pipeline.py 8.66 KB
Newer Older
1
2
# -*- coding: utf-8 -*-

3
4
5
import os
import numpy as np

6
7
from scipy import optimize

8
9
from mpi4py import MPI

10
11
from keepers import Loggable

12
13
14
15
from imagine.likelihoods import Likelihood
from imagine.magnetic_fields import MagneticFieldFactory
from imagine.observers import Observer
from imagine.priors import Prior
16
from imagine import pymultinest
17
from imagine.sample import Sample
18
19
20
21

comm = MPI.COMM_WORLD
size = comm.size
rank = comm.rank
22

23
24
25
WORK_TAG = 0
DIE_TAG = 1

26
27

class Pipeline(Loggable, object):
28
29
30
31
32
33
34
35
36
37
38
39
    """
    The pipeline
    - posses all the building blocks: magnetic_field, observer,
        likelihood and prior.
    - if multiple log-likelihoods and log-priors are given: sum the result
    - coordinates the repeated observation in order to compute an ensemble
    - controls which parameters of the magnetic field are tested
        (active parameters)


    """
    def __init__(self, magnetic_field_factory, observer, likelihood, prior,
40
                 active_variables=[], ensemble_size=1,
41
                 pymultinest_parameters={}, sample_callback=None):
42
        self.logger.debug("Setting up pipeline.")
43
        self.magnetic_field_factory = magnetic_field_factory
44
45
46
        self.observer = observer
        self.likelihood = likelihood
        self.prior = prior
47
        self.active_variables = active_variables
48
49
        self.ensemble_size = ensemble_size

50
        # setting defaults for pymultinest
51
52
53
54
55
        self.pymultinest_parameters = {'verbose': True,
                                       'n_iter_before_update': 1,
                                       'n_live_points': 100}
        self.pymultinest_parameters.update(pymultinest_parameters)

56
57
        self.sample_callback = sample_callback

58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
    @property
    def observer(self):
        return self._observer

    @observer.setter
    def observer(self, observer):
        if not isinstance(observer, Observer):
            raise TypeError("observer must be an instance of Observer-class.")
        self.logger.debug("Setting observer.")
        self._observer = observer

    @property
    def likelihood(self):
        return self._likelihood

    @likelihood.setter
    def likelihood(self, likelihood):
75
        self.logger.debug("Setting likelihood.")
76
        self._likelihood = ()
Theo Steininger's avatar
Theo Steininger committed
77
        if not (isinstance(likelihood, list) or
78
79
                isinstance(likelihood, tuple)):
            likelihood = [likelihood]
80
81
82
83
84
        for l in likelihood:
            if not isinstance(l, Likelihood):
                raise TypeError(
                    "likelihood must be an instance of Likelihood-class.")
            self._likelihood += (l,)
85
86
87
88
89
90
91

    @property
    def prior(self):
        return self._prior

    @prior.setter
    def prior(self, prior):
92
        self.logger.debug("Setting prior.")
93
94
95
96
        if not isinstance(prior, Prior):
            raise TypeError(
                "prior must be an instance of Prior-class.")
        self._prior = prior
97
98

    @property
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
    def magnetic_field_factory(self):
        return self._magnetic_field_factory

    @magnetic_field_factory.setter
    def magnetic_field_factory(self, magnetic_field_factory):
        if not isinstance(magnetic_field_factory, MagneticFieldFactory):
            raise TypeError(
                "magnetic_field_factory must be an instance of the "
                "MagneticFieldFactory-class.")
        self.logger.debug("Setting magnetic_field_factory.")
        self._magnetic_field_factory = magnetic_field_factory

    @property
    def active_variables(self):
        return self._active_variables

    @active_variables.setter
    def active_variables(self, active_variables):
        if not isinstance(active_variables, list):
            raise TypeError(
                    "active_variables must be a list.")
        self.logger.debug("Resetting active_variables to %s" %
                          str(active_variables))
        new_active = []
        for av in active_variables:
            new_active += [str(av)]
        self._active_variables = new_active
126
127
128
129
130
131
132
133
134
135
136
137
138
139

    @property
    def ensemble_size(self):
        return self._ensemble_size

    @ensemble_size.setter
    def ensemble_size(self, ensemble_size):

        ensemble_size = int(ensemble_size)
        if ensemble_size <= 0:
            raise ValueError("ensemble_size must be positive!")
        self.logger.debug("Setting ensemble size to %i." % ensemble_size)
        self._ensemble_size = ensemble_size

140
141
142
143
144
145
146
147
    def _multinest_likelihood(self, cube, ndim, nparams):
        cube_content = np.empty(ndim)
        for i in xrange(ndim):
            cube_content[i] = cube[i]
        if rank != 0:
            raise RuntimeError("_multinest_likelihood must only be called on "
                               "rank==0.")
        for i in xrange(1, size):
148
            comm.send(cube_content, dest=i, tag=WORK_TAG)
Theo Steininger's avatar
Theo Steininger committed
149
        self.logger.debug("Sent multinest-cube to nodes with rank > 0.")
150
151
152
153

        return self._core_likelihood(cube_content)

    def _listen_for_likelihood_calls(self):
154
155
156
157
158
159
160
161
        status = MPI.Status()
        while True:
            cube = comm.recv(source=0, tag=MPI.ANY_TAG, status=status)
            if status == DIE_TAG:
                self.logger.debug("Received DIE_TAG from rank 0.")
                break
            self.logger.debug("Received cube from rank 0.")
            self._core_likelihood(cube)
162
163

    def _core_likelihood(self, cube):
164
165
        self.logger.debug("Beginning Likelihood-calculation for %s." %
                          str(cube))
166
167
168
169
170
171
        # translate cube to variables
        variables = {}
        for i, av in enumerate(self.active_variables):
            variables[av] = cube[i]

        # create magnetic field
172
173
174
175
176
        self.logger.debug("Creating magnetic field.")
        b_field = self.magnetic_field_factory.generate(
                                              variables=variables,
                                              ensemble_size=self.ensemble_size)

177
        # create observables
178
        self.logger.debug("Creating observables.")
179
180
181
        observables = self.observer(b_field)

        # add up individual log-likelihood terms
182
        self.logger.debug("Evaluating likelihood(s).")
183
184
        likelihood = ()
        total_likelihood = 0
185
        for like in self.likelihood:
186
187
188
            current_likelihood = like(observables)
            likelihood += (current_likelihood, )
            total_likelihood += current_likelihood
189

Theo Steininger's avatar
Theo Steininger committed
190
        self.logger.info("Evaluated likelihood: %f for %s" %
191
192
193
194
195
196
197
198
199
200
201
                         (total_likelihood, str(cube)))

        if self.sample_callback is not None:
            self.logger.debug("Creating sample-object.")
            sample = Sample(variables=variables,
                            magnetic_field=b_field,
                            observables=observables,
                            likelihood=likelihood,
                            total_likelihood=total_likelihood)
            self.sample_callback(sample)

Theo Steininger's avatar
Theo Steininger committed
202
        return total_likelihood
203

204
    def __call__(self):
205
206
207

        if rank == 0:
            # kickstart pymultinest
Theo Steininger's avatar
Theo Steininger committed
208
            self.logger.info("Starting pymultinest.")
209
210
211
212
213
            if not os.path.exists("chains"):
                os.mkdir("chains")
            pymultinest.run(self._multinest_likelihood,
                            self.prior,
                            len(self.active_variables),
214
                            **self.pymultinest_parameters)
215
216
217
218
            self.logger.info("pymultinest finished.")
            for i in xrange(1, size):
                self.logger.debug("Sending DIE_TAG to rank %i." % i)
                comm.send(None, dest=i, tag=DIE_TAG)
219
220
221
        else:
            # let all other nodes listen for likelihood evaluations
            self._listen_for_likelihood_calls()
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246

    def find_minimum(self, starting_guess=None, **kwargs):
        if starting_guess is None:
            starting_guess = np.zeros(len(self.active_variables)) + 0.5

        if rank == 0:
            # kickstart pymultinest
            self.logger.info("Starting minimizer.")
            call_func = lambda z: self._multinest_likelihood(
                                                 z,
                                                 len(self.active_variables),
                                                 len(self.active_variables))
            minimum = optimize.fmin(func=call_func,
                                    x0=starting_guess,
                                    **kwargs)
            self.logger.info("Minimizer finished.")
            for i in xrange(1, size):
                self.logger.debug("Sending DIE_TAG to rank %i." % i)
                comm.send(None, dest=i, tag=DIE_TAG)
        else:
            minimum = None
            # let all other nodes listen for likelihood evaluations
            self._listen_for_likelihood_calls()
        minimum = comm.bcast(minimum, root=0)
        return minimum