descent_minimizers.py 15.1 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
#
Martin Reinecke's avatar
Martin Reinecke committed
14
# Copyright(C) 2013-2018 Max-Planck-Society
Theo Steininger's avatar
Theo Steininger committed
15
16
17
#
# NIFTy is being developed at the Max-Planck-Institut fuer Astrophysik
# and financially supported by the Studienstiftung des deutschen Volkes.
theos's avatar
theos committed
18

19
from __future__ import absolute_import, division, print_function
20

21
import numpy as np
Philipp Arras's avatar
Philipp Arras committed
22

23
from ..compat import *
24
from ..logger import logger
Martin Reinecke's avatar
Martin Reinecke committed
25
from .line_search_strong_wolfe import LineSearchStrongWolfe
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
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
from .minimizer import Minimizer


class DescentMinimizer(Minimizer):
    """ A base class used by gradient methods to find a local minimum.

    Descent minimization methods are used to find a local minimum of a scalar
    function by following a descent direction. This class implements the
    minimization procedure once a descent direction is known. The descent
    direction has to be implemented separately.

    Parameters
    ----------
    controller : IterationController
        Object that decides when to terminate the minimization.
    line_searcher : callable *optional*
        Function which infers the step size in the descent direction
        (default : LineSearchStrongWolfe()).
    """

    def __init__(self, controller, line_searcher=LineSearchStrongWolfe()):
        self._controller = controller
        self.line_searcher = line_searcher

    def __call__(self, energy):
        """ Performs the minimization of the provided Energy functional.

        Parameters
        ----------
        energy : Energy
           Energy object which provides value, gradient and metric at a
           specific position in parameter space.

        Returns
        -------
        Energy
            Latest `energy` of the minimization.
        int
            Can be controller.CONVERGED or controller.ERROR

        Notes
        -----
        The minimization is stopped if
            * the controller returns controller.CONVERGED or controller.ERROR,
            * a perfectly flat point is reached,
            * according to the line-search the minimum is found,
        """
        f_k_minus_1 = None
        controller = self._controller
        status = controller.start(energy)
        if status != controller.CONTINUE:
            return energy, status

        while True:
            # check if position is at a flat point
            if energy.gradient_norm == 0:
                return energy, controller.CONVERGED

            # compute a step length that reduces energy.value sufficiently
            new_energy, success = self.line_searcher.perform_line_search(
                energy=energy, pk=self.get_descent_direction(energy),
                f_k_minus_1=f_k_minus_1)
            if not success:
                self.reset()

            f_k_minus_1 = energy.value

            if new_energy.value > energy.value:
                logger.error("Error: Energy has increased")
                return energy, controller.ERROR

            if new_energy.value == energy.value:
                logger.warning(
                    "Warning: Energy has not changed. Assuming convergence...")
                return new_energy, controller.CONVERGED

            energy = new_energy
            status = self._controller.check(energy)
            if status != controller.CONTINUE:
                return energy, status

    def reset(self):
        pass

    def get_descent_direction(self, energy):
        """ Calculates the next descent direction.

        Parameters
        ----------
        energy : Energy
            An instance of the Energy class which shall be minimized. The
            position of `energy` is used as the starting point of minimization.

        Returns
        -------
        Field
           The descent direction.
        """
        raise NotImplementedError


class SteepestDescent(DescentMinimizer):
    """ Implementation of the steepest descent minimization scheme.

    Also known as 'gradient descent'. This algorithm simply follows the
    functional's gradient for minimization.
    """

    def get_descent_direction(self, energy):
        return -energy.gradient


Martin Reinecke's avatar
Martin Reinecke committed
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
class RelaxedNewton(DescentMinimizer):
    """ Calculates the descent direction according to a Newton scheme.

    The descent direction is determined by weighting the gradient at the
    current parameter position with the inverse local metric.
    """

    def __init__(self, controller, line_searcher=None):
        if line_searcher is None:
            line_searcher = LineSearchStrongWolfe(
                preferred_initial_step_size=1.)
        super(RelaxedNewton, self).__init__(controller=controller,
                                            line_searcher=line_searcher)

    def get_descent_direction(self, energy):
        return -energy.metric.inverse_times(energy.gradient)


156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
class NewtonCG(DescentMinimizer):
    """ Calculates the descent direction according to a Newton-CG scheme.

    Algorithm derived from SciPy sources.
    """

    def __init__(self, controller, line_searcher=None):
        if line_searcher is None:
            line_searcher = LineSearchStrongWolfe(
                preferred_initial_step_size=1.)
        super(NewtonCG, self).__init__(controller=controller,
                                       line_searcher=line_searcher)

    def get_descent_direction(self, energy):
        float64eps = np.finfo(np.float64).eps
        grad = energy.gradient
        maggrad = abs(grad).sum()
Martin Reinecke's avatar
Martin Reinecke committed
173
        termcond = np.min([0.5, np.sqrt(maggrad)]) * maggrad
174
175
176
177
178
        xsupi = energy.position*0
        ri = grad
        psupi = -ri
        dri0 = ri.vdot(ri)

Martin Reinecke's avatar
Martin Reinecke committed
179
        i = 0
180
181
182
        while True:
            if abs(ri).sum() <= termcond:
                return xsupi
Martin Reinecke's avatar
Martin Reinecke committed
183
            Ap = energy.apply_metric(psupi)
184
185
            # check curvature
            curv = psupi.vdot(Ap)
Martin Reinecke's avatar
Martin Reinecke committed
186
            if 0 <= curv <= 3*float64eps:
187
188
                return xsupi
            elif curv < 0:
Martin Reinecke's avatar
Martin Reinecke committed
189
                return xsupi if i > 0 else (dri0/curv) * grad
190
191
192
193
194
195
196
197
198
199
200
201
202
            alphai = dri0/curv
            xsupi = xsupi + alphai*psupi
            ri = ri + alphai*Ap
            dri1 = ri.vdot(ri)
            psupi = (dri1/dri0)*psupi - ri
            i += 1
            dri0 = dri1  # update numpy.dot(ri,ri) for next time.

        # curvature keeps increasing, bail out
        raise ValueError("Warning: CG iterations didn't converge. "
                         "The Hessian is not positive definite.")


203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
class L_BFGS(DescentMinimizer):
    def __init__(self, controller, line_searcher=LineSearchStrongWolfe(),
                 max_history_length=5):
        super(L_BFGS, self).__init__(controller=controller,
                                     line_searcher=line_searcher)
        self.max_history_length = max_history_length

    def __call__(self, energy):
        self.reset()
        return super(L_BFGS, self).__call__(energy)

    def reset(self):
        self._k = 0
        self._s = [None]*self.max_history_length
        self._y = [None]*self.max_history_length

    def get_descent_direction(self, energy):
        x = energy.position
        s = self._s
        y = self._y
        k = self._k
        maxhist = self.max_history_length
        gradient = energy.gradient

        nhist = min(k, maxhist)
        alpha = [None]*maxhist
        p = -gradient
        if k > 0:
            idx = (k-1) % maxhist
            s[idx] = x-self._lastx
            y[idx] = gradient-self._lastgrad
        if nhist > 0:
            for i in range(k-1, k-nhist-1, -1):
                idx = i % maxhist
                alpha[idx] = s[idx].vdot(p)/s[idx].vdot(y[idx])
                p = p - alpha[idx]*y[idx]
            idx = (k-1) % maxhist
            fact = s[idx].vdot(y[idx]) / y[idx].vdot(y[idx])
            if fact <= 0.:
                logger.error("L-BFGS curvature not positive definite!")
            p = p*fact
            for i in range(k-nhist, k):
                idx = i % maxhist
                beta = y[idx].vdot(p) / s[idx].vdot(y[idx])
                p = p + (alpha[idx]-beta)*s[idx]
        self._lastx = x
        self._lastgrad = gradient
        self._k += 1
        return p
theos's avatar
theos committed
252
253


254
class VL_BFGS(DescentMinimizer):
Martin Reinecke's avatar
Martin Reinecke committed
255
256
257
258
259
260
261
262
263
264
265
266
267
268
    """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.

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

Martin Reinecke's avatar
Martin Reinecke committed
269
270
    def __init__(self, controller, line_searcher=LineSearchStrongWolfe(),
                 max_history_length=5):
Martin Reinecke's avatar
cleanup    
Martin Reinecke committed
271
272
        super(VL_BFGS, self).__init__(controller=controller,
                                      line_searcher=line_searcher)
273
274
        self.max_history_length = max_history_length

275
    def __call__(self, energy):
276
        self._information_store = None
277
        return super(VL_BFGS, self).__call__(energy)
278

Martin Reinecke's avatar
stage 1    
Martin Reinecke committed
279
280
281
    def reset(self):
        self._information_store = None

282
    def get_descent_direction(self, energy):
283
284
        x = energy.position
        gradient = energy.gradient
285
286
287
288
        # initialize the information store if it doesn't already exist
        try:
            self._information_store.add_new_point(x, gradient)
        except AttributeError:
Martin Reinecke's avatar
Martin Reinecke committed
289
290
            self._information_store = _InformationStore(
                self.max_history_length, x0=x, gradient=gradient)
291
292
293
294

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

295
        descent_direction = delta[0] * b[0]
Martin Reinecke's avatar
Martin Reinecke committed
296
        for i in range(1, len(delta)):
297
            descent_direction = descent_direction + delta[i]*b[i]
298

299
        return descent_direction
theos's avatar
theos committed
300
301


Martin Reinecke's avatar
Martin Reinecke committed
302
class _InformationStore(object):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
303
    """Class for storing a list of past updates.
304

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
305
306
    Parameters
    ----------
Martin Reinecke's avatar
Martin Reinecke committed
307
    max_history_length : int
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
308
309
310
311
312
        Maximum number of stored past updates.
    x0 : Field
        Initial position in variable space.
    gradient : Field
        Gradient at position x0.
313

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
314
315
    Attributes
    ----------
Martin Reinecke's avatar
Martin Reinecke committed
316
    max_history_length : int
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
317
318
        Maximum number of stored past updates.
    s : List
Martin Reinecke's avatar
Martin Reinecke committed
319
        Circular buffer of past position differences, which are Fields.
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
320
    y : List
Martin Reinecke's avatar
Martin Reinecke committed
321
        Circular buffer of past gradient differences, which are Fields.
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
322
    last_x : Field
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
323
        Latest position in variable space.
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
324
    last_gradient : Field
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
325
        Gradient at latest position.
Martin Reinecke's avatar
Martin Reinecke committed
326
    k : int
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
327
        Number of updates that have taken place
Martin Reinecke's avatar
Martin Reinecke committed
328
    ss : numpy.ndarray
Martin Reinecke's avatar
Martin Reinecke committed
329
        2D circular buffer of scalar products between different elements of s.
Martin Reinecke's avatar
Martin Reinecke committed
330
    sy : numpy.ndarray
Martin Reinecke's avatar
Martin Reinecke committed
331
        2D circular buffer of scalar products between elements of s and y.
Martin Reinecke's avatar
Martin Reinecke committed
332
    yy : numpy.ndarray
Martin Reinecke's avatar
Martin Reinecke committed
333
        2D circular buffer of scalar products between different elements of y.
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
334
    """
Philipp Arras's avatar
Philipp Arras committed
335

336
337
    def __init__(self, max_history_length, x0, gradient):
        self.max_history_length = max_history_length
338
339
        self.s = [None]*max_history_length
        self.y = [None]*max_history_length
340
341
        self.last_x = x0
        self.last_gradient = gradient
theos's avatar
theos committed
342
        self.k = 0
343

Martin Reinecke's avatar
Martin Reinecke committed
344
        mmax = max_history_length
Martin Reinecke's avatar
Martin Reinecke committed
345
346
347
        self.ss = np.empty((mmax, mmax), dtype=np.float64)
        self.sy = np.empty((mmax, mmax), dtype=np.float64)
        self.yy = np.empty((mmax, mmax), dtype=np.float64)
348
349
350

    @property
    def history_length(self):
Martin Reinecke's avatar
Martin Reinecke committed
351
        """Returns the number of currently stored updates."""
352
353
354
355
        return min(self.k, self.max_history_length)

    @property
    def b(self):
356
357
        """Combines s, y and gradient to form the new base vectors b.

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
358
359
        Returns
        -------
Martin Reinecke's avatar
Martin Reinecke committed
360
        List
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
361
362
            List of new basis vectors.
        """
363
364
        result = []
        m = self.history_length
Martin Reinecke's avatar
Martin Reinecke committed
365
        mmax = self.max_history_length
366

Martin Reinecke's avatar
Martin Reinecke committed
367
        for i in range(m):
Martin Reinecke's avatar
Martin Reinecke committed
368
            result.append(self.s[(self.k-m+i) % mmax])
369

Martin Reinecke's avatar
Martin Reinecke committed
370
        for i in range(m):
Martin Reinecke's avatar
Martin Reinecke committed
371
            result.append(self.y[(self.k-m+i) % mmax])
372
373
374
375
376
377
378

        result.append(self.last_gradient)

        return result

    @property
    def b_dot_b(self):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
379
        """Generates the (2m+1) * (2m+1) scalar matrix.
380

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
381
        The i,j-th element of the matrix is a scalar product between the i-th
382
383
        and j-th base vector.

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
384
385
        Returns
        -------
Martin Reinecke's avatar
Martin Reinecke committed
386
        numpy.ndarray
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
387
388
            Scalar matrix.
        """
389
        m = self.history_length
Martin Reinecke's avatar
Martin Reinecke committed
390
        mmax = self.max_history_length
391
392
393
        k = self.k
        result = np.empty((2*m+1, 2*m+1), dtype=np.float)

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
394
        # update the stores
Martin Reinecke's avatar
Martin Reinecke committed
395
        k1 = (k-1) % mmax
Martin Reinecke's avatar
Martin Reinecke committed
396
        for i in range(m):
Martin Reinecke's avatar
Martin Reinecke committed
397
            kmi = (k-m+i) % mmax
Martin Reinecke's avatar
Martin Reinecke committed
398
399
400
            self.ss[kmi, k1] = self.ss[k1, kmi] = self.s[kmi].vdot(self.s[k1])
            self.yy[kmi, k1] = self.yy[k1, kmi] = self.y[kmi].vdot(self.y[k1])
            self.sy[kmi, k1] = self.s[kmi].vdot(self.y[k1])
Martin Reinecke's avatar
Martin Reinecke committed
401
        for j in range(m-1):
Martin Reinecke's avatar
Martin Reinecke committed
402
403
            kmj = (k-m+j) % mmax
            self.sy[k1, kmj] = self.s[k1].vdot(self.y[kmj])
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
404

Martin Reinecke's avatar
Martin Reinecke committed
405
        for i in range(m):
Martin Reinecke's avatar
Martin Reinecke committed
406
            kmi = (k-m+i) % mmax
Martin Reinecke's avatar
Martin Reinecke committed
407
            for j in range(m):
Martin Reinecke's avatar
Martin Reinecke committed
408
                kmj = (k-m+j) % mmax
Martin Reinecke's avatar
Martin Reinecke committed
409
410
411
                result[i, j] = self.ss[kmi, kmj]
                result[i, m+j] = result[m+j, i] = self.sy[kmi, kmj]
                result[m+i, m+j] = self.yy[kmi, kmj]
412

413
            sgrad_i = self.s[kmi].vdot(self.last_gradient)
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
414
            result[2*m, i] = result[i, 2*m] = sgrad_i
415

Martin Reinecke's avatar
fix    
Martin Reinecke committed
416
            ygrad_i = self.y[kmi].vdot(self.last_gradient)
Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
417
            result[2*m, m+i] = result[m+i, 2*m] = ygrad_i
418

Martin Reinecke's avatar
tweaks    
Martin Reinecke committed
419
        result[2*m, 2*m] = self.last_gradient.norm()
420
        return result
theos's avatar
theos committed
421
422

    @property
423
    def delta(self):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
424
        """Calculates the new scalar coefficients (deltas).
425

Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
426
427
        Returns
        -------
Martin Reinecke's avatar
Martin Reinecke committed
428
        List
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
429
430
            List of the new scalar coefficients (deltas).
        """
431
432
433
434
435
436
437
438
        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)

Martin Reinecke's avatar
Martin Reinecke committed
439
440
        for j in range(m-1, -1, -1):
            delta_b_b = sum([delta[l] * b_dot_b[l, j] for l in range(2*m+1)])
441
442
443
            alpha[j] = delta_b_b/b_dot_b[j, m+j]
            delta[m+j] -= alpha[j]

Martin Reinecke's avatar
Martin Reinecke committed
444
        for i in range(2*m+1):
445
446
            delta[i] *= b_dot_b[m-1, 2*m-1]/b_dot_b[2*m-1, 2*m-1]

Martin Reinecke's avatar
Martin Reinecke committed
447
        for j in range(m):
Martin Reinecke's avatar
Martin Reinecke committed
448
            delta_b_b = sum([delta[l]*b_dot_b[m+j, l] for l in range(2*m+1)])
449
450
451
452
453
            beta = delta_b_b/b_dot_b[j, m+j]
            delta[j] += (alpha[j] - beta)

        return delta

theos's avatar
theos committed
454
    def add_new_point(self, x, gradient):
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
455
456
        """Updates the s list and y list.

Martin Reinecke's avatar
Martin Reinecke committed
457
458
        Calculates the new position and gradient differences and enters them
        into the respective list.
Matevz, Sraml (sraml)'s avatar
Matevz, Sraml (sraml) committed
459
        """
Martin Reinecke's avatar
Martin Reinecke committed
460
461
462
        mmax = self.max_history_length
        self.s[self.k % mmax] = x - self.last_x
        self.y[self.k % mmax] = gradient - self.last_gradient
theos's avatar
theos committed
463

464
465
        self.last_x = x
        self.last_gradient = gradient
theos's avatar
theos committed
466

Martin Reinecke's avatar
fixes    
Martin Reinecke committed
467
        self.k += 1