energy_operators.py 12.4 KB
Newer Older
Martin Reinecke's avatar
Martin Reinecke committed
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) 2013-2019 Max-Planck-Society
Martin Reinecke's avatar
Martin Reinecke committed
15
#
16
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik.
Martin Reinecke's avatar
Martin Reinecke committed
17

Philipp Arras's avatar
Philipp Arras committed
18
19
import numpy as np

Philipp Arras's avatar
Philipp Arras committed
20
from .. import utilities
Martin Reinecke's avatar
Martin Reinecke committed
21
from ..domain_tuple import DomainTuple
Philipp Arras's avatar
Philipp Arras committed
22
23
from ..field import Field
from ..linearization import Linearization
Philipp Arras's avatar
Philipp Arras committed
24
25
from ..sugar import makeDomain, makeOp
from .linear_operator import LinearOperator
Martin Reinecke's avatar
Martin Reinecke committed
26
from .operator import Operator
Martin Reinecke's avatar
fix    
Martin Reinecke committed
27
from .sampling_enabler import SamplingEnabler
Philipp Arras's avatar
Philipp Arras committed
28
from .sandwich_operator import SandwichOperator
Martin Reinecke's avatar
Martin Reinecke committed
29
from .simple_linear_operators import VdotOperator
Martin Reinecke's avatar
Martin Reinecke committed
30
31
32


class EnergyOperator(Operator):
Philipp Arras's avatar
Philipp Arras committed
33
    """Operator which has a scalar domain as target domain.
34

Martin Reinecke's avatar
Martin Reinecke committed
35
    It is intended as an objective function for field inference.
36

Philipp Arras's avatar
Philipp Arras committed
37
38
39
40
41
    Examples
    --------
     - Information Hamiltonian, i.e. negative-log-probabilities.
     - Gibbs free energy, i.e. an averaged Hamiltonian, aka Kullbach-Leibler
       divergence.
42
    """
Martin Reinecke's avatar
Martin Reinecke committed
43
44
45
46
    _target = DomainTuple.scalar_domain()


class SquaredNormOperator(EnergyOperator):
Philipp Arras's avatar
Philipp Arras committed
47
    """Computes the L2-norm of the output of an operator.
48

Philipp Arras's avatar
Philipp Arras committed
49
50
51
52
    Parameters
    ----------
    domain : Domain, DomainTuple or tuple of Domain
        Target domain of the operator in which the L2-norm shall be computed.
Martin Reinecke's avatar
Martin Reinecke committed
53
    """
Philipp Arras's avatar
Philipp Arras committed
54

Martin Reinecke's avatar
Martin Reinecke committed
55
56
57
58
    def __init__(self, domain):
        self._domain = domain

    def apply(self, x):
59
        self._check_input(x)
Martin Reinecke's avatar
Martin Reinecke committed
60
        if isinstance(x, Linearization):
Martin Reinecke's avatar
Martin Reinecke committed
61
            val = Field.scalar(x.val.vdot(x.val))
Martin Reinecke's avatar
Martin Reinecke committed
62
            jac = VdotOperator(2*x.val)(x.jac)
63
            return x.new(val, jac)
Martin Reinecke's avatar
Martin Reinecke committed
64
        return Field.scalar(x.vdot(x))
Martin Reinecke's avatar
Martin Reinecke committed
65
66
67


class QuadraticFormOperator(EnergyOperator):
Philipp Arras's avatar
Philipp Arras committed
68
69
70
71
72
    """Computes the L2-norm of a Field or MultiField with respect to a
    specific metric `endo`.

    .. math ::
        E(f) = \\frac12 f^\\dagger \\text{endo}(f)
73
74
75

    Parameters
    ----------
Philipp Arras's avatar
Philipp Arras committed
76
77
    endo : EndomorphicOperator
         Kernel of quadratic form.
Martin Reinecke's avatar
Martin Reinecke committed
78
    """
Philipp Arras's avatar
Philipp Arras committed
79
80

    def __init__(self, endo):
Martin Reinecke's avatar
Martin Reinecke committed
81
        from .endomorphic_operator import EndomorphicOperator
Philipp Arras's avatar
Philipp Arras committed
82
        if not isinstance(endo, EndomorphicOperator):
Martin Reinecke's avatar
Martin Reinecke committed
83
            raise TypeError("op must be an EndomorphicOperator")
Philipp Arras's avatar
Philipp Arras committed
84
85
        self._op = endo
        self._domain = endo.domain
Martin Reinecke's avatar
Martin Reinecke committed
86
87

    def apply(self, x):
88
        self._check_input(x)
Martin Reinecke's avatar
Martin Reinecke committed
89
        if isinstance(x, Linearization):
Martin Reinecke's avatar
Martin Reinecke committed
90
91
            t1 = self._op(x.val)
            jac = VdotOperator(t1)(x.jac)
Martin Reinecke's avatar
Martin Reinecke committed
92
            val = Field.scalar(0.5*x.val.vdot(t1))
93
            return x.new(val, jac)
Martin Reinecke's avatar
Martin Reinecke committed
94
        return Field.scalar(0.5*x.vdot(self._op(x)))
Martin Reinecke's avatar
Martin Reinecke committed
95
96
97


class GaussianEnergy(EnergyOperator):
Martin Reinecke's avatar
Martin Reinecke committed
98
    """Class for energies of fields with Gaussian probability distribution.
99

Philipp Arras's avatar
Philipp Arras committed
100
    Represents up to constants in :math:`m`:
Martin Reinecke's avatar
Martin Reinecke committed
101

Philipp Arras's avatar
Philipp Arras committed
102
103
    .. math ::
        E(f) = - \\log G(f-m, D) = 0.5 (f-m)^\\dagger D^{-1} (f-m),
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
104

Philipp Arras's avatar
Philipp Arras committed
105
106
    an information energy for a Gaussian distribution with mean m and
    covariance D.
107

Philipp Arras's avatar
Philipp Arras committed
108
109
110
111
112
113
114
115
116
117
118
119
120
    Parameters
    ----------
    mean : Field
        Mean of the Gaussian. Default is 0.
    covariance : LinearOperator
        Covariance of the Gaussian. Default is the identity operator.
    domain : Domain, DomainTuple of tuple of Domain
        Operator domain. By default it is inferred from `mean` or
        `covariance` if specified

    Note
    ----
    At least one of the arguments has to be provided.
Martin Reinecke's avatar
Martin Reinecke committed
121
    """
Martin Reinecke's avatar
Martin Reinecke committed
122

Martin Reinecke's avatar
Martin Reinecke committed
123
    def __init__(self, mean=None, covariance=None, domain=None):
Philipp Arras's avatar
Philipp Arras committed
124
125
126
127
128
129
130
131
        if mean is not None and not isinstance(mean, Field):
            raise TypeError
        if covariance is not None and not isinstance(covariance,
                                                     LinearOperator):
            raise TypeError
        if domain is not None:
            domain = DomainTuple.make(domain)

Martin Reinecke's avatar
Martin Reinecke committed
132
133
134
135
136
137
138
139
140
141
        self._domain = None
        if mean is not None:
            self._checkEquivalence(mean.domain)
        if covariance is not None:
            self._checkEquivalence(covariance.domain)
        if domain is not None:
            self._checkEquivalence(domain)
        if self._domain is None:
            raise ValueError("no domain given")
        self._mean = mean
Martin Reinecke's avatar
Martin Reinecke committed
142
143
144
145
        if covariance is None:
            self._op = SquaredNormOperator(self._domain).scale(0.5)
        else:
            self._op = QuadraticFormOperator(covariance.inverse)
Martin Reinecke's avatar
Martin Reinecke committed
146
147
148
        self._icov = None if covariance is None else covariance.inverse

    def _checkEquivalence(self, newdom):
Martin Reinecke's avatar
fix    
Martin Reinecke committed
149
        newdom = makeDomain(newdom)
Martin Reinecke's avatar
Martin Reinecke committed
150
        if self._domain is None:
Philipp Arras's avatar
Philipp Arras committed
151
            self._domain = newdom
Martin Reinecke's avatar
Martin Reinecke committed
152
        else:
Philipp Arras's avatar
Philipp Arras committed
153
            if self._domain != newdom:
Martin Reinecke's avatar
Martin Reinecke committed
154
155
156
                raise ValueError("domain mismatch")

    def apply(self, x):
157
        self._check_input(x)
Philipp Arras's avatar
Philipp Arras committed
158
        residual = x if self._mean is None else x - self._mean
Philipp Arras's avatar
Changes    
Philipp Arras committed
159
        res = self._op(residual).real
160
        if not isinstance(x, Linearization) or not x.want_metric:
Martin Reinecke's avatar
Martin Reinecke committed
161
162
163
164
165
166
            return res
        metric = SandwichOperator.make(x.jac, self._icov)
        return res.add_metric(metric)


class PoissonianEnergy(EnergyOperator):
Philipp Arras's avatar
Philipp Arras committed
167
168
    """Class for likelihood Hamiltonians of expected count field constrained
    by Poissonian count data.
169

Philipp Arras's avatar
Philipp Arras committed
170
    Represents up to an f-independent term :math:`log(d!)`:
171

Philipp Arras's avatar
Philipp Arras committed
172
173
    .. math ::
        E(f) = -\\log \\text{Poisson}(d|f) = \\sum f - d^\\dagger \\log(f),
174

Philipp Arras's avatar
Philipp Arras committed
175
    where f is a :class:`Field` in data space with the expectation values for
Martin Reinecke's avatar
Martin Reinecke committed
176
    the counts.
Philipp Arras's avatar
Philipp Arras committed
177
178
179
180
181
182

    Parameters
    ----------
    d : Field
        Data field with counts. Needs to have integer dtype and all field
        values need to be non-negative.
Martin Reinecke's avatar
Martin Reinecke committed
183
    """
Philipp Arras's avatar
Philipp Arras committed
184

185
    def __init__(self, d):
Philipp Arras's avatar
Philipp Arras committed
186
187
188
189
        if not isinstance(d, Field) or not np.issubdtype(d.dtype, np.integer):
            raise TypeError
        if np.any(d.local_data < 0):
            raise ValueError
190
191
        self._d = d
        self._domain = DomainTuple.make(d.domain)
Martin Reinecke's avatar
Martin Reinecke committed
192
193

    def apply(self, x):
194
        self._check_input(x)
Martin Reinecke's avatar
Martin Reinecke committed
195
196
        res = x.sum() - x.log().vdot(self._d)
        if not isinstance(x, Linearization):
Martin Reinecke's avatar
Martin Reinecke committed
197
            return Field.scalar(res)
198
199
        if not x.want_metric:
            return res
Martin Reinecke's avatar
Martin Reinecke committed
200
201
202
        metric = SandwichOperator.make(x.jac, makeOp(1./x.val))
        return res.add_metric(metric)

203

204
class InverseGammaLikelihood(EnergyOperator):
Philipp Arras's avatar
Philipp Arras committed
205
206
    """
    FIXME
207
    """
Philipp Arras's avatar
Philipp Arras committed
208

Martin Reinecke's avatar
Martin Reinecke committed
209
    def __init__(self, d):
Philipp Arras's avatar
Philipp Arras committed
210
211
        if not isinstance(d, Field):
            raise TypeError
212
213
        self._d = d
        self._domain = DomainTuple.make(d.domain)
214
215

    def apply(self, x):
216
        self._check_input(x)
Philipp Frank's avatar
Philipp Frank committed
217
        res = 0.5*(x.log().sum() + (1./x).vdot(self._d))
218
219
        if not isinstance(x, Linearization):
            return Field.scalar(res)
220
221
        if not x.want_metric:
            return res
222
223
224
225
        metric = SandwichOperator.make(x.jac, makeOp(0.5/(x.val**2)))
        return res.add_metric(metric)


Martin Reinecke's avatar
Martin Reinecke committed
226
class BernoulliEnergy(EnergyOperator):
Philipp Arras's avatar
Philipp Arras committed
227
    """Computes likelihood energy of expected event frequency constrained by
228
229
    event data.

Philipp Arras's avatar
Philipp Arras committed
230
231
232
233
234
235
236
    .. math ::
        E(f) = -\\log \\text{Bernoulli}(d|f)
             = -d^\\dagger \\log f  - (1-d)^\\dagger \\log(1-f),

    where f is a field defined on `d.domain` with the expected
    frequencies of events.

237
238
    Parameters
    ----------
Martin Reinecke's avatar
Martin Reinecke committed
239
    d : Field
Philipp Arras's avatar
Philipp Arras committed
240
        Data field with events (1) or non-events (0).
Martin Reinecke's avatar
Martin Reinecke committed
241
    """
Philipp Arras's avatar
Philipp Arras committed
242

243
    def __init__(self, d):
Philipp Arras's avatar
Philipp Arras committed
244
245
246
247
248
        print(d.dtype)
        if not isinstance(d, Field) or not np.issubdtype(d.dtype, np.integer):
            raise TypeError
        if not np.all(np.logical_or(d.local_data == 0, d.local_data == 1)):
            raise ValueError
Martin Reinecke's avatar
Martin Reinecke committed
249
        self._d = d
250
        self._domain = DomainTuple.make(d.domain)
Martin Reinecke's avatar
Martin Reinecke committed
251
252

    def apply(self, x):
253
        self._check_input(x)
Philipp Arras's avatar
Philipp Arras committed
254
        v = -(x.log().vdot(self._d) + (1. - x).log().vdot(1. - self._d))
Martin Reinecke's avatar
Martin Reinecke committed
255
        if not isinstance(x, Linearization):
Martin Reinecke's avatar
Martin Reinecke committed
256
            return Field.scalar(v)
257
258
        if not x.want_metric:
            return v
Philipp Arras's avatar
Philipp Arras committed
259
        met = makeOp(1./(x.val*(1. - x.val)))
Martin Reinecke's avatar
Martin Reinecke committed
260
261
262
263
264
        met = SandwichOperator.make(x.jac, met)
        return v.add_metric(met)


class Hamiltonian(EnergyOperator):
Philipp Arras's avatar
Philipp Arras committed
265
266
    """Computes an information Hamiltonian in its standard form, i.e. with the
    prior being a Gaussian with unit covariance.
267

Philipp Arras's avatar
Philipp Arras committed
268
    Let the likelihood energy be :math:`E_{lh}`. Then this operator computes:
269

Philipp Arras's avatar
Philipp Arras committed
270
271
    .. math ::
         H(f) = 0.5 f^\\dagger f + E_{lh}(f):
272

Martin Reinecke's avatar
Martin Reinecke committed
273
    Other field priors can be represented via transformations of a white
274
275
    Gaussian field into a field with the desired prior probability structure.

Martin Reinecke's avatar
Martin Reinecke committed
276
    By implementing prior information this way, the field prior is represented
277
278
279
    by a generative model, from which NIFTy can draw samples and infer a field
    using the Maximum a Posteriori (MAP) or the Variational Bayes (VB) method.

Philipp Arras's avatar
Philipp Arras committed
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
    The metric of this operator can be used as covariance for drawing Gaussian
    samples.

    Parameters
    ----------
    lh : EnergyOperator
        The likelihood energy.
    ic_samp : IterationController
        Tells an internal :class:`SamplingEnabler` which convergence criterium
        to use to draw Gaussian samples.


    See also
    --------
    `Encoding prior knowledge in the structure of the likelihood`,
    Jakob Knollmüller, Torsten A. Ensslin,
Martin Reinecke's avatar
Martin Reinecke committed
296
    `<https://arxiv.org/abs/1812.04403>`_
Martin Reinecke's avatar
Martin Reinecke committed
297
    """
Philipp Arras's avatar
Philipp Arras committed
298

Martin Reinecke's avatar
Martin Reinecke committed
299
300
301
302
    def __init__(self, lh, ic_samp=None):
        self._lh = lh
        self._prior = GaussianEnergy(domain=lh.domain)
        self._ic_samp = ic_samp
Martin Reinecke's avatar
Martin Reinecke committed
303
        self._domain = lh.domain
Martin Reinecke's avatar
Martin Reinecke committed
304
305

    def apply(self, x):
306
        self._check_input(x)
Philipp Arras's avatar
Philipp Arras committed
307
308
309
        if (self._ic_samp is None or not isinstance(x, Linearization)
                or not x.want_metric):
            return self._lh(x) + self._prior(x)
Martin Reinecke's avatar
Martin Reinecke committed
310
        else:
311
            lhx, prx = self._lh(x), self._prior(x)
Martin Reinecke's avatar
Martin Reinecke committed
312
313
            mtr = SamplingEnabler(lhx.metric, prx.metric.inverse,
                                  self._ic_samp, prx.metric.inverse)
Philipp Arras's avatar
Philipp Arras committed
314
            return (lhx + prx).add_metric(mtr)
Martin Reinecke's avatar
Martin Reinecke committed
315

Philipp Arras's avatar
Philipp Arras committed
316
317
318
319
320
    def __repr__(self):
        subs = 'Likelihood:\n{}'.format(utilities.indent(self._lh.__repr__()))
        subs += '\nPrior: Quadratic{}'.format(self._lh.domain.keys())
        return 'Hamiltonian:\n' + utilities.indent(subs)

Martin Reinecke's avatar
Martin Reinecke committed
321

Martin Reinecke's avatar
Martin Reinecke committed
322
class AveragedEnergy(EnergyOperator):
Philipp Arras's avatar
Philipp Arras committed
323
    """Computes Kullbach-Leibler (KL) divergence or Gibbs free energies.
324

Philipp Arras's avatar
Philipp Arras committed
325
326
327
    A sample-averaged energy, e.g. an Hamiltonian, approximates the relevant
    part of a KL to be used in Variational Bayes inference if the samples are
    drawn from the approximating Gaussian:
Martin Reinecke's avatar
Martin Reinecke committed
328

Philipp Arras's avatar
Philipp Arras committed
329
330
    .. math ::
        \\text{KL}(m) = \\frac1{\\#\{v_i\}} \\sum_{v_i} H(m+v_i),
331

Philipp Arras's avatar
Philipp Arras committed
332
333
    where :math:`v_i` are the residual samples and :math:`m` is the mean field
    around which the samples are drawn.
Martin Reinecke's avatar
Martin Reinecke committed
334

335
336
337
    Parameters
    ----------
    h: Hamiltonian
Philipp Arras's avatar
Philipp Arras committed
338
       The energy to be averaged.
Martin Reinecke's avatar
Martin Reinecke committed
339
    res_samples : iterable of Fields
Philipp Arras's avatar
Philipp Arras committed
340
341
       Set of residual sample points to be added to mean field for approximate
       estimation of the KL.
342

Philipp Arras's avatar
Philipp Arras committed
343
344
    Note
    ----
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
345
346
347
348
    Having symmetrized residual samples, with both v_i and -v_i being present
    ensures that the distribution mean is exactly represented. This reduces
    sampling noise and helps the numerics of the KL minimization process in the
    variational Bayes inference.
Philipp Arras's avatar
Philipp Arras committed
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369

    See also
    --------
    Let :math:`Q(f) = G(f-m,D)` be the Gaussian distribution
    which is used to approximate the accurate posterior :math:`P(f|d)` with
    information Hamiltonian
    :math:`H(d,f) = -\\log P(d,f) = -\\log P(f|d) + \\text{const}`. In
    Variational Bayes one needs to optimize the KL divergence between those
    two distributions for m. It is:

    :math:`KL(Q,P) = \\int Df Q(f) \\log Q(f)/P(f)\\\\
    = \\left< \\log Q(f) \\right>_Q(f) - \\left< \\log P(f) \\right>_Q(f)\\\\
    = \\text{const} + \\left< H(f) \\right>_G(f-m,D)`

    in essence the information Hamiltonian averaged over a Gaussian
    distribution centered on the mean m.

    :class:`AveragedEnergy(h)` approximates
    :math:`\\left< H(f) \\right>_{G(f-m,D)}` if the residuals
    :math:`f-m` are drawn from a Gaussian distribution with covariance
    :math:`D`.
Martin Reinecke's avatar
Martin Reinecke committed
370
    """
Philipp Arras's avatar
Philipp Arras committed
371

Martin Reinecke's avatar
Martin Reinecke committed
372
373
    def __init__(self, h, res_samples):
        self._h = h
Martin Reinecke's avatar
Martin Reinecke committed
374
        self._domain = h.domain
Martin Reinecke's avatar
Martin Reinecke committed
375
376
377
        self._res_samples = tuple(res_samples)

    def apply(self, x):
378
        self._check_input(x)
Philipp Arras's avatar
Philipp Arras committed
379
380
        mymap = map(lambda v: self._h(x + v), self._res_samples)
        return utilities.my_sum(mymap)*(1./len(self._res_samples))