metric_gaussian_kl.py 5.69 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
# 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/>.
#
# Copyright(C) 2013-2019 Max-Planck-Society
#
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.

from .. import utilities
Philipp Arras's avatar
Philipp Arras committed
19
20
21
from ..linearization import Linearization
from ..operators.energy_operators import StandardHamiltonian
from .energy import Energy
22
23
24


class MetricGaussianKL(Energy):
Martin Reinecke's avatar
Martin Reinecke committed
25
26
27
    """Provides the sampled Kullback-Leibler divergence between a distribution
    and a Metric Gaussian.

Philipp Arras's avatar
Docs    
Philipp Arras committed
28
29
30
31
    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
32
    Kullback-Leibler divergence is minimized. This estimate is obtained by
Philipp Arras's avatar
Docs    
Philipp Arras committed
33
34
35
36
37
    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.
38
39
40
41

    Parameters
    ----------
    mean : Field
Philipp Arras's avatar
Docs    
Philipp Arras committed
42
        Mean of the Gaussian probability distribution.
Jakob Knollmueller's avatar
Jakob Knollmueller committed
43
    hamiltonian : StandardHamiltonian
Philipp Arras's avatar
Docs    
Philipp Arras committed
44
        Hamiltonian of the approximated probability distribution.
45
    n_samples : integer
Philipp Arras's avatar
Docs    
Philipp Arras committed
46
        Number of samples used to stochastically estimate the KL.
47
    constants : list
Philipp Arras's avatar
Docs    
Philipp Arras committed
48
49
        List of parameter keys that are kept constant during optimization.
        Default is no constants.
50
    point_estimates : list
Philipp Arras's avatar
Docs    
Philipp Arras committed
51
52
53
        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.
54
55
    mirror_samples : boolean
        Whether the negative of the drawn samples are also used,
Philipp Arras's avatar
Docs    
Philipp Arras committed
56
        as they are equally legitimate samples. If true, the number of used
57
        samples doubles. Mirroring samples stabilizes the KL estimate as
Philipp Arras's avatar
Docs    
Philipp Arras committed
58
        extreme sample variation is counterbalanced. Default is False.
Philipp Arras's avatar
Philipp Arras committed
59
60
    _samples : None
        Only a parameter for internal uses. Typically not to be set by users.
61

Philipp Arras's avatar
Docs    
Philipp Arras committed
62
63
64
65
66
67
68
69
70
    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
    --------
    Metric Gaussian Variational Inference (FIXME in preparation)
71
72
    """

Martin Reinecke's avatar
typo    
Martin Reinecke committed
73
    def __init__(self, mean, hamiltonian, n_samples, constants=[],
Philipp Arras's avatar
Philipp Arras committed
74
                 point_estimates=[], mirror_samples=False,
75
76
                 _samples=None):
        super(MetricGaussianKL, self).__init__(mean)
Philipp Arras's avatar
Philipp Arras committed
77
78
79

        if not isinstance(hamiltonian, StandardHamiltonian):
            raise TypeError
80
        if hamiltonian.domain is not mean.domain:
Philipp Arras's avatar
Philipp Arras committed
81
82
83
84
85
86
            raise ValueError
        if not isinstance(n_samples, int):
            raise TypeError
        self._constants = list(constants)
        self._point_estimates = list(point_estimates)
        if not isinstance(mirror_samples, bool):
87
            raise TypeError
Philipp Arras's avatar
Philipp Arras committed
88

89
        self._hamiltonian = hamiltonian
Philipp Arras's avatar
Philipp Arras committed
90

91
92
93
94
        if _samples is None:
            met = hamiltonian(Linearization.make_partial_var(
                mean, point_estimates, True)).metric
            _samples = tuple(met.draw_sample(from_inverse=True)
Martin Reinecke's avatar
typo    
Martin Reinecke committed
95
                             for _ in range(n_samples))
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
            if mirror_samples:
                _samples += tuple(-s for s in _samples)
        self._samples = _samples

        self._lin = Linearization.make_partial_var(mean, constants)
        v, g = None, None
        for s in self._samples:
            tmp = self._hamiltonian(self._lin+s)
            if v is None:
                v = tmp.val.local_data[()]
                g = tmp.gradient
            else:
                v += tmp.val.local_data[()]
                g = g + tmp.gradient
        self._val = v / len(self._samples)
        self._grad = g * (1./len(self._samples))
        self._metric = None

    def at(self, position):
        return MetricGaussianKL(position, self._hamiltonian, 0,
Philipp Arras's avatar
Philipp Arras committed
116
                                self._constants, self._point_estimates,
117
118
119
120
121
122
123
124
125
126
127
128
129
                                _samples=self._samples)

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

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

    def _get_metric(self):
        if self._metric is None:
            lin = self._lin.with_want_metric()
Martin Reinecke's avatar
Martin Reinecke committed
130
131
            mymap = map(lambda v: self._hamiltonian(lin+v).metric,
                        self._samples)
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
            self._metric = utilities.my_sum(mymap)
            self._metric = self._metric.scale(1./len(self._samples))

    def apply_metric(self, x):
        self._get_metric()
        return self._metric(x)

    @property
    def metric(self):
        self._get_metric()
        return self._metric

    @property
    def samples(self):
        return self._samples

    def __repr__(self):
        return 'KL ({} samples):\n'.format(len(
Philipp Arras's avatar
Philipp Arras committed
150
            self._samples)) + utilities.indent(self._hamiltonian.__repr__())