vl_bfgs.py 8.36 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
# 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/>.
Theo Steininger's avatar
Theo Steininger committed
13
14
15
16
17
#
# Copyright(C) 2013-2017 Max-Planck-Society
#
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik
# and financially supported by the Studienstiftung des deutschen Volkes.
Theo Steininger's avatar
Theo Steininger committed
18

19
20
import numpy as np

21
from .descent_minimizer import DescentMinimizer
22
from .line_searching import LineSearchStrongWolfe
Theo Steininger's avatar
Theo Steininger committed
23
24


25
class VL_BFGS(DescentMinimizer):
26
27
    def __init__(self, line_searcher=LineSearchStrongWolfe(), callback=None,
                 convergence_tolerance=1E-4, convergence_level=3,
28
                 iteration_limit=None, max_history_length=5):
29
30
31
32
33
34
35
36
37
38

        super(VL_BFGS, self).__init__(
                                line_searcher=line_searcher,
                                callback=callback,
                                convergence_tolerance=convergence_tolerance,
                                convergence_level=convergence_level,
                                iteration_limit=iteration_limit)

        self.max_history_length = max_history_length

39
    def __call__(self, energy):
40
        self._information_store = None
41
        return super(VL_BFGS, self).__call__(energy)
42

43
44
45
46
47
48
49
50
51
    def get_descend_direction(self, energy):
        """Implementation of the Vector-free L-BFGS minimization scheme.

        Find the descent direction by using the inverse Hessian.
        Instead of storing the whole matrix, it stores only the last few
        updates, which are used to do operations requiring the inverse
        Hessian product. The updates are represented in a new basis to optimize
        the algorithm.

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
52
53
        Parameters
        ----------
54
55
56
57
        energy : Energy
            An instance of the Energy class which shall be minized. The
            position of `energy` is used as the starting point of minization.

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
58
59
60
        Returns
        -------
        descend_direction : Field
61
62
63
64
65
66
67
            Returns the descent direction.

        References
        ----------
        W. Chen, Z. Wang, J. Zhou, "Large-scale L-BFGS using MapReduce", 2014,
        Microsoft

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
68
        """
69

70
71
        x = energy.position
        gradient = energy.gradient
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
        # initialize the information store if it doesn't already exist
        try:
            self._information_store.add_new_point(x, gradient)
        except AttributeError:
            self._information_store = InformationStore(self.max_history_length,
                                                       x0=x,
                                                       gradient=gradient)

        b = self._information_store.b
        delta = self._information_store.delta

        descend_direction = delta[0] * b[0]
        for i in xrange(1, len(delta)):
            descend_direction += delta[i] * b[i]

        return descend_direction
Theo Steininger's avatar
Theo Steininger committed
88
89
90


class InformationStore(object):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
91
    """Class for storing a list of past updates.
92

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
93
94
95
96
97
98
99
100
    Parameters
    ----------
    max_history_length : integer
        Maximum number of stored past updates.
    x0 : Field
        Initial position in variable space.
    gradient : Field
        Gradient at position x0.
101

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
102
103
104
105
106
107
108
109
110
    Attributes
    ----------
    max_history_length : integer
        Maximum number of stored past updates.
    s : List
        List of past position differences, which are Fields.
    y : List
        List of past gradient differences, which are Fields.
    last_x : Field
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
111
        Latest position in variable space.
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
112
    last_gradient : Field
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
113
        Gradient at latest position.
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
114
    k : integer
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
115
        Number of updates that have taken place
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
116
117
118
119
120
121
    _ss_store : dictionary
        Dictionary of scalar products between different elements of s.
    _sy_store : dictionary
        Dictionary of scalar products between elements of s and y.
    _yy_store : dictionary
        Dictionary of scalar products between different elements of y.
122

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
123
    """
124
125
    def __init__(self, max_history_length, x0, gradient):
        self.max_history_length = max_history_length
126
127
        self.s = [None]*max_history_length
        self.y = [None]*max_history_length
128
129
        self.last_x = x0.copy()
        self.last_gradient = gradient.copy()
Theo Steininger's avatar
Theo Steininger committed
130
        self.k = 0
131

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
132
        hl = max_history_length
Martin Reinecke's avatar
fix    
Martin Reinecke committed
133
134
135
        self._ss_store = np.empty((hl, hl), dtype=np.float64)
        self._sy_store = np.empty((hl, hl), dtype=np.float64)
        self._yy_store = np.empty((hl, hl), dtype=np.float64)
136
137
138

    @property
    def history_length(self):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
139
        """Returns the number of currently stored updates.
140

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
141
        """
142
143
144
145
        return min(self.k, self.max_history_length)

    @property
    def b(self):
146
147
        """Combines s, y and gradient to form the new base vectors b.

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
148
149
150
151
        Returns
        -------
        result : List
            List of new basis vectors.
152

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
153
        """
154
155
156
157
158
159
        result = []
        m = self.history_length
        k = self.k

        s = self.s
        for i in xrange(m):
160
            result.append(s[(k-m+i) % m])
161
162
163

        y = self.y
        for i in xrange(m):
164
            result.append(y[(k-m+i) % m])
165
166
167
168
169
170
171

        result.append(self.last_gradient)

        return result

    @property
    def b_dot_b(self):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
172
        """Generates the (2m+1) * (2m+1) scalar matrix.
173

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
174
        The i,j-th element of the matrix is a scalar product between the i-th
175
176
        and j-th base vector.

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
177
178
        Returns
        -------
179
        result : numpy.ndarray
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
180
            Scalar matrix.
181

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
182
        """
183
184
185
186
        m = self.history_length
        k = self.k
        result = np.empty((2*m+1, 2*m+1), dtype=np.float)

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
187
        # update the stores
Martin Reinecke's avatar
fix    
Martin Reinecke committed
188
189
        if (m > 0):
            k1 = (k-1) % m
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
190

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
191
        for i in xrange(m):
Martin Reinecke's avatar
fix    
Martin Reinecke committed
192
193
            kmi = (k-m+i) % m
            self._ss_store[kmi, k1] = self._ss_store[k1, kmi] \
194
                = self.s[kmi].vdot(self.s[k1])
Martin Reinecke's avatar
fix    
Martin Reinecke committed
195
            self._yy_store[kmi, k1] = self._yy_store[k1, kmi] \
196
197
                = self.y[kmi].vdot(self.y[k1])
            self._sy_store[kmi, k1] = self.s[kmi].vdot(self.y[k1])
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
198
        for j in xrange(m-1):
199
200
            kmj = (k-m+j) % m
            self._sy_store[k1, kmj] = self.s[k1].vdot(self.y[kmj])
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
201

202
        for i in xrange(m):
Martin Reinecke's avatar
fix    
Martin Reinecke committed
203
            kmi = (k-m+i) % m
204
            for j in xrange(m):
Martin Reinecke's avatar
fix    
Martin Reinecke committed
205
                kmj = (k-m+j) % m
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
206
207
                result[i, j] = self._ss_store[kmi, kmj]
                result[i, m+j] = result[m+j, i] = self._sy_store[kmi, kmj]
Martin Reinecke's avatar
fix    
Martin Reinecke committed
208
                result[m+i, m+j] = self._yy_store[kmi, kmj]
209

210
            sgrad_i = self.s[kmi].vdot(self.last_gradient)
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
211
            result[2*m, i] = result[i, 2*m] = sgrad_i
212

Martin Reinecke's avatar
fix    
Martin Reinecke committed
213
            ygrad_i = self.y[kmi].vdot(self.last_gradient)
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
214
            result[2*m, m+i] = result[m+i, 2*m] = ygrad_i
215

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
216
        result[2*m, 2*m] = self.last_gradient.norm()
217
218

        return result
Theo Steininger's avatar
Theo Steininger committed
219
220

    @property
221
    def delta(self):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
222
        """Calculates the new scalar coefficients (deltas).
223

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
224
225
226
227
        Returns
        -------
        delta : List
            List of the new scalar coefficients (deltas).
228

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
229
        """
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
        m = self.history_length
        b_dot_b = self.b_dot_b

        delta = np.zeros(2*m+1, dtype=np.float)
        delta[2*m] = -1

        alpha = np.empty(m, dtype=np.float)

        for j in xrange(m-1, -1, -1):
            delta_b_b = sum([delta[l] * b_dot_b[l, j] for l in xrange(2*m+1)])
            alpha[j] = delta_b_b/b_dot_b[j, m+j]
            delta[m+j] -= alpha[j]

        for i in xrange(2*m+1):
            delta[i] *= b_dot_b[m-1, 2*m-1]/b_dot_b[2*m-1, 2*m-1]

Martin Reinecke's avatar
fix    
Martin Reinecke committed
246
        for j in xrange(m):
247
248
249
250
251
252
            delta_b_b = sum([delta[l]*b_dot_b[m+j, l] for l in xrange(2*m+1)])
            beta = delta_b_b/b_dot_b[j, m+j]
            delta[j] += (alpha[j] - beta)

        return delta

Theo Steininger's avatar
Theo Steininger committed
253
    def add_new_point(self, x, gradient):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
254
255
        """Updates the s list and y list.

256
257
258
        Calculates the new position and gradient differences and adds them to
        the respective list.

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
259
        """
260
261
262
        m = self.max_history_length
        self.s[self.k % m] = x - self.last_x
        self.y[self.k % m] = gradient - self.last_gradient
Theo Steininger's avatar
Theo Steininger committed
263

264
265
        self.last_x = x.copy()
        self.last_gradient = gradient.copy()
Theo Steininger's avatar
Theo Steininger committed
266

Martin Reinecke's avatar
fixes    
Martin Reinecke committed
267
        self.k += 1