critical_power_energy.py 5.67 KB
Newer Older
1
2
from ...energies.energy import Energy
from ...operators.smoothness_operator import SmoothnessOperator
3
from ...operators.diagonal_operator import DiagonalOperator
4
5
from ...operators.linear_operator import LinearOperator
from ...operators.power_projection_operator import PowerProjection
6
from . import CriticalPowerCurvature
7
from ...memoization import memo
8
from ...minimization import ConjugateGradient
9

10
11
from ...sugar import generate_posterior_sample
from ... import Field, exp
12

13

14
class CriticalPowerEnergy(Energy):
Jakob Knollmueller's avatar
Jakob Knollmueller committed
15
    """The Energy of the power spectrum according to the critical filter.
16

17
18
19
20
    It describes the energy of the logarithmic amplitudes of the power spectrum
    of a Gaussian random field under reconstruction uncertainty with smoothness
    and inverse gamma prior. It is used to infer the correlation structure of a
    correlated signal.
21
22
23

    Parameters
    ----------
Jakob Knollmueller's avatar
Jakob Knollmueller committed
24
25
26
27
28
29
30
31
32
    position : Field,
        The current position of this energy.
    m : Field,
        The map whichs power spectrum has to be inferred
    D : EndomorphicOperator,
        The curvature of the Gaussian encoding the posterior covariance.
        If not specified, the map is assumed to be no reconstruction.
        default : None
    alpha : float
33
34
        The spectral prior of the inverse gamma distribution. 1.0 corresponds
        to non-informative.
Jakob Knollmueller's avatar
Jakob Knollmueller committed
35
36
        default : 1.0
    q : float
37
38
39
40
41
        The cutoff parameter of the inverse gamma distribution. 0.0 corresponds
        to non-informative.
        default : 0.0
    smoothness_prior : float
        Controls the strength of the smoothness prior
Jakob Knollmueller's avatar
Jakob Knollmueller committed
42
43
44
45
        default : 0.0
    logarithmic : boolean
        Whether smoothness acts on linear or logarithmic scale.
    samples : integer
46
47
        Number of samples used for the estimation of the uncertainty
        corrections.
Jakob Knollmueller's avatar
Jakob Knollmueller committed
48
49
50
51
52
        default : 3
    w : Field
        The contribution from the map with or without uncertainty. It is used
        to pass on the result of the costly sampling during the minimization.
        default : None
53
54
55
    inverter : ConjugateGradient
        The inversion strategy to invert the curvature and to generate samples.
        default : None
56
57
    """

58
59
    # ---Overwritten properties and methods---

60
    def __init__(self, position, m, D=None, alpha=1.0, q=0.,
61
                 smoothness_prior=0., logarithmic=True, samples=3, w=None,
62
63
64
65
                 inverter=None, gradient=None, curvature=None):
        super(CriticalPowerEnergy, self).__init__(position=position,
                                                  gradient=gradient,
                                                  curvature=curvature)
66
        self.m = m
67
68
        self.D = D
        self.samples = samples
69
70
        self.alpha = Field(self.position.domain, val=alpha)
        self.q = Field(self.position.domain, val=q)
71
        self.T = SmoothnessOperator(domain=self.position.domain[0],
72
                                    strength=smoothness_prior,
Jakob Knollmueller's avatar
Jakob Knollmueller committed
73
                                    logarithmic=logarithmic)
74
        self.rho = self.position.domain[0].rho
75
        self.P = PowerProjection(domain=self.m.domain,target=self.position.domain)
76
        self._w = w if w is not None else None
77
78
        if inverter is None:
            preconditioner = DiagonalOperator(self._theta.domain,
79
                                              diagonal=self._theta,
80
81
82
                                              copy=False)
            inverter = ConjugateGradient(preconditioner=preconditioner)
        self._inverter = inverter
83
        self.one = Field(self.position.domain,val=1.)
84
85
86
87

    @property
    def inverter(self):
        return self._inverter
88

89
90
    # ---Mandatory properties and methods---

91
    def at(self, position):
92
93
        return self.__class__(position, self.m, D=self.D, alpha=self.alpha,
                              q=self.q, smoothness_prior=self.smoothness_prior,
94
                              logarithmic=self.logarithmic,
95
                              w=self.w, samples=self.samples,
96
                              inverter=self.inverter)
97
98

    @property
99
    @memo
100
    def value(self):
101
102
        energy = self.one.vdot(self._theta)
        energy += self.position.vdot(self.one/2.)
103
        energy += 0.5 * self.position.vdot(self._Tt)
104
105
106
        return energy.real

    @property
107
    @memo
108
    def gradient(self):
109
110
        gradient = -self._theta
        gradient += (self.one/2.)
111
        gradient += self._Tt
112
        gradient.val = gradient.val.real
113
114
115
        return gradient

    @property
116
    @memo
117
    def curvature(self):
118
        return CriticalPowerCurvature(theta=self._theta, T=self.T,
119
                                      inverter=self.inverter)
120

121
122
123
124
125
126
127
128
129
130
    # ---Added properties and methods---

    @property
    def logarithmic(self):
        return self.T.logarithmic

    @property
    def smoothness_prior(self):
        return self.T.strength

131
132
133
    @property
    def w(self):
        if self._w is None:
Theo Steininger's avatar
Theo Steininger committed
134
            self.logger.info("Initializing w")
135
136
137
            w = Field(domain=self.position.domain, val=0., dtype=self.m.dtype)
            if self.D is not None:
                for i in range(self.samples):
Theo Steininger's avatar
Theo Steininger committed
138
                    self.logger.info("Drawing sample %i" % i)
139
140
                    posterior_sample = generate_posterior_sample(
                                                            self.m, self.D)
141
142
                    w += self.P(abs(posterior_sample) ** 2)

143
144
                w /= float(self.samples)
            else:
145
146
                w = self.P(abs(self.m)**2)
            self._w = w
147
        return self._w
148

149
150
151
    @property
    @memo
    def _theta(self):
152
        return exp(-self.position) * (self.q + self.w / 2.)
153
154
155
156
157
158
159
160
161
162

    @property
    @memo
    def _rho_prime(self):
        return self.alpha - 1. + self.rho / 2.

    @property
    @memo
    def _Tt(self):
        return self.T(self.position)