metric_gaussian_kl.py 10.4 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.
#
14
# Copyright(C) 2013-2020 Max-Planck-Society
15
16
17
18
#
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.

from .. import utilities
Philipp Arras's avatar
Philipp Arras committed
19
20
from ..linearization import Linearization
from ..operators.energy_operators import StandardHamiltonian
21
from ..operators.endomorphic_operator import EndomorphicOperator
Philipp Arras's avatar
Philipp Arras committed
22
from .energy import Energy
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
import numpy as np
from ..probing import approximation2endo
from ..sugar import makeOp, full
from ..field import Field
from ..multi_field import MultiField
from .. import random


def _shareRange(nwork, nshares, myshare):
    nbase = nwork//nshares
    additional = nwork % nshares
    lo = myshare*nbase + min(myshare, additional)
    hi = lo + nbase + int(myshare < additional)
    return lo, hi


def np_allreduce_sum(comm, arr):
    if comm is None:
        return arr
    from mpi4py import MPI
    arr = np.array(arr)
    res = np.empty_like(arr)
    comm.Allreduce(arr, res, MPI.SUM)
    return res


def allreduce_sum_field(comm, fld):
    if comm is None:
        return fld
    if isinstance(fld, Field):
        return Field(fld.domain, np_allreduce_sum(fld.val))
    res = tuple(
        Field(f.domain, np_allreduce_sum(comm, f.val))
        for f in fld.values())
    return MultiField(fld.domain, res)


class KLMetric(EndomorphicOperator):
    def __init__(self, KL):
        self._KL = KL
        self._capability = self.TIMES | self.ADJOINT_TIMES
        self._domain = KL.position.domain

    def apply(self, x, mode):
        self._check_input(x, mode)
        return self._KL.apply_metric(x)

    def draw_sample(self, from_inverse=False, dtype=np.float64):
        return self._KL.metric_sample(from_inverse, dtype)
72
73
74


class MetricGaussianKL(Energy):
Martin Reinecke's avatar
Martin Reinecke committed
75
76
77
    """Provides the sampled Kullback-Leibler divergence between a distribution
    and a Metric Gaussian.

Philipp Arras's avatar
Docs    
Philipp Arras committed
78
79
80
81
    A Metric Gaussian is used to approximate another probability distribution.
    It is a Gaussian distribution that uses the Fisher information metric of
    the other distribution at the location of its mean to approximate the
    variance. In order to infer the mean, a stochastic estimate of the
Martin Reinecke's avatar
Martin Reinecke committed
82
    Kullback-Leibler divergence is minimized. This estimate is obtained by
Philipp Arras's avatar
Docs    
Philipp Arras committed
83
84
85
86
87
    sampling the Metric Gaussian at the current mean. During minimization
    these samples are kept constant; only the mean is updated. Due to the
    typically nonlinear structure of the true distribution these samples have
    to be updated eventually by intantiating `MetricGaussianKL` again. For the
    true probability distribution the standard parametrization is assumed.
88
    The samples of this class can be distributed among MPI tasks.
89
90
91
92

    Parameters
    ----------
    mean : Field
Philipp Arras's avatar
Docs    
Philipp Arras committed
93
        Mean of the Gaussian probability distribution.
Jakob Knollmueller's avatar
Jakob Knollmueller committed
94
    hamiltonian : StandardHamiltonian
Philipp Arras's avatar
Docs    
Philipp Arras committed
95
        Hamiltonian of the approximated probability distribution.
96
    n_samples : integer
Philipp Arras's avatar
Docs    
Philipp Arras committed
97
        Number of samples used to stochastically estimate the KL.
98
    constants : list
Philipp Arras's avatar
Docs    
Philipp Arras committed
99
100
        List of parameter keys that are kept constant during optimization.
        Default is no constants.
101
    point_estimates : list
Philipp Arras's avatar
Docs    
Philipp Arras committed
102
103
104
        List of parameter keys for which no samples are drawn, but that are
        (possibly) optimized for, corresponding to point estimates of these.
        Default is to draw samples for the complete domain.
105
106
    mirror_samples : boolean
        Whether the negative of the drawn samples are also used,
Philipp Arras's avatar
Docs    
Philipp Arras committed
107
        as they are equally legitimate samples. If true, the number of used
108
        samples doubles. Mirroring samples stabilizes the KL estimate as
Philipp Arras's avatar
Docs    
Philipp Arras committed
109
        extreme sample variation is counterbalanced. Default is False.
110
111
112
    napprox : int
        Number of samples for computing preconditioner for sampling. No
        preconditioning is done by default.
113
114
115
116
117
    use_mpi : bool
        whether MPI should be used.
        If MPI is enabled, samples will be distributed as evenly as possible
        across MPI.COMM_WORLD. If `mirror_samples` is set, then a sample and
        its mirror image will always reside on the same task.
Philipp Arras's avatar
Philipp Arras committed
118
119
    _samples : None
        Only a parameter for internal uses. Typically not to be set by users.
120

121
122
FIXME: lh_sampling_dtype not documented!

Philipp Arras's avatar
Docs    
Philipp Arras committed
123
124
125
126
127
128
129
130
    Note
    ----
    The two lists `constants` and `point_estimates` are independent from each
    other. It is possible to sample along domains which are kept constant
    during minimization and vice versa.

    See also
    --------
131
132
    `Metric Gaussian Variational Inference`, Jakob Knollmüller,
    Torsten A. Enßlin, `<https://arxiv.org/abs/1901.11033>`_
133
134
    """

Martin Reinecke's avatar
typo    
Martin Reinecke committed
135
    def __init__(self, mean, hamiltonian, n_samples, constants=[],
Philipp Arras's avatar
Philipp Arras committed
136
                 point_estimates=[], mirror_samples=False,
137
138
                 napprox=0, use_mpi=False, _samples=None,
                 lh_sampling_dtype=np.float64):
139
        super(MetricGaussianKL, self).__init__(mean)
Philipp Arras's avatar
Philipp Arras committed
140
141
142

        if not isinstance(hamiltonian, StandardHamiltonian):
            raise TypeError
143
        if hamiltonian.domain is not mean.domain:
Philipp Arras's avatar
Philipp Arras committed
144
145
146
            raise ValueError
        if not isinstance(n_samples, int):
            raise TypeError
147
148
        self._constants = tuple(constants)
        self._point_estimates = tuple(point_estimates)
Philipp Arras's avatar
Philipp Arras committed
149
        if not isinstance(mirror_samples, bool):
150
            raise TypeError
Philipp Arras's avatar
Philipp Arras committed
151

152
        self._hamiltonian = hamiltonian
Philipp Arras's avatar
Philipp Arras committed
153

154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
        self._use_mpi = bool(use_mpi)
        if self._use_mpi:
            from mpi4py import MPI
            self._comm = MPI.COMM_WORLD
            self._ntask = self._comm.Get_size()
            self._rank = self._comm.Get_rank()
            self._master = (self._rank == 0)
        else:
            self._comm, self._ntask, self._rank, self._master = None, 1, 0, True

        self._n_samples = int(n_samples)
        self._lo, self._hi = _shareRange(self._n_samples, self._ntask, self._rank)
        self._mirror_samples = bool(mirror_samples)
        self._n_eff_samples = self._n_samples
        if self._mirror_samples:
            self._n_eff_samples *= 2

171
172
        if _samples is None:
            met = hamiltonian(Linearization.make_partial_var(
173
174
                mean, self._point_estimates, True)).metric
# FIXME: should this be ">=1" instead of ">1"?
175
176
            if napprox > 1:
                met._approximation = makeOp(approximation2endo(met, napprox))
177
178
179
180
181
182
183
184
185
186
187
            _samples = []
            sseq = random.spawn_sseq(self._n_samples)
            for i in range(self._lo, self._hi):
                random.push_sseq(sseq[i])
                _samples.append(met.draw_sample(from_inverse=True,
                                                dtype=lh_sampling_dtype))
                random.pop_sseq()
            _samples = tuple(_samples)
        else:
            if len(_samples) != self._n_samples:
                raise ValueError("# of samples mismatch")
188
        self._samples = _samples
189
        self._lin = Linearization.make_partial_var(mean, self._constants)
190
        v, g = None, None
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
        if len(self._samples) == 0:  # hack if there are too many MPI tasks
            tmp = self._hamiltonian(self._lin)
            v = 0. * tmp.val.val
            g = 0. * tmp.gradient
        else:
            for s in self._samples:
                tmp = self._hamiltonian(self._lin+s)
                if self._mirror_samples:
                    tmp = tmp + self._hamiltonian(self._lin-s)
                if v is None:
                    v = tmp.val.val_rw()
                    g = tmp.gradient
                else:
                    v += tmp.val.val
                    g = g + tmp.gradient
        self._val = np_allreduce_sum(self._comm, v)[()] / self._n_eff_samples
        self._grad = allreduce_sum_field(self._comm, g) / self._n_eff_samples
208
        self._metric = None
209
        self._sampdt = lh_sampling_dtype
210
211

    def at(self, position):
212
213
214
215
        return MetricGaussianKL(
            position, self._hamiltonian, self._n_samples, self._constants,
            self._point_estimates, self._mirror_samples, use_mpi=self._use_mpi,
            _samples=self._samples, lh_sampling_dtype=self._sampdt)
216
217
218
219
220
221
222
223
224
225

    @property
    def value(self):
        return self._val

    @property
    def gradient(self):
        return self._grad

    def _get_metric(self):
226
        lin = self._lin.with_want_metric()
227
        if self._metric is None:
228
229
230
231
232
233
234
            if len(self._samples) == 0:  # hack if there are too many MPI tasks
                self._metric = self._hamiltonian(lin).metric.scale(0.)
            else:
                mymap = map(lambda v: self._hamiltonian(lin+v).metric,
                            self._samples)
                self.unscaled_metric = utilities.my_sum(mymap)
                self._metric = self.unscaled_metric.scale(1./self._n_eff_samples)
235
236
237

    def apply_metric(self, x):
        self._get_metric()
238
        return allreduce_sum_field(self._comm, self._metric(x))
239
240
241

    @property
    def metric(self):
242
        return KLMetric(self)
243
244
245

    @property
    def samples(self):
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
        if self._comm is not None:
            res = _comm.allgather(self._samples)
            res = [item for sublist in res for item in sublist]
        else:
            res = self._samples
        if self._mirror_samples:
            res = res + tuple(-item for item in res)
        return res

    def unscaled_metric_sample(self, from_inverse=False, dtype=np.float64):
        if from_inverse:
            raise NotImplementedError()
        lin = self._lin.with_want_metric()
        samp = full(self._hamiltonian.domain, 0.)
        sseq = random.spawn_sseq(n_samples)
        for i, v in enumerate(self._samples):
            random.push_sseq(sseq[self._lo+i])
            samp = samp + self._hamiltonian(lin+v).metric.draw_sample(from_inverse=False, dtype=dtype)
            if self._mirror_samples:
                samp = samp + self._hamiltonian(lin-v).metric.draw_sample(from_inverse=False, dtype=dtype)
            random.pop_sseq()
        return allreduce_sum_field(self._comm, samp)
268

269
270
    def metric_sample(self, from_inverse=False, dtype=np.float64):
        return self.unscaled_metric_sample(from_inverse, dtype)/self._n_eff_samples