__init__.py 10.8 KB
Newer Older
1
# -*- coding: utf-8 -*-
2
# Copyright (C) 2014 Laboratoire d'Astrophysique de Marseille, AMU
3
# Copyright (C) 2013 Centre de données Astrophysiques de Marseille
4 5
# Copyright (C) 2013-2014 Institute of Astronomy
# Copyright (C) 2013-2014 Yannick Roehlly <yannick@iaora.eu>
6
# Licensed under the CeCILL-v2 licence - see Licence_CeCILL_V2-en.txt
7
# Author: Yannick Roehlly, Médéric Boquien & Denis Burgarella
8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27

"""
Probability Density Function analysis module
============================================

This module builds the probability density functions (PDF) of the SED
parameters to compute their moments.

The models corresponding to all possible combinations of parameters are
computed and their fluxes in the same filters as the observations are
integrated. These fluxes are compared to the observed ones to compute the
χ² value of the fitting. This χ² give a probability that is associated with
the model values for the parameters.

At the end, for each parameter, the probability-weighted mean and standard
deviation are computed and the best fitting model (the one with the least
reduced χ²) is given for each observation.

"""

28
from collections import OrderedDict
29
import ctypes
30
import multiprocessing as mp
31 32 33 34 35
from multiprocessing.sharedctypes import RawArray
import time

import numpy as np

36 37
from ...utils import read_table
from .. import AnalysisModule, complete_obs_table
38
from .utils import save_results, analyse_chi2
39
from ...warehouse import SedWarehouse
40
from .workers import sed as worker_sed
41 42
from .workers import init_sed as init_worker_sed
from .workers import init_analysis as init_worker_analysis
43
from .workers import analysis as worker_analysis
44 45
from ..utils import backup_dir
from ...handlers.parameters_handler import ParametersHandler
46

47

48
# Tolerance threshold under which any flux or error is considered as 0.
49
TOLERANCE = 1e-12
50 51 52 53 54


class PdfAnalysis(AnalysisModule):
    """PDF analysis module"""

55
    parameter_list = OrderedDict([
56
        ("variables", (
57
            "cigale_string_list()",
58 59 60
            "List of the physical properties to estimate. Leave empty to "
            "analyse all the physical properties (not recommended when there "
            "are many models).",
61
            ["sfh.sfr", "sfh.sfr10Myrs", "sfh.sfr100Myrs"]
62 63
        )),
        ("save_best_sed", (
64
            "boolean()",
65 66 67
            "If true, save the best SED for each observation to a file.",
            False
        )),
68
        ("save_chi2", (
69
            "boolean()",
70
            "If true, for each observation and each analysed variable save "
71
            "the reduced chi2.",
72 73 74
            False
        )),
        ("save_pdf", (
75
            "boolean()",
76 77
            "If true, for each observation and each analysed variable save "
            "the probability density function.",
78 79
            False
        )),
80
        ("lim_flag", (
81
            "boolean()",
82 83 84
            "If true, for each object check whether upper limits are present "
            "and analyse them.",
            False
85 86
        )),
        ("mock_flag", (
87
            "boolean()",
88 89 90
            "If true, for each object we create a mock object "
            "and analyse them.",
            False
91 92 93
        ))
    ])

94
    def process(self, conf):
95 96
        """Process with the psum analysis.

97 98 99 100 101
        The analysis is done in two steps which can both run on multiple
        processors to run faster. The first step is to compute all the fluxes
        associated with each model as well as ancillary data such as the SED
        information. The second step is to carry out the analysis of each
        object, considering all models at once.
102 103 104

        Parameters
        ----------
105 106
        conf: dictionary
            Contents of pcigale.ini in the form of a dictionary
107 108

        """
109
        np.seterr(invalid='ignore')
110

111 112
        print("Initialising the analysis module... ")

113
        # Rename the output directory if it exists
114
        backup_dir()
115

116
        # Initalise variables from input arguments.
117 118 119 120
        variables = conf['analysis_params']["variables"]
        variables_nolog = [variable[:-4] if variable.endswith('_log') else
                           variable for variable in variables]
        n_variables = len(variables)
121 122 123
        save = {key: conf['analysis_params']["save_{}".format(key)] for key in
                ["best_sed", "chi2", "pdf"]}
        lim_flag = conf['analysis_params']["lim_flag"]
124

125
        filters = [name for name in conf['bands'] if not
126
                   name.endswith('_err')]
127
        n_filters = len(filters)
128 129 130

        # Read the observation table and complete it by adding error where
        # none is provided and by adding the systematic deviation.
131
        obs_table = complete_obs_table(read_table(conf['data_file']),
132
                                       conf['bands'], filters, TOLERANCE,
133
                                       lim_flag)
134
        n_obs = len(obs_table)
135

136
        z = np.array(conf['sed_modules_params']['redshifting']['redshift'])
137 138 139 140 141 142

        # The parameters handler allows us to retrieve the models parameters
        # from a 1D index. This is useful in that we do not have to create
        # a list of parameters as they are computed on-the-fly. It also has
        # nice goodies such as finding the index of the first parameter to
        # have changed between two indices or the number of models.
143
        params = ParametersHandler(conf)
144 145
        n_params = params.size

146
        # Retrieve an arbitrary SED to obtain the list of output parameters
147
        warehouse = SedWarehouse()
148
        sed = warehouse.get_sed(conf['sed_modules'], params.from_index(0))
149 150 151
        info = list(sed.info.keys())
        info.sort()
        n_info = len(info)
152 153
        del warehouse, sed

154 155
        print("Computing the models fluxes...")

156 157 158 159 160 161 162
        # Arrays where we store the data related to the models. For memory
        # efficiency reasons, we use RawArrays that will be passed in argument
        # to the pool. Each worker will fill a part of the RawArrays. It is
        # important that there is no conflict and that two different workers do
        # not write on the same section.
        # We put the shape in a tuple along with the RawArray because workers
        # need to know the shape to create the numpy array from the RawArray.
163
        model_fluxes = (RawArray(ctypes.c_double, n_params * n_filters),
164
                        (n_filters, n_params))
165
        model_variables = (RawArray(ctypes.c_double, n_params * n_variables),
166
                           (n_variables, n_params))
167

168
        initargs = (params, filters, variables_nolog, model_fluxes,
169
                    model_variables, time.time(), mp.Value('i', 0))
170
        if conf['cores'] == 1:  # Do not create a new process
171
            init_worker_sed(*initargs)
172 173
            for idx in range(n_params):
                worker_sed(idx)
174
        else:  # Compute the models in parallel
175
            with mp.Pool(processes=conf['cores'], initializer=init_worker_sed,
176
                         initargs=initargs) as pool:
177
                pool.map(worker_sed, range(n_params))
178

179
        print("\nAnalysing models...")
180

181 182
        # We use RawArrays for the same reason as previously
        analysed_averages = (RawArray(ctypes.c_double, n_obs * n_variables),
Médéric Boquien's avatar
Médéric Boquien committed
183
                             (n_obs, n_variables))
184
        analysed_std = (RawArray(ctypes.c_double, n_obs * n_variables),
Médéric Boquien's avatar
Médéric Boquien committed
185
                        (n_obs, n_variables))
186
        best_fluxes = (RawArray(ctypes.c_double, n_obs * n_filters),
Médéric Boquien's avatar
Médéric Boquien committed
187
                       (n_obs, n_filters))
188 189 190 191 192
        best_parameters = (RawArray(ctypes.c_double, n_obs * n_info),
                           (n_obs, n_info))
        best_chi2 = (RawArray(ctypes.c_double, n_obs), (n_obs))
        best_chi2_red = (RawArray(ctypes.c_double, n_obs), (n_obs))

193
        initargs = (params, filters, variables, z, model_fluxes,
194 195 196 197
                    model_variables, time.time(), mp.Value('i', 0),
                    analysed_averages, analysed_std, best_fluxes,
                    best_parameters, best_chi2, best_chi2_red, save, lim_flag,
                    n_obs)
198
        if conf['cores'] == 1:  # Do not create a new process
199
            init_worker_analysis(*initargs)
200 201
            for idx, obs in enumerate(obs_table):
                worker_analysis(idx, obs)
202
        else:  # Analyse observations in parallel
203 204
            with mp.Pool(processes=conf['cores'],
                         initializer=init_worker_analysis,
205
                         initargs=initargs) as pool:
206
                pool.starmap(worker_analysis, enumerate(obs_table))
207

208
        analyse_chi2(best_chi2_red)
209

210 211
        print("\nSaving results...")

212 213 214
        save_results("results", obs_table['id'], variables, analysed_averages,
                     analysed_std, best_chi2, best_chi2_red, best_parameters,
                     best_fluxes, filters, info)
215

216
        if conf['analysis_params']['mock_flag'] is True:
217 218 219

            print("\nMock analysis...")

220 221 222 223
            # For the mock analysis we do not save the ancillary files
            for k in save:
                save[k] = False

224
            obs_fluxes = np.array([obs_table[name] for name in filters]).T
Médéric Boquien's avatar
Médéric Boquien committed
225 226
            obs_errors = np.array([obs_table[name + "_err"] for name in
                                   filters]).T
227
            mock_fluxes = obs_fluxes.copy()
228 229
            bestmod_fluxes = np.ctypeslib.as_array(best_fluxes[0])
            bestmod_fluxes = bestmod_fluxes.reshape(best_fluxes[1])
230 231 232 233
            wdata = np.where((obs_fluxes > TOLERANCE) &
                             (obs_errors > TOLERANCE))
            mock_fluxes[wdata] = np.random.normal(bestmod_fluxes[wdata],
                                                  obs_errors[wdata])
Médéric Boquien's avatar
Médéric Boquien committed
234

235
            mock_table = obs_table.copy()
236 237
            for idx, name in enumerate(filters):
                mock_table[name] = mock_fluxes[:, idx]
Médéric Boquien's avatar
Médéric Boquien committed
238

239
            initargs = (params, filters, variables, z, model_fluxes,
240 241 242 243
                        model_variables, time.time(), mp.Value('i', 0),
                        analysed_averages, analysed_std, best_fluxes,
                        best_parameters, best_chi2, best_chi2_red, save,
                        lim_flag, n_obs)
244
            if conf['cores'] == 1:  # Do not create a new process
245 246 247 248
                init_worker_analysis(*initargs)
                for idx, mock in enumerate(mock_table):
                    worker_analysis(idx, mock)
            else:  # Analyse observations in parallel
249 250
                with mp.Pool(processes=conf['cores'],
                             initializer=init_worker_analysis,
251 252 253
                             initargs=initargs) as pool:
                    pool.starmap(worker_analysis, enumerate(mock_table))

254
            print("\nSaving results...")
255

256
            save_results("results_mock", mock_table['id'], variables,
257 258 259
                         analysed_averages, analysed_std, best_chi2,
                         best_chi2_red, best_parameters, best_fluxes, filters,
                         info)
260 261

        print("Run completed!")
262

263 264
# AnalysisModule to be returned by get_module
Module = PdfAnalysis