,
# fall back to rave_params for backward compatibility.
spec = getattr(self.star, 'spectroscopic_params', None)
if spec is None:
spec = getattr(self.star, 'rave_params', None)
spec_source = spec.get('source', 'spectroscopic') if spec else None
# Per-field null-safe helper: a spectroscopic field is usable only if
# the spec dict exists and both the value and its error are not None.
# This handles partial spec dicts (e.g. Hypatia [Fe/H]-only results
# where teff/logg are None) by falling back per-field to the
# population/default prior rather than crashing.
def _spec_has(field):
return (spec is not None and
spec.get(field) is not None and
spec.get(field + '_err') is not None)
# Teff: hot-star grids (TLUSTY) need a wide uniform prior — the FGK
# population prior assigns ~zero probability above ~12000 K, so it
# cannot reach the O/B regime. Applied for standalone TLUSTY fits.
if (getattr(self, '_grid', '') or '').lower() == 'tlusty':
defaults['teff'] = st.uniform(loc=15000, scale=40000) # 15-55 kK
self.prior_sources['teff'] = 'tlusty_uniform'
logger.info('Using uniform Teff prior for the TLUSTY hot-star grid')
# Teff: spectroscopic star-specific or population prior
elif _spec_has('teff'):
defaults['teff'] = st.norm(loc=spec['teff'],
scale=spec['teff_err'])
self.prior_sources['teff'] = spec_source
logger.info('Using %s Teff prior (star-specific)', spec_source)
else:
with open(priorsdir + '/teff_ppf.pkl', 'rb') as jar:
defaults['teff'] = pickle.load(jar)
self.prior_sources['teff'] = 'rave_population'
logger.info('Using population Teff prior (no spectroscopic match)')
# logg: spectroscopic, isochrone, or population prior
if _spec_has('logg'):
defaults['logg'] = st.norm(loc=spec['logg'],
scale=spec['logg_err'])
self.prior_sources['logg'] = spec_source
logger.info('Using %s logg prior (star-specific)', spec_source)
elif self.star.get_logg:
defaults['logg'] = st.norm(
loc=self.star.logg, scale=self.star.logg_e)
self.prior_sources['logg'] = 'isochrone'
logger.info('Using isochrone logg estimate')
else:
defaults['logg'] = st.uniform(loc=3.5, scale=2.5)
self.prior_sources['logg'] = 'uniform_default'
logger.info('Using population logg prior')
# [Fe/H]: spectroscopic or population prior
if _spec_has('feh'):
defaults['z'] = st.norm(loc=spec['feh'], scale=spec['feh_err'])
self.prior_sources['z'] = spec_source
logger.info('Using %s [Fe/H] prior (star-specific)', spec_source)
else:
defaults['z'] = st.norm(loc=-0.125, scale=0.234)
self.prior_sources['z'] = 'population'
logger.info('Using population [Fe/H] prior')
# Distance prior setup.
if not self._norm:
if self.star.dist != -1:
_d_scale = 3 * self.star.dist_e
defaults['dist'] = st.truncnorm(
a=-self.star.dist / _d_scale, b=1e100,
loc=self.star.dist, scale=_d_scale,
)
self.prior_sources['dist'] = 'gaia'
else:
defaults['dist'] = st.uniform(loc=1, scale=3000)
self.prior_sources['dist'] = 'uniform_default'
# Radius prior setup.
defaults['rad'] = st.uniform(loc=0.05, scale=100)
self.prior_sources['rad'] = 'uniform_default'
# Normalization prior setup.
else:
# up = 1 / 1e-30
# defaults['norm'] = st.truncnorm(a=0, b=up, loc=1e-20, scale=1e-10)
defaults['norm'] = st.uniform(loc=0, scale=1e-20)
self.prior_sources['norm'] = 'uniform_default'
# Extinction prior setup.
if self.star.Av == 0.:
av_idx = 4 if self._norm else 5
self.coordinator[av_idx] = 1
self.fixed[av_idx] = 0
defaults['Av'] = None
self.prior_sources['Av'] = 'fixed_zero'
else:
if self.star.Av_e is None:
defaults['Av'] = st.uniform(loc=0, scale=self.star.Av)
self.prior_sources['Av'] = 'dustmap_uniform'
else:
mu = self.star.Av
sig = self.star.Av_e
low = 0
up = 100
b, a = (up - mu) / sig, (low - mu) / sig
defaults['Av'] = st.truncnorm(loc=mu, scale=sig, a=a, b=b)
self.prior_sources['Av'] = 'dustmap'
# Noise model prior setup.
mask = self.star.filter_mask
flxs = self.star.flux[mask]
errs = self.star.flux_er[mask]
for filt, flx, flx_e in zip(self.star.filter_names[mask], flxs, errs):
p_ = get_noise_name(filt) + '_noise'
mu = 0
sigma = flx_e * 10
b = (1 - flx) / flx_e
defaults[p_] = st.truncnorm(loc=mu, scale=sigma, a=0, b=b)
# defaults[p_] = st.uniform(loc=0, scale=5)
order = np.append(order, p_)
return defaults
[docs]
def create_priors_from_setup(self):
"""Create priors from the manual setup."""
prior_dict = dict()
keys = self.prior_setup.keys()
noise = []
mask = self.star.filter_mask
flxs = self.star.flux[mask]
errs = self.star.flux_er[mask]
for filt, flx, flx_e in zip(self.star.filter_names[mask], flxs, errs):
p_ = get_noise_name(filt) + '_noise'
noise.append(p_)
prior_out = 'Parameter\tPrior\tValues\n'
if 'norm' in keys and ('rad' in keys or 'dist' in keys):
er = PriorError('rad or dist', 1)
er.log(self.out_folder + '/output.log')
er.__raise__()
for k in keys:
if type(self.prior_setup[k]) == str:
if self.prior_setup[k] == 'default':
prior_dict[k] = self.default_priors[k]
prior_out += k + '\tdefault\n'
if self.prior_setup[k].lower() == 'rave':
# RAVE prior only available for teff and logg. It's already
# the default for [Fe/H]
if k == 'logg' or k == 'teff':
if k == 'teff':
with open(priorsdir + '/teff_ppf.pkl', 'rb') as jar:
prior_dict[k] = pickle.load(jar)
else: # k == 'logg'
with open(priorsdir + '/logg_ppf.pkl', 'rb') as jar:
prior_dict[k] = pickle.load(jar)
prior_out += k + '\tRAVE\n'
else:
prior = self.prior_setup[k][0]
if prior == 'fixed':
value = self.prior_setup[k][1]
idx = np.where(k == order)[0]
self.coordinator[idx] = 1
self.fixed[idx] = value
prior_out += k + '\tfixed\t{}\n'.format(value)
if hasattr(self, 'prior_sources'):
self.prior_sources[k] = 'user_fixed'
if prior == 'normal':
mu = self.prior_setup[k][1]
sig = self.prior_setup[k][2]
prior_dict[k] = st.norm(loc=mu, scale=sig)
prior_out += k + '\tnormal\t{}\t{}\n'.format(mu, sig)
if hasattr(self, 'prior_sources'):
self.prior_sources[k] = 'user_normal'
if prior == 'truncnorm':
mu = self.prior_setup[k][1]
sig = self.prior_setup[k][2]
low = self.prior_setup[k][3]
up = self.prior_setup[k][4]
b, a = (up - mu) / sig, (low - mu) / sig
prior_dict[k] = st.truncnorm(a=a, b=b, loc=mu, scale=sig)
prior_out += k
prior_out += '\ttruncatednormal\t{}\t{}\t{}\t{}\n'.format(
mu, sig, low, up)
if hasattr(self, 'prior_sources'):
self.prior_sources[k] = 'user_truncnorm'
if prior == 'uniform':
low = self.prior_setup[k][1]
up = self.prior_setup[k][2]
prior_dict[k] = st.uniform(loc=low, scale=up - low)
prior_out += k + '\tuniform\t{}\t{}\n'.format(low, up)
if hasattr(self, 'prior_sources'):
self.prior_sources[k] = 'user_uniform'
for par in noise:
prior_dict[par] = self.default_priors[par]
ff = open(self.out_folder + '/prior.dat', 'w')
ff.write(prior_out)
ff.close()
del ff
self.priors = prior_dict
pass
[docs]
def show_priors(self, return_string=False):
"""Display priors in statistical notation.
Parameters
----------
return_string : bool, optional
If True, return the formatted string instead of printing.
Default: False
Returns
-------
str or None
Formatted prior information if return_string=True, else None
"""
global order
_T = '\t\t\t'
c = np.random.choice(colors)
# Physical parameters to display (skip noise priors)
physical_params = []
for param_name in self.priors.keys():
if param_name.endswith('_noise'):
continue
physical_params.append(param_name)
# Format each prior
prior_strs = {}
for param_name in physical_params:
prior_obj = self.priors[param_name]
param_idx = np.where(order == param_name)[0]
if len(param_idx) > 0 and self.coordinator[param_idx[0]] == 1:
fixed_value = self.fixed[param_idx[0]]
prior_str = f"Fixed({fixed_value:.4f})"
else:
prior_str = self._format_prior_notation(param_name, prior_obj)
prior_strs[param_name] = prior_str
# Sort parameters in logical order
param_order = ['teff', 'logg', 'z', 'dist', 'rad', 'norm', 'Av']
sorted_params = [p for p in param_order if p in physical_params]
sorted_params.extend([p for p in physical_params if p not in param_order])
# Build table
lines = []
lines.append(f'{_T}{"Parameter":24s} {"Prior"}')
lines.append(f'{_T}{"-" * 52}')
for param_name in sorted_params:
lines.append(f'{_T}{param_name:24s} {prior_strs[param_name]}')
output = '\n'.join(lines)
if return_string:
return output
else:
print(colored(output, c))
print()
return None
def _format_prior_notation(self, param_name, prior_obj):
"""Format scipy.stats distribution in statistical notation.
Parameters
----------
param_name : str
Parameter name (e.g., 'teff', 'logg')
prior_obj : scipy.stats distribution
The prior distribution object
Returns
-------
str
Formatted statistical notation
"""
# RAVE population prior: loaded as an InterpolatedUnivariateSpline
# from teff_ppf.pkl / logg_ppf.pkl. Detect by type since manual
# prior setups don't populate prior_sources.
if 'Spline' in type(prior_obj).__name__:
return "RAVE (population)"
if hasattr(self, 'prior_sources') and param_name in self.prior_sources:
if self.prior_sources[param_name] == 'rave_population':
return "RAVE (population)"
# Frozen scipy distributions expose .dist (the class) and .kwds/.args
# (the frozen parameters). Use the underlying class name to identify
# the distribution.
if hasattr(prior_obj, 'dist'):
inner_name = type(prior_obj.dist).__name__
kwds = dict(getattr(prior_obj, 'kwds', {}) or {})
args = list(getattr(prior_obj, 'args', ()) or ())
else:
inner_name = type(prior_obj).__name__
kwds, args = {}, []
def _pop(key, idx, default=0.0):
if key in kwds:
return kwds[key]
if idx < len(args):
return args[idx]
return default
try:
if 'norm_gen' in inner_name and 'truncnorm' not in inner_name:
mu = _pop('loc', 0)
sigma = _pop('scale', 1, 1.0)
return f"N({mu:.3f}, {sigma:.3f})"
if 'uniform_gen' in inner_name:
a = _pop('loc', 0)
scale = _pop('scale', 1, 1.0)
b = a + scale
return f"U({a:.2f}, {b:.2f})"
if 'truncnorm_gen' in inner_name:
a_std = _pop('a', 0)
b_std = _pop('b', 1)
mu = _pop('loc', 2)
sigma = _pop('scale', 3, 1.0)
a_act = mu + a_std * sigma
b_act = mu + b_std * sigma
a_str = "-inf" if a_std < -1e10 else f"{a_act:.2f}"
b_str = "inf" if b_std > 1e10 else f"{b_act:.2f}"
return f"TN({mu:.2f}, {sigma:.2f}, [{a_str}, {b_str}])"
return inner_name
except (AttributeError, TypeError):
return f"{inner_name} (custom)"
[docs]
def fit(self):
"""Run fitting routine."""
if self._engine == 'multinest':
self.fit_multinest()
else:
self.fit_dynesty()
elapsed_time = execution_time(self.start)
end(self.coordinator, elapsed_time,
self.out_folder, self._engine, self.norm, )
pass
[docs]
def fit_bma(self):
"""Perform the fit with different models and the average the output.
Only works with dynesty.
"""
if len(self.star.filter_names[self.star.filter_mask]) <= 5:
print(colored('\t\t\tNOT ENOUGH POINTS TO MAKE THE FIT! !', 'red'))
return
global interpolator, _BMA_FITTER
n_jobs = min(max(int(self.n_grid_jobs), 1), len(self._grids))
if n_jobs > 1:
# Across-grid parallelism: the grids are independent fits, so run
# several at once (one core each). Each worker writes its own
# ``_out.pkl``, which we read back below for averaging.
ncpu = os.cpu_count() or 1
if n_jobs > ncpu:
logger.warning(
'n_grid_jobs=%d exceeds the %d available CPU cores; '
'this will oversubscribe the machine.', n_jobs, ncpu)
_BMA_FITTER = self
print(f'\t\t\tFITTING {len(self._grids)} MODELS, '
f'{n_jobs} CONCURRENTLY')
with Pool(n_jobs) as pool:
for gr, err in pool.map(_bma_grid_worker,
range(len(self._grids))):
if err is not None:
dump_out = self.out_folder + '/' + gr + '_DUMP.pkl'
DynestyError(dump_out, gr, err).__raise__()
else:
for intp, gr in zip(self._interpolators, self._grids):
interpolator = intp
self.grid = gr
out_file = self.out_folder + '/' + gr + '_out.pkl'
print('\t\t\tFITTING MODEL : ' + gr)
try:
self.fit_dynesty(out_file=out_file)
except ValueError as e:
dump_out = self.out_folder + '/' + gr + '_DUMP.pkl'
pickle.dump(self.sampler.results, open(dump_out, 'wb'))
DynestyError(dump_out, gr, e).__raise__()
continue
# Now that the fitting finished, read the outputs and average
# the posteriors
outs = []
for g in self._grids:
in_folder = f'{self.out_folder}/{g}_out.pkl'
outs.append(in_folder)
# with open(in_folder, 'rb') as out:
# outs.append(pickle.load(out))
c = np.random.choice(self.colors)
avgd = self.bayesian_model_average(outs, self._grids, self._norm,
self.n_samples, c)
self.save_bma(avgd)
elapsed_time = execution_time(self.start)
end(self.coordinator, elapsed_time, self.out_folder,
'Bayesian Model Averaging', self.norm)
pass
def _bma_dynesty(self, intp, grid):
global interpolator
interpolator = intp
# Parallel parallelized routine experiment
if self.experimental:
if self._dynamic:
with Pool(self._threads) as executor:
sampler = dynesty.DynamicNestedSampler(
dynesty_loglike_bma, pt_dynesty, self.ndim,
bound=self._bound, sample=self._sample, pool=executor,
queue_size=self._threads, logl_args=([intp])
)
sampler.run_nested(dlogz_init=self._dlogz,
nlive_batch=self._nlive,
wt_kwargs={'pfrac': .95})
else:
with Pool(self._threads) as executor:
sampler = dynesty.NestedSampler(
dynesty_loglike_bma, pt_dynesty, self.ndim,
nlive=self._nlive, bound=self._bound,
sample=self._sample, pool=executor,
queue_size=self._threads, logl_args=([intp])
)
sampler.run_nested(dlogz=self._dlogz)
elif self._dynamic:
sampler = dynesty.DynamicNestedSampler(
dynesty_loglike_bma, pt_dynesty, self.ndim,
bound=self._bound, sample=self._sample, logl_args=([intp])
)
sampler.run_nested(dlogz_init=self._dlogz,
nlive_init=self._nlive,
wt_kwargs={'pfrac': .95})
else:
try:
self.sampler = dynesty.NestedSampler(
dynesty_loglike_bma, pt_dynesty, self.ndim,
nlive=self._nlive, bound=self._bound,
sample=self._sample,
logl_args=([intp])
)
self.sampler.run_nested(dlogz=self._dlogz)
except Error:
dump_out = self.out_folder + '/' + grid + '_DUMP.pkl'
pickle.dump(self.sampler.results, open(dump_out, 'wb'))
er = DynestyError(dump_out, grid)
er.log(self.out + '/output.log')
er.__raise__()
results = self.sampler.results
out_file = self.out_folder + '/' + grid + '_out.pkl'
self.save(out_file, results=results)
pass
[docs]
def fit_multinest(self, out_file=None):
"""Run MultiNest."""
# Set up some globals
global mask, flux, flux_er, filts, wave
mask = star.filter_mask
flux = star.flux[mask]
flux_er = star.flux_er[mask]
filts = star.filter_names[mask]
wave = star.wave[mask]
path = self.out_folder + '/mnest/'
create_dir(path) # Create multinest path.
pymultinest.run(
multinest_log_like, pt_multinest, self.ndim,
n_params=self.ndim,
sampling_efficiency=0.8,
evidence_tolerance=self._dlogz,
n_live_points=self._nlive,
outputfiles_basename=path + 'chains',
max_modes=100,
verbose=self.verbose,
resume=False
)
if out_file is None:
out_file = f'{self.out_folder}/{self._grid}_out.pkl'
self.save(out_file=out_file)
pass
[docs]
def fit_dynesty(self, out_file=None):
"""Run dynesty."""
# Set up some globals
global mask, flux, flux_er, filts, wave
mask = star.filter_mask
flux = star.flux[mask]
flux_er = star.flux_er[mask]
filts = star.filter_names[mask]
wave = star.wave[mask]
if self._dynamic:
if self._threads > 1:
with Pool(self._threads) as executor:
self.sampler = dynesty.DynamicNestedSampler(
dynesty_log_like, pt_dynesty, self.ndim,
bound=self._bound, sample=self._sample,
pool=executor, walks=self.walks,
queue_size=self._threads
)
self.sampler.run_nested(dlogz_init=self._dlogz,
nlive_init=self._nlive,
wt_kwargs={'pfrac': 1})
else:
self.sampler = dynesty.DynamicNestedSampler(
dynesty_log_like, pt_dynesty, self.ndim, walks=self.walks,
bound=self._bound, sample=self._sample
)
self.sampler.run_nested(dlogz_init=self._dlogz,
nlive_init=self._nlive,
wt_kwargs={'pfrac': 1})
else:
if self._threads > 1:
with Pool(self._threads) as executor:
self.sampler = dynesty.NestedSampler(
dynesty_log_like, pt_dynesty, self.ndim,
nlive=self._nlive, bound=self._bound,
sample=self._sample,
pool=executor, walks=self.walks,
queue_size=self._threads,
)
self.sampler.run_nested(dlogz=self._dlogz)
else:
self.sampler = dynesty.NestedSampler(
dynesty_log_like, pt_dynesty, self.ndim, walks=self.walks,
nlive=self._nlive, bound=self._bound,
sample=self._sample
)
self.sampler.run_nested(dlogz=self._dlogz)
results = self.sampler.results
if out_file is None:
out_file = f'{self.out_folder}/{self._grid}_out.pkl'
self.save(out_file, results=results)
pass
[docs]
def save(self, out_file, results=None):
"""Save multinest/dynesty output and relevant information.
Saves a dictionary as a pickle file. The dictionary contains the
following:
lnZ: The global evidence.
lnZerr: The global evidence error.
posterior_samples: A dictionary containing the samples of each
parameter (even if it's fixed), the evidence,
log likelihood, the prior, and the posterior
for each set of sampled parameters.
fixed: An array with the fixed parameter values.
coordinator : An array with the status of each parameter (1 for fixed
0 for free)
best_fit: The best fit is chosen to be the median of each sample.
It also includes the log likelihood of the best fit.
star: The Star object containing the information of the star (name,
magnitudes, fluxes, coordinates, etc)
engine: The fitting engine used (i.e. MultiNest or Dynesty)
Also creates a log file with the best fit parameters and 1 sigma
error bars.
"""
out = dict()
logdat = '#Parameter\tmedian\tupper\tlower\t3sig_CI\n'
log_out = self.out_folder + '/' + 'best_fit.dat'
if self._engine == 'multinest':
lnz, lnzer, posterior_samples = self.multinest_results(
self.out_folder,
self.ndim
)
else:
lnz, lnzer, posterior_samples = self.dynesty_results(results)
mask = self.star.filter_mask
# Save global evidence
if self._engine == 'dynesty':
out['dynesty'] = results
out['global_lnZ'] = lnz
out['global_lnZerr'] = lnzer
# Create raw samples holder
out['posterior_samples'] = dict()
j = 0
k = 0 # filter counter
for i, param in enumerate(order):
if not self.coordinator[i]:
samples = posterior_samples[:, j]
if 'noise' in param:
filt = self.star.filter_names[mask][k]
flx = self.star.flux[mask][k]
_, samples = flux_to_mag(flx, samples, filt)
k += 1
out['posterior_samples'][param] = samples
j += 1
else:
out['posterior_samples'][param] = self.fixed[i]
# Save loglike, priors and posteriors.
out['posterior_samples']['loglike'] = np.zeros(
posterior_samples.shape[0]
)
# If normalization constant was fitted, create a distribution of radii
# only if there's a distance available.
if use_norm and star.dist != -1:
rad = self._get_rad(
out['posterior_samples']['norm'], star.dist, star.dist_e
)
out['posterior_samples']['rad'] = rad
# Create a distribution of masses.
logg_samp = out['posterior_samples']['logg']
rad_samp = out['posterior_samples']['rad']
mass_samp = self._get_mass(logg_samp, rad_samp)
out['posterior_samples']['grav_mass'] = mass_samp
# Create a distribution of luminosities.
teff_samp = out['posterior_samples']['teff']
lum_samp = self._get_lum(teff_samp, rad_samp)
out['posterior_samples']['lum'] = lum_samp
# Create a distribution of angular diameters.
if not use_norm:
dist_samp = out['posterior_samples']['dist']
ad_samp = self._get_angular_diameter(rad_samp, dist_samp)
out['posterior_samples']['AD'] = ad_samp
for i in range(posterior_samples.shape[0]):
theta = build_params(
posterior_samples[i, :], flux, flux_er, filts, coordinator,
fixed, self.norm)
out['posterior_samples']['loglike'][i] = log_likelihood(
theta, flux, flux_er, wave, filts, interpolator, self.norm,
av_law)
# Best fit
# The logic is as follows:
# Calculate KDE for each marginalized posterior distributions
# Find peak
# peak is best fit.
# do only if not bma
if not self.bma:
out['best_fit_averaged'] = dict()
out['uncertainties_averaged'] = dict()
out['confidence_interval_averaged'] = dict()
best_theta = np.zeros(order.shape[0])
for i, param in enumerate(order):
if not self.coordinator[i]:
if 'noise' in param:
continue
samp = out['posterior_samples'][param]
if param == 'z':
logdat = out_filler(samp, logdat, param, '[Fe/H]', out)
elif param == 'norm':
logdat = out_filler(samp, logdat, param, '(R/D)^2',
out, fmt='e')
if star.dist != 1:
logdat = out_filler(
out['posterior_samples']['rad'], logdat, 'rad',
'R', out
)
else:
logdat = out_filler(samp, logdat, param, param, out)
else:
logdat = out_filler(
0, logdat, param, param, out, fixed=self.fixed[i]
)
best_theta[i] = out['best_fit_averaged'][param]
# Add derived mass to best fit dictionary.
samp = out['posterior_samples']['grav_mass']
logdat = out_filler(
samp, logdat, 'grav_mass', 'grav_mass', out
)
# Add derived luminosity to best fit dictionary.
samp = out['posterior_samples']['lum']
logdat = out_filler(samp, logdat, 'lum', 'lum', out)
# Add derived angular diameter to best fit dictionary.
if not use_norm:
samp = out['posterior_samples']['AD']
logdat = out_filler(samp, logdat, 'AD', 'AD', out)
for i, param in enumerate(order):
if not self.coordinator[i]:
if 'noise' not in param:
continue
samp = out['posterior_samples'][param]
logdat = out_filler(samp, logdat, param,
param, out, fmt='f')
# Fill in best loglike, prior and posterior.
out['best_fit_averaged']['loglike'] = log_likelihood(
best_theta, flux, flux_er, wave,
filts, interpolator, self.norm, av_law
)
# Spectral type
# Load Mamajek spt table
mamajek_spt = np.loadtxt(
filesdir + '/mamajek_spt.dat', dtype=str, usecols=[0])
mamajek_temp = np.loadtxt(
filesdir + '/mamajek_spt.dat', usecols=[1])
# Find spt
spt_idx = np.argmin(abs(mamajek_temp - out['best_fit_averaged']['teff']))
spt = mamajek_spt[spt_idx]
out['spectral_type'] = spt
# Utilities for plotting.
out['fixed'] = self.fixed
out['coordinator'] = self.coordinator
out['star'] = self.star
out['engine'] = self._engine
out['norm'] = self.norm
out['model_grid'] = self.grid
out['av_law'] = av_law
if not self.bma:
with open(log_out, 'w') as logfile:
logfile.write(logdat)
pickle.dump(out, open(out_file, 'wb'))
pass
[docs]
def save_bma(self, avgd):
"""Save BMA output and relevant information.
Saves a dictionary as a pickle file. The dictionary contains the
following:
lnZ : The global evidences.
posterior_samples : A dictionary containing the samples of each
parameter (even if it's fixed), the evidence,
log likelihood, the prior, and the posterior
for each set of sampled parameters.
fixed : An array with the fixed parameter values.
coordinator : An array with the status of each parameter (1 for fixed
0 for free)
best_fit : The best fit is chosen to be the median of each sample.
It also includes the log likelihood of the best fit.
star : The Star object containing the information of the star (name,
magnitudes, fluxes, coordinates, etc)
Also creates a log file with the best fit parameters, 1 sigma
error bars and 3 sigma CIs.
"""
out = dict()
logdat_samples = '#Parameter\tmedian\tupper\tlower\t3sig_low\t3sig_up\n'
logdat_average = '#Parameter\tmedian\tupper\tlower\t3sig_low\t3sig_up\n'
log_out_samples = f'{self.out_folder}/best_fit_sample.dat'
log_out_average = f'{self.out_folder}/best_fit_average.dat'
prob_out = f'{self.out_folder}/model_probabilities.dat'
synth_out = f'{self.out_folder}/synthetic_fluxes.dat'
# Save global evidence of each model.
out['lnZ'] = avgd['evidences']
# Save original samples.
out['originals'] = avgd['originals']
# Save weights.
out['weights'] = avgd['weights']
# Create raw samples holder
out['weighted_samples'] = dict()
out['weighted_average'] = dict()
j = 0
for i, par in enumerate(order):
if not self.coordinator[i]:
out['weighted_samples'][par] = avgd['weighted_samples'][par]
out['weighted_average'][par] = avgd['weighted_average'][par]
j += 1
else:
out['weighted_samples'][par] = self.fixed[i]
out['weighted_average'][par] = self.fixed[i]
# If normalization constant was fitted, create a distribution of radii.
if use_norm and star.dist != -1:
rad_sampled = self._get_rad(
out['weighted_samples']['norm'], star.dist, star.dist_e
)
rad_averageed = self._get_rad(
out['weighted_average']['norm'], star.dist, star.dist_e
)
out['weighted_samples']['rad'] = rad_sampled
out['weighted_average']['rad'] = rad_averageed
# Create a distribution of masses
logg_samp = out['weighted_samples']['logg']
logg_average = out['weighted_average']['logg']
rad_samp = out['weighted_samples']['rad']
rad_average = out['weighted_average']['rad']
mass_samp = self._get_mass(logg_samp, rad_samp)
mass_average = self._get_mass(logg_average, rad_average)
out['weighted_samples']['grav_mass'] = mass_samp
out['weighted_average']['grav_mass'] = mass_average
# Create a distribution of luminosities.
teff_samp = out['weighted_samples']['teff']
teff_average = out['weighted_average']['teff']
lum_samp = self._get_lum(teff_samp, rad_samp)
lum_average = self._get_lum(teff_average, rad_average)
out['weighted_samples']['lum'] = lum_samp
out['weighted_average']['lum'] = lum_average
# Create a distribution of angular diameters.
if not use_norm:
dist_samp = out['weighted_samples']['dist']
dist_average = out['weighted_average']['dist']
ad_samp = self._get_angular_diameter(rad_samp, dist_samp)
ad_average = self._get_angular_diameter(rad_average, dist_average)
out['weighted_samples']['AD'] = ad_samp
out['weighted_average']['AD'] = ad_average
# Best fit
# The logic is as follows:
# Calculate KDE for each marginalized posterior distributions
# Find peak
# peak is best fit.
out['best_fit_samples'] = dict()
out['uncertainties_samples'] = dict()
out['confidence_interval_samples'] = dict()
out['best_fit_averaged'] = dict()
out['uncertainties_averaged'] = dict()
out['confidence_interval_averaged'] = dict()
for i, param in enumerate(order):
if not self.coordinator[i]:
if 'noise' in param:
continue
samp = out['weighted_samples'][param]
sampw = out['weighted_average'][param]
if param == 'z':
logdat_samples = out_filler(samp, logdat_samples, param,
'[Fe/H]', out, method='samples')
logdat_average = out_filler(sampw, logdat_average, param,
'[Fe/H]', out,
method='averaged')
elif param == 'norm':
logdat_samples = out_filler(samp, logdat_samples, param,
'(R/D)^2', out, fmt='e',
method='samples')
logdat_average = out_filler(sampw, logdat_average, param,
'(R/D)^2', out, fmt='e',
method='averaged')
if star.dist != 1:
logdat_samples = out_filler(
out['weighted_samples']['rad'], logdat_samples,
'rad', 'R', out, method='samples'
)
logdat_average = out_filler(
out['weighted_average']['rad'], logdat_average,
'rad', 'R', out, method='averaged'
)
else:
logdat_samples = out_filler(samp, logdat_samples, param,
param, out, method='samples')
logdat_average = out_filler(sampw, logdat_average, param,
param, out, method='averaged')
else:
logdat_samples = out_filler(
0, logdat_samples, param, param, out, fixed=self.fixed[i],
method='samples'
)
logdat_average = out_filler(
0, logdat_average, param, param, out, fixed=self.fixed[i],
method='averaged'
)
# Add derived mass to best fit dictionary.
samp = out['weighted_samples']['grav_mass']
sampw = out['weighted_average']['grav_mass']
logdat_samples = out_filler(samp, logdat_samples, 'grav_mass',
'grav_mass', out, method='samples')
logdat_average = out_filler(sampw, logdat_average, 'grav_mass',
'grav_mass', out, method='averaged')
# Add derived luminosity to best fit dictionary.
samp = out['weighted_samples']['lum']
sampw = out['weighted_average']['lum']
logdat_samples = out_filler(samp, logdat_samples, 'lum', 'lum', out,
method='samples')
logdat_average = out_filler(sampw, logdat_average, 'lum', 'lum', out,
method='averaged')
# Add derived angular diameter to best fit dictionary.
if not use_norm:
samp = out['weighted_samples']['AD']
sampw = out['weighted_average']['AD']
logdat_samples = out_filler(samp, logdat_samples, 'AD', 'AD', out,
method='samples')
logdat_average = out_filler(sampw, logdat_average, 'AD', 'AD', out,
method='averaged')
# Add estimated age to best fit dictionary. This is done with the wider
# sampled distribution instead of the averaged one in order to save time
if self.estimate_age:
age_samp, mass_samp, eep_samp = self._run_estimate_age(
out['best_fit_samples'],
out['uncertainties_samples'],
c=choice(self.colors)
)
# Create new thingy for MIST samples. Sadly now everything done before
# this update will be incompatible :(
out['mist_samples'] = dict()
out['mist_samples']['age'] = age_samp
out['mist_samples']['iso_mass'] = mass_samp
out['mist_samples']['eep'] = eep_samp
logdat_samples = out_filler(age_samp, logdat_samples, 'age', 'age', out,
method='samples')
logdat_samples = out_filler(mass_samp, logdat_samples, 'iso_mass',
'iso_mass', out, method='samples')
logdat_samples = out_filler(eep_samp, logdat_samples, 'eep', 'eep', out,
method='samples')
# Ugly... but faster than doing the kde 2x times...
age = out['best_fit_samples']['age']
age_unc = out['uncertainties_samples']['age']
age_ci = out['confidence_interval_samples']['age']
iso_mass = out['best_fit_samples']['iso_mass']
iso_mass_unc = out['uncertainties_samples']['iso_mass']
iso_mass_ci = out['confidence_interval_samples']['iso_mass']
eep = out['best_fit_samples']['eep']
eep_unc = out['uncertainties_samples']['eep']
eep_ci = out['confidence_interval_samples']['eep']
out['best_fit_averaged']['age'] = age
out['best_fit_averaged']['iso_mass'] = iso_mass
out['best_fit_averaged']['eep'] = eep
out['uncertainties_averaged']['age'] = age_unc
out['uncertainties_averaged']['iso_mass'] = iso_mass_unc
out['uncertainties_averaged']['eep'] = eep_unc
out['confidence_interval_averaged']['age'] = age_ci
out['confidence_interval_averaged']['iso_mass'] = iso_mass_ci
out['confidence_interval_averaged']['eep'] = eep_ci
logdat_average += f'age\t{age:.4f}\t'
logdat_average += f'{age_unc[1]:.4f}\t{age_unc[0]:.4f}\t'
logdat_average += f'{age_ci[0]:.4f}\t{age_ci[1]}\n'
logdat_average += f'iso_mas\t{iso_mass:.4f}\t'
logdat_average += f'{iso_mass_unc[1]:.4f}\t{iso_mass_unc[0]:.4f}\t'
logdat_average += f'{iso_mass_ci[0]:.4f}\t{iso_mass_ci[1]}\n'
logdat_average += f'eep\t{eep:.4f}\t'
logdat_average += f'{eep_unc[1]:.4f}\t{eep_unc[0]:.4f}\t'
logdat_average += f'{eep_ci[0]:.4f}\t{eep_ci[1]}\n'
else:
out['mist_samples'] = {}
###
probdat = ''
max_prob = 0
max_prob_mod = ''
for k in avgd['weights'].keys():
if avgd['weights'][k] > max_prob:
max_prob = avgd['weights'][k]
max_prob_mod = k
probdat += f'{k}_probability\t{avgd["weights"][k]:.4f}\n'
# Get synthetic mag and fluxes for highest probability model.
synth_mod = 'btsettl'
intp = self.load_interpolator(synth_mod)
ogteff = avgd['originals'][max_prob_mod]['teff']
oglogg = avgd['originals'][max_prob_mod]['logg']
ogfeh = avgd['originals'][max_prob_mod]['z']
fluxes = np.zeros((len(oglogg), len(filter_names)))
for i, t, g, z in zip(range(len(oglogg)), ogteff, oglogg, ogfeh):
fluxes[i, :] = get_interpolated_flux(t, g, z, filter_names, intp)
synthdat = 'Filter\tFlux\n'
for i, filt in enumerate(filter_names):
samp = fluxes[:, i]
# xx, pdf = estimate_pdf(samp)
# cdf = estimate_cdf(samp, hdr=True)
b, _, _ = credibility_interval(samp, alpha=1)
synthdat += f'{filt}\t{b:.6e}\n'
for i, param in enumerate(order):
if not self.coordinator[i]:
if 'noise' not in param:
continue
samp = out['weighted_samples'][param]
sampw = out['weighted_average'][param]
logdat_samples = out_filler(samp, logdat_samples, param, param,
out, fmt='f', method='samples')
logdat_average = out_filler(sampw, logdat_average, param, param,
out, fmt='f', method='averaged')
out['fixed'] = self.fixed
out['coordinator'] = self.coordinator
out['star'] = self.star
out['norm'] = self.norm
out['engine'] = 'Bayesian Model Averaging'
out['av_law'] = av_law
# Spectral type
# Load Mamajek spt table
mamajek_spt, mamajek_temp = np.loadtxt(f'{filesdir}/mamajek_spt.dat',
dtype=str, usecols=[0, 1],
unpack=True)
mamajek_temp = mamajek_temp.astype(float)
# Find spt
spt_idx = np.argmin(
abs(mamajek_temp - out['best_fit_averaged']['teff']))
spt = mamajek_spt[spt_idx]
out['spectral_type'] = spt
out_file = f'{self.out_folder}/BMA.pkl'
with open(log_out_samples, 'w') as logfile:
logfile.write(logdat_samples)
with open(log_out_average, 'w') as logfile:
logfile.write(logdat_average)
with open(prob_out, 'w') as logfile:
logfile.write(probdat)
with open(synth_out, 'w') as logfile:
logfile.write(synthdat)
pickle.dump(out, open(out_file, 'wb'))
self.out = out
nc_file = f'{self.out_folder}/ariadne_result.nc'
self.to_netcdf(nc_file)
[docs]
@staticmethod
def bayesian_model_average(outputs, grids, norm, nsamples, c='white'):
"""Perform Bayesian Model Averaging.
Parameters
----------
outputs: array_like
An array or list of output pickle files from the individual model
fits.
grids: array_like
An array or list of the model grids used to model.
norm: bool
Flag indicating if the normalization factor was fit for.
nsamples: int
The number of samples to sample from the averaged distribution.
c: str, optional
Termcolor color.
"""
evidences = []
post_samples = []
model_posteriors = []
# Read and extract model posterior information.
for o in outputs:
with open(o, 'rb') as f:
model_posteriors.append(pickle.load(f))
# Extract the evidences of each model.
for o in model_posteriors:
evidences.append(o['global_lnZ'])
post_samples.append(o['posterior_samples'])
# Convert evidences to weights/probabilities via a softmax of the
# log-evidences. Shift by the MAX (not the min) so the largest exponent
# is 0: softmax is shift-invariant, so the weights are unchanged in the
# normal regime, but shifting by the min lets exp() overflow to inf ->
# nan when one grid's log-evidence dominates (delta lnZ > ~709).
evidences = np.array(evidences)
weights = evidences - evidences.max()
weights = [np.exp(e) / np.exp(weights).sum() for e in weights]
weights = np.array(weights)
# We're not averaging these
ban = ['loglike', 'priors', 'posteriors']
if norm: # if normalization was used, we won't average the radius
ban.append('rad')
# Create an output dictionary for the averaged samples.
out = dict()
out['originals'] = dict()
out['weights'] = dict()
# Populate the dict with the probabilities and individual model
# posteriors
for i, o in enumerate(model_posteriors):
out['weights'][o['model_grid']] = weights[i]
out['originals'][o['model_grid']] = o['posterior_samples']
out['weighted_samples'] = dict()
out['weighted_average'] = dict()
print(colored('\t\t*** AVERAGING POSTERIOR SAMPLES ***', c))
for k in tqdm(post_samples[0].keys()):
if k in ban: # Skip things that are not main parameters.
continue
try: # Skip fixed params.
len(post_samples[0][k])
except TypeError:
continue
traces = []
extended_weights = []
out['weighted_samples'][k] = np.zeros(nsamples)
out['weighted_average'][k] = np.zeros(nsamples)
for i, o in enumerate(post_samples):
# This is for the weighted sampling.
traces.append(o[k])
extended_weights.append(np.ones(len(o[k])) * weights[i])
# This is for the weighted averaging.
# The weighted averaging consists of taking the weighted
# average of the posterior samples, supersampled to maked them
# coincide in length.
weighted_samples = choice(o[k], nsamples) * weights[i]
out['weighted_average'][k] += weighted_samples
# Do the weighted sampling. For this we're going to estimate the
# KDE of the 'master' posterior built by taking random samples
# from each model where the number of samples is proportional
# to the model's relative probability. This is equivalent
# to taking the KDE of each model and then do the weighted average
avg_kde = st.gaussian_kde(np.concatenate(traces),
weights=np.concatenate(extended_weights))
out['weighted_samples'][k] = avg_kde.resample(nsamples)[0]
# Now we save the evidences.
out['evidences'] = dict()
for e, g in zip(evidences, grids):
out['evidences'][g] = e
return out
[docs]
@staticmethod
def multinest_results(out_folder, ndim):
"""Extract posterior samples, global evidence and its error."""
path = f'{out_folder}/mnest/chains'
output = pymultinest.Analyzer(outputfiles_basename=path, n_params=ndim)
posterior_samples = output.get_equal_weighted_posterior()[:, :-1]
lnz = output.get_stats()['global evidence']
lnzer = output.get_stats()['global evidence error']
return lnz, lnzer, posterior_samples
[docs]
@staticmethod
def dynesty_results(results):
"""Extract posterior samples, global evidence and its error."""
weights = np.exp(results['logwt'] - results['logz'][-1])
posterior_samples = resample_equal(results.samples, weights)
lnz = results.logz[-1]
lnzer = results.logzerr[-1]
return lnz, lnzer, posterior_samples
@staticmethod
def _get_mass(logg, rad):
"""Calculate mass from logg and radius."""
# Solar logg = 4.437
# g = g_Sol * M / R**2
mass = logg + 2 * np.log10(rad) - 4.437
mass = 10 ** mass
return mass
@staticmethod
def _get_lum(teff, rad):
sb = sigma_sb.to(u.solLum / u.K ** 4 / u.solRad ** 2).value
L = 4 * np.pi * rad ** 2 * sb * teff ** 4
return L
@staticmethod
def _get_rad(samples, dist, dist_e):
"""Calculate radius from the normalization constant and distance."""
norm = samples
# Create a synthetic distribution for distance.
# N = (R / D) ** 2
b, a = (np.inf - dist) / dist_e, (0 - dist) / dist_e
d = st.truncnorm(loc=dist, scale=dist_e, a=a, b=b).rvs(
size=norm.shape[0])
n = np.sqrt(norm)
r = n * d # This is in pc
r *= u.pc.to(u.solRad) # Transform to Solar radii
return r
@staticmethod
def _get_angular_diameter(rad, dist):
diameter = 2 * rad
ad = (diameter / (dist * u.pc.to(u.solRad))) * u.rad.to(u.marcsec)
return ad
[docs]
@staticmethod
def load_interpolator(model):
"""Load a DFInterpolator."""
if model.lower() == 'phoenix':
with open(gridsdir + '/Phoenixv2_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
elif model.lower() == 'btsettl':
with open(gridsdir + '/BTSettl_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
elif model.lower() == 'btnextgen':
with open(gridsdir + '/BTNextGen_DF.pkl', 'rb') as inp:
df = DFInterpolator(pd.read_pickle(inp))
elif model.lower() == 'btcond':
with open(gridsdir + '/BTCond_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
elif model.lower() == 'ck04':
with open(gridsdir + '/CK04_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
elif model.lower() == 'kurucz':
with open(gridsdir + '/Kurucz_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
elif model.lower() == 'coelho':
with open(gridsdir + '/Coelho_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
elif model.lower() == 'bosz':
with open(gridsdir + '/BOSZ_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
elif model.lower() == 'sphinx':
with open(gridsdir + '/SPHINX_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
elif model.lower() == 'tlusty':
with open(gridsdir + '/TLUSTY_DF.pkl', 'rb') as intp:
df = DFInterpolator(pd.read_pickle(intp))
return df
def _run_estimate_age(self, bf, unc, c='white'):
"""Estimate age using MIST isochrones.
Parameters
----------
bf: dict
A dictionary with the best fit parameters.
unc: dict
A dictionary with the uncertainties.
"""
print(
colored(
'\t\t*** ESTIMATING AGE AND MASS USING MIST ISOCHRONES ***', c
)
)
params = dict() # params for isochrones.
for i, k in enumerate(order):
if k == 'logg' or 'noise' in k:
continue
if k == 'teff':
par = 'Teff'
if k == 'z':
par = 'feh'
if k == 'dist':
par = 'distance'
if k == 'rad':
par = 'radius'
if k == 'Av':
par = 'AV'
if not self.coordinator[i]:
if k != 'norm':
params[par] = (bf[k], max(unc[k]))
if k == 'norm' and self.star.dist != -1:
params['distance'] = (self.star.dist, self.star.dist_e)
if par == 'distance':
err = max(unc[k])
params['parallax'] = (1000 / bf[k], 1000 * err / bf[k] ** 2)
else:
continue
params['mass'] = (bf['grav_mass'], max(unc['grav_mass']))
if self.star.lum != 0 and self.star.lum_e != 0:
params['logL'] = (np.log10(bf['lum']),
abs(np.log10(max(unc['lum']))))
mags = self.star.mags[iso_mask == 1]
mags_e = self.star.mag_errs[iso_mask == 1]
used_bands = []
for m, e, b in zip(mags, mags_e, iso_bands):
if m != 0:
params[b] = (m, e)
used_bands.append(b)
age_samp, mass_samp, eep_samp = estimate(
used_bands, params, logg=False,
out_folder=self.out_folder,
threads=self._threads, dlogz=self.isochrone_dlogz)
return age_samp, mass_samp, eep_samp
[docs]
def to_dict(self):
"""Return BMA results as a structured dict matching the ecosystem spec.
Returns a dictionary with keys ``posterior``, ``observed_data``, and
``sample_stats``, each containing numpy arrays ready for conversion to
``arviz.InferenceData``.
Raises
------
RuntimeError
If the fitter has not been run yet (``self.out`` is not populated).
"""
if not hasattr(self, 'out') or self.out is None:
raise RuntimeError(
'No results available. Run the fitter before calling to_dict().'
)
out = self.out
ws = out.get('weighted_samples', {})
# Determine number of draws from any free parameter
n_draws = None
for v in ws.values():
if hasattr(v, '__len__') and len(v) > 1:
n_draws = len(v)
break
if n_draws is None:
n_draws = 100_000 # fallback
def _to_chain(arr_or_scalar):
"""Reshape to (1, n_draws) for single-chain arviz convention."""
if hasattr(arr_or_scalar, '__len__') and len(arr_or_scalar) > 1:
return np.array(arr_or_scalar).reshape(1, -1)
return np.full((1, n_draws), float(arr_or_scalar))
# Map internal names → ecosystem spec names
param_map = {
'teff': 'Teff',
'logg': 'logg',
'z': 'feh',
'rad': 'radius',
'lum': 'luminosity',
'dist': 'distance',
'Av': 'Av',
}
posterior = {}
for internal, spec in param_map.items():
if internal in ws:
posterior[spec] = _to_chain(ws[internal])
# Observed photometry
mask = self.star.filter_mask
observed_data = {
'wavelength': self.star.wave[mask], # already in micron
'flux': self.star.flux[mask],
'flux_err': self.star.flux_er[mask],
}
# Model evidence and weights
weights = out.get('weights', {})
lnZ = out.get('lnZ', {})
model_names = list(weights.keys())
weight_vals = np.array([weights[m] for m in model_names])
lnZ_vals = np.array([lnZ.get(m, 0.0) for m in model_names])
bma_log_evidence = float(np.sum(weight_vals * lnZ_vals))
sample_stats = {
'log_evidence': bma_log_evidence,
'model_names': model_names,
'model_weights': weight_vals,
}
# Per-model posteriors from originals
originals = out.get('originals', {})
model_posteriors = {}
for model_name, model_samples in originals.items():
n_model = None
for v in model_samples.values():
if hasattr(v, '__len__') and len(v) > 1:
n_model = len(v)
break
if n_model is None:
continue
def _to_model_chain(a, nd=n_model):
if hasattr(a, '__len__') and len(a) > 1:
return np.array(a).reshape(1, -1)
return np.full((1, nd), float(a))
mp = {}
for internal, spec in param_map.items():
if internal in model_samples:
mp[spec] = _to_model_chain(model_samples[internal])
if 'loglike' in model_samples:
mp['log_likelihood'] = _to_model_chain(
model_samples['loglike'])
model_posteriors[model_name] = mp
# MIST isochrone samples
mist = out.get('mist_samples', {})
mist_posterior = {}
for k in ('age', 'iso_mass', 'eep'):
if k in mist and hasattr(mist[k], '__len__') and len(mist[k]) > 1:
mist_posterior[k] = _to_chain(mist[k])
# Filter names and bandwidths for SED reconstruction
observed_data['filter_names'] = np.array(
self.star.filter_names[mask].tolist()
if hasattr(self.star.filter_names[mask], 'tolist')
else list(self.star.filter_names[mask])
)
observed_data['bandwidths'] = self.star.bandpass[mask]
# Model SED fluxes (injected by worker from SEDPlotter)
model_sed = {}
if 'model_sed' in out:
ms = out['model_sed']
if 'model_flux' in ms:
model_sed['model_flux'] = np.array(ms['model_flux'])
# Summary statistics for quick display
summary = {}
for src_key in ('best_fit_averaged', 'uncertainties_averaged',
'confidence_interval_averaged'):
if src_key in out:
summary[src_key] = {}
for k, v in out[src_key].items():
if hasattr(v, '__len__'):
summary[src_key][k] = np.array(
[float(x) for x in v])
else:
summary[src_key][k] = float(v)
return {
'posterior': posterior,
'observed_data': observed_data,
'sample_stats': sample_stats,
'model_posteriors': model_posteriors,
'mist_posterior': mist_posterior,
'model_sed': model_sed,
'summary': summary,
}
[docs]
def to_netcdf(self, path):
"""Export BMA results as arviz DataTree in netCDF4 format.
Produces an ``arviz.InferenceData``-compatible ``.nc`` file
following the ecosystem output spec.
Parameters
----------
path : str
Output file path (e.g. ``'ariadne_result.nc'``).
"""
import arviz as az
d = self.to_dict()
ss = d['sample_stats']
data = {
'posterior': d['posterior'],
'observed_data': d['observed_data'],
'constant_data': {
'log_evidence': np.array(ss['log_evidence']),
'model_weights': ss['model_weights'],
'model_names': np.array(ss['model_names']),
},
}
# Add per-model posteriors as separate groups
for model_name, mp in d.get('model_posteriors', {}).items():
data[f'posterior_{model_name}'] = mp
# Add MIST isochrone posteriors
if d.get('mist_posterior'):
data['posterior'].update(d['mist_posterior'])
# Add model SED fluxes to observed_data
if d.get('model_sed') and 'model_flux' in d['model_sed']:
data['observed_data']['model_flux'] = d['model_sed']['model_flux']
# Add summary stats to constant_data
for stat_key in ('best_fit_averaged', 'uncertainties_averaged',
'confidence_interval_averaged'):
if stat_key in d.get('summary', {}):
for param, val in d['summary'][stat_key].items():
key = f'{stat_key}__{param}'
data['constant_data'][key] = np.array(val)
dt = az.from_dict(data)
dt.to_netcdf(path)
logger.info(f'Wrote InferenceData to {path}')
#####################
# Dynesty and multinest wrappers
# Set by fit_bma before forking the grid-parallel Pool; read by the workers.
_BMA_FITTER = None
def _bma_grid_worker(index):
"""Fit one BMA model grid single-threaded and write its ``_out.pkl``.
Runs in a forked worker process: the star photometry, priors, coordinator
and other module globals are inherited from the parent, so only the
per-grid interpolator is swapped in. The fit result is delivered via the
on-disk pickle (the parent reads it back for averaging), so nothing needs
to be pickled back to the parent. Returns ``(grid, None)`` on success or
``(grid, error)`` on failure.
"""
global interpolator
fitter = _BMA_FITTER
intp = fitter._interpolators[index]
grid = fitter._grids[index]
out_file = f'{fitter.out_folder}/{grid}_out.pkl'
# Forked processes inherit the parent's RNG state; reseed so each grid
# draws independently.
np.random.seed()
interpolator = intp
fitter._threads = 1 # grids are the unit of parallelism; no inner Pool
fitter._grid = grid # used by save() for out['model_grid']
try:
fitter.fit_dynesty(out_file=out_file)
except ValueError as e:
dump_out = f'{fitter.out_folder}/{grid}_DUMP.pkl'
pickle.dump(fitter.sampler.results, open(dump_out, 'wb'))
return grid, e
return grid, None
def dynesty_loglike_bma(cube, interpolator):
"""Dynesty log likelihood wrapper for BMA."""
theta = build_params(cube, coordinator, fixed, use_norm)
return log_likelihood(theta, star, interpolator, use_norm, av_law)
def dynesty_log_like(cube):
"""Dynesty log likelihood wrapper."""
theta = build_params(
cube, flux, flux_er, filts, coordinator, fixed, use_norm
)
return log_likelihood(theta, flux, flux_er, wave,
filts, interpolator, use_norm, av_law)
def pt_dynesty(cube):
"""Dynesty prior transform."""
return prior_transform_dynesty(cube, flux, flux_er, filts, prior_dict,
coordinator, use_norm)
def multinest_log_like(cube, ndim, nparams):
"""Multinest log likelihood wrapper."""
theta = [cube[i] for i in range(ndim)]
theta = build_params(
theta, flux, flux_er, filts, coordinator, fixed, use_norm
)
return log_likelihood(theta, flux, flux_er, wave,
filts, interpolator, use_norm, av_law)
def pt_multinest(cube, ndim, nparams):
"""Multinest prior transform."""
prior_transform_multinest(cube, flux, flux_er, filts, prior_dict,
coordinator, use_norm)