opener-opinion-detector-basic 2.0.0 → 2.0.1
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/README.md +2 -0
- data/ext/hack/Rakefile +0 -2
- data/lib/opener/opinion_detector_basic/version.rb +1 -1
- data/opener-opinion-detector-basic.gemspec +0 -1
- data/task/compile.rake +1 -1
- data/task/requirements.rake +0 -1
- metadata +2 -142
- data/core/vendor/src/crfsuite/AUTHORS +0 -1
- data/core/vendor/src/crfsuite/COPYING +0 -27
- data/core/vendor/src/crfsuite/ChangeLog +0 -103
- data/core/vendor/src/crfsuite/INSTALL +0 -236
- data/core/vendor/src/crfsuite/Makefile.am +0 -19
- data/core/vendor/src/crfsuite/Makefile.in +0 -783
- data/core/vendor/src/crfsuite/README +0 -183
- data/core/vendor/src/crfsuite/aclocal.m4 +0 -9018
- data/core/vendor/src/crfsuite/autogen.sh +0 -38
- data/core/vendor/src/crfsuite/compile +0 -143
- data/core/vendor/src/crfsuite/config.guess +0 -1502
- data/core/vendor/src/crfsuite/config.h.in +0 -198
- data/core/vendor/src/crfsuite/config.sub +0 -1714
- data/core/vendor/src/crfsuite/configure +0 -14273
- data/core/vendor/src/crfsuite/configure.in +0 -149
- data/core/vendor/src/crfsuite/crfsuite.sln +0 -42
- data/core/vendor/src/crfsuite/depcomp +0 -630
- data/core/vendor/src/crfsuite/example/chunking.py +0 -49
- data/core/vendor/src/crfsuite/example/crfutils.py +0 -179
- data/core/vendor/src/crfsuite/example/ner.py +0 -270
- data/core/vendor/src/crfsuite/example/pos.py +0 -78
- data/core/vendor/src/crfsuite/example/template.py +0 -88
- data/core/vendor/src/crfsuite/frontend/Makefile.am +0 -29
- data/core/vendor/src/crfsuite/frontend/Makefile.in +0 -640
- data/core/vendor/src/crfsuite/frontend/dump.c +0 -116
- data/core/vendor/src/crfsuite/frontend/frontend.vcxproj +0 -129
- data/core/vendor/src/crfsuite/frontend/iwa.c +0 -273
- data/core/vendor/src/crfsuite/frontend/iwa.h +0 -65
- data/core/vendor/src/crfsuite/frontend/learn.c +0 -439
- data/core/vendor/src/crfsuite/frontend/main.c +0 -137
- data/core/vendor/src/crfsuite/frontend/option.c +0 -93
- data/core/vendor/src/crfsuite/frontend/option.h +0 -86
- data/core/vendor/src/crfsuite/frontend/readdata.h +0 -38
- data/core/vendor/src/crfsuite/frontend/reader.c +0 -136
- data/core/vendor/src/crfsuite/frontend/tag.c +0 -427
- data/core/vendor/src/crfsuite/genbinary.sh.in +0 -15
- data/core/vendor/src/crfsuite/include/Makefile.am +0 -11
- data/core/vendor/src/crfsuite/include/Makefile.in +0 -461
- data/core/vendor/src/crfsuite/include/crfsuite.h +0 -1063
- data/core/vendor/src/crfsuite/include/crfsuite.hpp +0 -555
- data/core/vendor/src/crfsuite/include/crfsuite_api.hpp +0 -400
- data/core/vendor/src/crfsuite/include/os.h +0 -61
- data/core/vendor/src/crfsuite/install-sh +0 -520
- data/core/vendor/src/crfsuite/lib/cqdb/COPYING +0 -28
- data/core/vendor/src/crfsuite/lib/cqdb/Makefile.am +0 -21
- data/core/vendor/src/crfsuite/lib/cqdb/Makefile.in +0 -549
- data/core/vendor/src/crfsuite/lib/cqdb/cqdb.vcxproj +0 -86
- data/core/vendor/src/crfsuite/lib/cqdb/include/cqdb.h +0 -524
- data/core/vendor/src/crfsuite/lib/cqdb/src/cqdb.c +0 -587
- data/core/vendor/src/crfsuite/lib/cqdb/src/lookup3.c +0 -976
- data/core/vendor/src/crfsuite/lib/crf/Makefile.am +0 -46
- data/core/vendor/src/crfsuite/lib/crf/Makefile.in +0 -721
- data/core/vendor/src/crfsuite/lib/crf/crf.vcxproj +0 -216
- data/core/vendor/src/crfsuite/lib/crf/src/crf1d.h +0 -353
- data/core/vendor/src/crfsuite/lib/crf/src/crf1d_context.c +0 -705
- data/core/vendor/src/crfsuite/lib/crf/src/crf1d_encode.c +0 -943
- data/core/vendor/src/crfsuite/lib/crf/src/crf1d_feature.c +0 -352
- data/core/vendor/src/crfsuite/lib/crf/src/crf1d_model.c +0 -994
- data/core/vendor/src/crfsuite/lib/crf/src/crf1d_tag.c +0 -550
- data/core/vendor/src/crfsuite/lib/crf/src/crfsuite.c +0 -492
- data/core/vendor/src/crfsuite/lib/crf/src/crfsuite_internal.h +0 -236
- data/core/vendor/src/crfsuite/lib/crf/src/crfsuite_train.c +0 -272
- data/core/vendor/src/crfsuite/lib/crf/src/dataset.c +0 -106
- data/core/vendor/src/crfsuite/lib/crf/src/dictionary.c +0 -118
- data/core/vendor/src/crfsuite/lib/crf/src/holdout.c +0 -80
- data/core/vendor/src/crfsuite/lib/crf/src/logging.c +0 -91
- data/core/vendor/src/crfsuite/lib/crf/src/logging.h +0 -48
- data/core/vendor/src/crfsuite/lib/crf/src/params.c +0 -335
- data/core/vendor/src/crfsuite/lib/crf/src/params.h +0 -80
- data/core/vendor/src/crfsuite/lib/crf/src/quark.c +0 -172
- data/core/vendor/src/crfsuite/lib/crf/src/quark.h +0 -46
- data/core/vendor/src/crfsuite/lib/crf/src/rumavl.c +0 -1107
- data/core/vendor/src/crfsuite/lib/crf/src/rumavl.h +0 -160
- data/core/vendor/src/crfsuite/lib/crf/src/train_arow.c +0 -408
- data/core/vendor/src/crfsuite/lib/crf/src/train_averaged_perceptron.c +0 -242
- data/core/vendor/src/crfsuite/lib/crf/src/train_l2sgd.c +0 -507
- data/core/vendor/src/crfsuite/lib/crf/src/train_lbfgs.c +0 -338
- data/core/vendor/src/crfsuite/lib/crf/src/train_passive_aggressive.c +0 -435
- data/core/vendor/src/crfsuite/lib/crf/src/vecmath.h +0 -341
- data/core/vendor/src/crfsuite/ltmain.sh +0 -8413
- data/core/vendor/src/crfsuite/missing +0 -376
- data/core/vendor/src/crfsuite/swig/Makefile.am +0 -13
- data/core/vendor/src/crfsuite/swig/Makefile.in +0 -365
- data/core/vendor/src/crfsuite/swig/crfsuite.cpp +0 -2
- data/core/vendor/src/crfsuite/swig/export.i +0 -32
- data/core/vendor/src/crfsuite/swig/python/README +0 -92
- data/core/vendor/src/crfsuite/swig/python/crfsuite.py +0 -329
- data/core/vendor/src/crfsuite/swig/python/export_wrap.cpp +0 -14355
- data/core/vendor/src/crfsuite/swig/python/export_wrap.h +0 -63
- data/core/vendor/src/crfsuite/swig/python/prepare.sh +0 -9
- data/core/vendor/src/crfsuite/swig/python/sample_tag.py +0 -52
- data/core/vendor/src/crfsuite/swig/python/sample_train.py +0 -68
- data/core/vendor/src/crfsuite/swig/python/setup.py +0 -44
- data/core/vendor/src/crfsuite/win32/stdint.h +0 -679
- data/core/vendor/src/liblbfgs/AUTHORS +0 -1
- data/core/vendor/src/liblbfgs/COPYING +0 -22
- data/core/vendor/src/liblbfgs/ChangeLog +0 -120
- data/core/vendor/src/liblbfgs/INSTALL +0 -231
- data/core/vendor/src/liblbfgs/Makefile.am +0 -10
- data/core/vendor/src/liblbfgs/Makefile.in +0 -638
- data/core/vendor/src/liblbfgs/NEWS +0 -0
- data/core/vendor/src/liblbfgs/README +0 -71
- data/core/vendor/src/liblbfgs/aclocal.m4 +0 -6985
- data/core/vendor/src/liblbfgs/autogen.sh +0 -38
- data/core/vendor/src/liblbfgs/config.guess +0 -1411
- data/core/vendor/src/liblbfgs/config.h.in +0 -64
- data/core/vendor/src/liblbfgs/config.sub +0 -1500
- data/core/vendor/src/liblbfgs/configure +0 -21146
- data/core/vendor/src/liblbfgs/configure.in +0 -107
- data/core/vendor/src/liblbfgs/depcomp +0 -522
- data/core/vendor/src/liblbfgs/include/lbfgs.h +0 -745
- data/core/vendor/src/liblbfgs/install-sh +0 -322
- data/core/vendor/src/liblbfgs/lbfgs.sln +0 -26
- data/core/vendor/src/liblbfgs/lib/Makefile.am +0 -24
- data/core/vendor/src/liblbfgs/lib/Makefile.in +0 -499
- data/core/vendor/src/liblbfgs/lib/arithmetic_ansi.h +0 -133
- data/core/vendor/src/liblbfgs/lib/arithmetic_sse_double.h +0 -294
- data/core/vendor/src/liblbfgs/lib/arithmetic_sse_float.h +0 -298
- data/core/vendor/src/liblbfgs/lib/lbfgs.c +0 -1371
- data/core/vendor/src/liblbfgs/lib/lib.vcxproj +0 -95
- data/core/vendor/src/liblbfgs/ltmain.sh +0 -6426
- data/core/vendor/src/liblbfgs/missing +0 -353
- data/core/vendor/src/liblbfgs/sample/Makefile.am +0 -15
- data/core/vendor/src/liblbfgs/sample/Makefile.in +0 -433
- data/core/vendor/src/liblbfgs/sample/sample.c +0 -81
- data/core/vendor/src/liblbfgs/sample/sample.cpp +0 -126
- data/core/vendor/src/liblbfgs/sample/sample.vcxproj +0 -105
- data/core/vendor/src/svm_light/LICENSE.txt +0 -59
- data/core/vendor/src/svm_light/Makefile +0 -105
- data/core/vendor/src/svm_light/kernel.h +0 -40
- data/core/vendor/src/svm_light/svm_classify.c +0 -197
- data/core/vendor/src/svm_light/svm_common.c +0 -985
- data/core/vendor/src/svm_light/svm_common.h +0 -301
- data/core/vendor/src/svm_light/svm_hideo.c +0 -1062
- data/core/vendor/src/svm_light/svm_learn.c +0 -4147
- data/core/vendor/src/svm_light/svm_learn.h +0 -169
- data/core/vendor/src/svm_light/svm_learn_main.c +0 -397
- data/core/vendor/src/svm_light/svm_loqo.c +0 -211
- data/task/c.rake +0 -36
- data/task/submodules.rake +0 -5
@@ -1,1371 +0,0 @@
|
|
1
|
-
/*
|
2
|
-
* Limited memory BFGS (L-BFGS).
|
3
|
-
*
|
4
|
-
* Copyright (c) 1990, Jorge Nocedal
|
5
|
-
* Copyright (c) 2007-2010 Naoaki Okazaki
|
6
|
-
* All rights reserved.
|
7
|
-
*
|
8
|
-
* Permission is hereby granted, free of charge, to any person obtaining a copy
|
9
|
-
* of this software and associated documentation files (the "Software"), to deal
|
10
|
-
* in the Software without restriction, including without limitation the rights
|
11
|
-
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
12
|
-
* copies of the Software, and to permit persons to whom the Software is
|
13
|
-
* furnished to do so, subject to the following conditions:
|
14
|
-
*
|
15
|
-
* The above copyright notice and this permission notice shall be included in
|
16
|
-
* all copies or substantial portions of the Software.
|
17
|
-
*
|
18
|
-
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
19
|
-
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
20
|
-
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
21
|
-
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
22
|
-
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
23
|
-
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
24
|
-
* THE SOFTWARE.
|
25
|
-
*/
|
26
|
-
|
27
|
-
/* $Id$ */
|
28
|
-
|
29
|
-
/*
|
30
|
-
This library is a C port of the FORTRAN implementation of Limited-memory
|
31
|
-
Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) method written by Jorge Nocedal.
|
32
|
-
The original FORTRAN source code is available at:
|
33
|
-
http://www.ece.northwestern.edu/~nocedal/lbfgs.html
|
34
|
-
|
35
|
-
The L-BFGS algorithm is described in:
|
36
|
-
- Jorge Nocedal.
|
37
|
-
Updating Quasi-Newton Matrices with Limited Storage.
|
38
|
-
<i>Mathematics of Computation</i>, Vol. 35, No. 151, pp. 773--782, 1980.
|
39
|
-
- Dong C. Liu and Jorge Nocedal.
|
40
|
-
On the limited memory BFGS method for large scale optimization.
|
41
|
-
<i>Mathematical Programming</i> B, Vol. 45, No. 3, pp. 503-528, 1989.
|
42
|
-
|
43
|
-
The line search algorithms used in this implementation are described in:
|
44
|
-
- John E. Dennis and Robert B. Schnabel.
|
45
|
-
<i>Numerical Methods for Unconstrained Optimization and Nonlinear
|
46
|
-
Equations</i>, Englewood Cliffs, 1983.
|
47
|
-
- Jorge J. More and David J. Thuente.
|
48
|
-
Line search algorithm with guaranteed sufficient decrease.
|
49
|
-
<i>ACM Transactions on Mathematical Software (TOMS)</i>, Vol. 20, No. 3,
|
50
|
-
pp. 286-307, 1994.
|
51
|
-
|
52
|
-
This library also implements Orthant-Wise Limited-memory Quasi-Newton (OWL-QN)
|
53
|
-
method presented in:
|
54
|
-
- Galen Andrew and Jianfeng Gao.
|
55
|
-
Scalable training of L1-regularized log-linear models.
|
56
|
-
In <i>Proceedings of the 24th International Conference on Machine
|
57
|
-
Learning (ICML 2007)</i>, pp. 33-40, 2007.
|
58
|
-
|
59
|
-
I would like to thank the original author, Jorge Nocedal, who has been
|
60
|
-
distributing the effieicnt and explanatory implementation in an open source
|
61
|
-
licence.
|
62
|
-
*/
|
63
|
-
|
64
|
-
#ifdef HAVE_CONFIG_H
|
65
|
-
#include <config.h>
|
66
|
-
#endif/*HAVE_CONFIG_H*/
|
67
|
-
|
68
|
-
#include <stdint.h>
|
69
|
-
#include <stdio.h>
|
70
|
-
#include <stdlib.h>
|
71
|
-
#include <math.h>
|
72
|
-
|
73
|
-
#include <lbfgs.h>
|
74
|
-
|
75
|
-
#ifdef _MSC_VER
|
76
|
-
#define inline __inline
|
77
|
-
#endif/*_MSC_VER*/
|
78
|
-
|
79
|
-
#if defined(USE_SSE) && defined(__SSE2__) && LBFGS_FLOAT == 64
|
80
|
-
/* Use SSE2 optimization for 64bit double precision. */
|
81
|
-
#include "arithmetic_sse_double.h"
|
82
|
-
|
83
|
-
#elif defined(USE_SSE) && defined(__SSE__) && LBFGS_FLOAT == 32
|
84
|
-
/* Use SSE optimization for 32bit float precision. */
|
85
|
-
#include "arithmetic_sse_float.h"
|
86
|
-
|
87
|
-
#else
|
88
|
-
/* No CPU specific optimization. */
|
89
|
-
#include "arithmetic_ansi.h"
|
90
|
-
|
91
|
-
#endif
|
92
|
-
|
93
|
-
#define min2(a, b) ((a) <= (b) ? (a) : (b))
|
94
|
-
#define max2(a, b) ((a) >= (b) ? (a) : (b))
|
95
|
-
#define max3(a, b, c) max2(max2((a), (b)), (c));
|
96
|
-
|
97
|
-
struct tag_callback_data {
|
98
|
-
int n;
|
99
|
-
void *instance;
|
100
|
-
lbfgs_evaluate_t proc_evaluate;
|
101
|
-
lbfgs_progress_t proc_progress;
|
102
|
-
};
|
103
|
-
typedef struct tag_callback_data callback_data_t;
|
104
|
-
|
105
|
-
struct tag_iteration_data {
|
106
|
-
lbfgsfloatval_t alpha;
|
107
|
-
lbfgsfloatval_t *s; /* [n] */
|
108
|
-
lbfgsfloatval_t *y; /* [n] */
|
109
|
-
lbfgsfloatval_t ys; /* vecdot(y, s) */
|
110
|
-
};
|
111
|
-
typedef struct tag_iteration_data iteration_data_t;
|
112
|
-
|
113
|
-
static const lbfgs_parameter_t _defparam = {
|
114
|
-
6, 1e-5, 0, 1e-5,
|
115
|
-
0, LBFGS_LINESEARCH_DEFAULT, 40,
|
116
|
-
1e-20, 1e20, 1e-4, 0.9, 0.9, 1.0e-16,
|
117
|
-
0.0, 0, -1,
|
118
|
-
};
|
119
|
-
|
120
|
-
/* Forward function declarations. */
|
121
|
-
|
122
|
-
typedef int (*line_search_proc)(
|
123
|
-
int n,
|
124
|
-
lbfgsfloatval_t *x,
|
125
|
-
lbfgsfloatval_t *f,
|
126
|
-
lbfgsfloatval_t *g,
|
127
|
-
lbfgsfloatval_t *s,
|
128
|
-
lbfgsfloatval_t *stp,
|
129
|
-
const lbfgsfloatval_t* xp,
|
130
|
-
const lbfgsfloatval_t* gp,
|
131
|
-
lbfgsfloatval_t *wa,
|
132
|
-
callback_data_t *cd,
|
133
|
-
const lbfgs_parameter_t *param
|
134
|
-
);
|
135
|
-
|
136
|
-
static int line_search_backtracking(
|
137
|
-
int n,
|
138
|
-
lbfgsfloatval_t *x,
|
139
|
-
lbfgsfloatval_t *f,
|
140
|
-
lbfgsfloatval_t *g,
|
141
|
-
lbfgsfloatval_t *s,
|
142
|
-
lbfgsfloatval_t *stp,
|
143
|
-
const lbfgsfloatval_t* xp,
|
144
|
-
const lbfgsfloatval_t* gp,
|
145
|
-
lbfgsfloatval_t *wa,
|
146
|
-
callback_data_t *cd,
|
147
|
-
const lbfgs_parameter_t *param
|
148
|
-
);
|
149
|
-
|
150
|
-
static int line_search_backtracking_owlqn(
|
151
|
-
int n,
|
152
|
-
lbfgsfloatval_t *x,
|
153
|
-
lbfgsfloatval_t *f,
|
154
|
-
lbfgsfloatval_t *g,
|
155
|
-
lbfgsfloatval_t *s,
|
156
|
-
lbfgsfloatval_t *stp,
|
157
|
-
const lbfgsfloatval_t* xp,
|
158
|
-
const lbfgsfloatval_t* gp,
|
159
|
-
lbfgsfloatval_t *wp,
|
160
|
-
callback_data_t *cd,
|
161
|
-
const lbfgs_parameter_t *param
|
162
|
-
);
|
163
|
-
|
164
|
-
static int line_search_morethuente(
|
165
|
-
int n,
|
166
|
-
lbfgsfloatval_t *x,
|
167
|
-
lbfgsfloatval_t *f,
|
168
|
-
lbfgsfloatval_t *g,
|
169
|
-
lbfgsfloatval_t *s,
|
170
|
-
lbfgsfloatval_t *stp,
|
171
|
-
const lbfgsfloatval_t* xp,
|
172
|
-
const lbfgsfloatval_t* gp,
|
173
|
-
lbfgsfloatval_t *wa,
|
174
|
-
callback_data_t *cd,
|
175
|
-
const lbfgs_parameter_t *param
|
176
|
-
);
|
177
|
-
|
178
|
-
static int update_trial_interval(
|
179
|
-
lbfgsfloatval_t *x,
|
180
|
-
lbfgsfloatval_t *fx,
|
181
|
-
lbfgsfloatval_t *dx,
|
182
|
-
lbfgsfloatval_t *y,
|
183
|
-
lbfgsfloatval_t *fy,
|
184
|
-
lbfgsfloatval_t *dy,
|
185
|
-
lbfgsfloatval_t *t,
|
186
|
-
lbfgsfloatval_t *ft,
|
187
|
-
lbfgsfloatval_t *dt,
|
188
|
-
const lbfgsfloatval_t tmin,
|
189
|
-
const lbfgsfloatval_t tmax,
|
190
|
-
int *brackt
|
191
|
-
);
|
192
|
-
|
193
|
-
static lbfgsfloatval_t owlqn_x1norm(
|
194
|
-
const lbfgsfloatval_t* x,
|
195
|
-
const int start,
|
196
|
-
const int n
|
197
|
-
);
|
198
|
-
|
199
|
-
static void owlqn_pseudo_gradient(
|
200
|
-
lbfgsfloatval_t* pg,
|
201
|
-
const lbfgsfloatval_t* x,
|
202
|
-
const lbfgsfloatval_t* g,
|
203
|
-
const int n,
|
204
|
-
const lbfgsfloatval_t c,
|
205
|
-
const int start,
|
206
|
-
const int end
|
207
|
-
);
|
208
|
-
|
209
|
-
static void owlqn_project(
|
210
|
-
lbfgsfloatval_t* d,
|
211
|
-
const lbfgsfloatval_t* sign,
|
212
|
-
const int start,
|
213
|
-
const int end
|
214
|
-
);
|
215
|
-
|
216
|
-
|
217
|
-
#if defined(USE_SSE) && (defined(__SSE__) || defined(__SSE2__))
|
218
|
-
static int round_out_variables(int n)
|
219
|
-
{
|
220
|
-
n += 7;
|
221
|
-
n /= 8;
|
222
|
-
n *= 8;
|
223
|
-
return n;
|
224
|
-
}
|
225
|
-
#endif/*defined(USE_SSE)*/
|
226
|
-
|
227
|
-
lbfgsfloatval_t* lbfgs_malloc(int n)
|
228
|
-
{
|
229
|
-
#if defined(USE_SSE) && (defined(__SSE__) || defined(__SSE2__))
|
230
|
-
n = round_out_variables(n);
|
231
|
-
#endif/*defined(USE_SSE)*/
|
232
|
-
return (lbfgsfloatval_t*)vecalloc(sizeof(lbfgsfloatval_t) * n);
|
233
|
-
}
|
234
|
-
|
235
|
-
void lbfgs_free(lbfgsfloatval_t *x)
|
236
|
-
{
|
237
|
-
vecfree(x);
|
238
|
-
}
|
239
|
-
|
240
|
-
void lbfgs_parameter_init(lbfgs_parameter_t *param)
|
241
|
-
{
|
242
|
-
memcpy(param, &_defparam, sizeof(*param));
|
243
|
-
}
|
244
|
-
|
245
|
-
int lbfgs(
|
246
|
-
int n,
|
247
|
-
lbfgsfloatval_t *x,
|
248
|
-
lbfgsfloatval_t *ptr_fx,
|
249
|
-
lbfgs_evaluate_t proc_evaluate,
|
250
|
-
lbfgs_progress_t proc_progress,
|
251
|
-
void *instance,
|
252
|
-
lbfgs_parameter_t *_param
|
253
|
-
)
|
254
|
-
{
|
255
|
-
int ret;
|
256
|
-
int i, j, k, ls, end, bound;
|
257
|
-
lbfgsfloatval_t step;
|
258
|
-
|
259
|
-
/* Constant parameters and their default values. */
|
260
|
-
lbfgs_parameter_t param = (_param != NULL) ? (*_param) : _defparam;
|
261
|
-
const int m = param.m;
|
262
|
-
|
263
|
-
lbfgsfloatval_t *xp = NULL;
|
264
|
-
lbfgsfloatval_t *g = NULL, *gp = NULL, *pg = NULL;
|
265
|
-
lbfgsfloatval_t *d = NULL, *w = NULL, *pf = NULL;
|
266
|
-
iteration_data_t *lm = NULL, *it = NULL;
|
267
|
-
lbfgsfloatval_t ys, yy;
|
268
|
-
lbfgsfloatval_t xnorm, gnorm, beta;
|
269
|
-
lbfgsfloatval_t fx = 0.;
|
270
|
-
lbfgsfloatval_t rate = 0.;
|
271
|
-
line_search_proc linesearch = line_search_morethuente;
|
272
|
-
|
273
|
-
/* Construct a callback data. */
|
274
|
-
callback_data_t cd;
|
275
|
-
cd.n = n;
|
276
|
-
cd.instance = instance;
|
277
|
-
cd.proc_evaluate = proc_evaluate;
|
278
|
-
cd.proc_progress = proc_progress;
|
279
|
-
|
280
|
-
#if defined(USE_SSE) && (defined(__SSE__) || defined(__SSE2__))
|
281
|
-
/* Round out the number of variables. */
|
282
|
-
n = round_out_variables(n);
|
283
|
-
#endif/*defined(USE_SSE)*/
|
284
|
-
|
285
|
-
/* Check the input parameters for errors. */
|
286
|
-
if (n <= 0) {
|
287
|
-
return LBFGSERR_INVALID_N;
|
288
|
-
}
|
289
|
-
#if defined(USE_SSE) && (defined(__SSE__) || defined(__SSE2__))
|
290
|
-
if (n % 8 != 0) {
|
291
|
-
return LBFGSERR_INVALID_N_SSE;
|
292
|
-
}
|
293
|
-
if ((uintptr_t)(const void*)x % 16 != 0) {
|
294
|
-
return LBFGSERR_INVALID_X_SSE;
|
295
|
-
}
|
296
|
-
#endif/*defined(USE_SSE)*/
|
297
|
-
if (param.epsilon < 0.) {
|
298
|
-
return LBFGSERR_INVALID_EPSILON;
|
299
|
-
}
|
300
|
-
if (param.past < 0) {
|
301
|
-
return LBFGSERR_INVALID_TESTPERIOD;
|
302
|
-
}
|
303
|
-
if (param.delta < 0.) {
|
304
|
-
return LBFGSERR_INVALID_DELTA;
|
305
|
-
}
|
306
|
-
if (param.min_step < 0.) {
|
307
|
-
return LBFGSERR_INVALID_MINSTEP;
|
308
|
-
}
|
309
|
-
if (param.max_step < param.min_step) {
|
310
|
-
return LBFGSERR_INVALID_MAXSTEP;
|
311
|
-
}
|
312
|
-
if (param.ftol < 0.) {
|
313
|
-
return LBFGSERR_INVALID_FTOL;
|
314
|
-
}
|
315
|
-
if (param.linesearch == LBFGS_LINESEARCH_BACKTRACKING_WOLFE ||
|
316
|
-
param.linesearch == LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE) {
|
317
|
-
if (param.wolfe <= param.ftol || 1. <= param.wolfe) {
|
318
|
-
return LBFGSERR_INVALID_WOLFE;
|
319
|
-
}
|
320
|
-
}
|
321
|
-
if (param.gtol < 0.) {
|
322
|
-
return LBFGSERR_INVALID_GTOL;
|
323
|
-
}
|
324
|
-
if (param.xtol < 0.) {
|
325
|
-
return LBFGSERR_INVALID_XTOL;
|
326
|
-
}
|
327
|
-
if (param.max_linesearch <= 0) {
|
328
|
-
return LBFGSERR_INVALID_MAXLINESEARCH;
|
329
|
-
}
|
330
|
-
if (param.orthantwise_c < 0.) {
|
331
|
-
return LBFGSERR_INVALID_ORTHANTWISE;
|
332
|
-
}
|
333
|
-
if (param.orthantwise_start < 0 || n < param.orthantwise_start) {
|
334
|
-
return LBFGSERR_INVALID_ORTHANTWISE_START;
|
335
|
-
}
|
336
|
-
if (param.orthantwise_end < 0) {
|
337
|
-
param.orthantwise_end = n;
|
338
|
-
}
|
339
|
-
if (n < param.orthantwise_end) {
|
340
|
-
return LBFGSERR_INVALID_ORTHANTWISE_END;
|
341
|
-
}
|
342
|
-
if (param.orthantwise_c != 0.) {
|
343
|
-
switch (param.linesearch) {
|
344
|
-
case LBFGS_LINESEARCH_BACKTRACKING:
|
345
|
-
linesearch = line_search_backtracking_owlqn;
|
346
|
-
break;
|
347
|
-
default:
|
348
|
-
/* Only the backtracking method is available. */
|
349
|
-
return LBFGSERR_INVALID_LINESEARCH;
|
350
|
-
}
|
351
|
-
} else {
|
352
|
-
switch (param.linesearch) {
|
353
|
-
case LBFGS_LINESEARCH_MORETHUENTE:
|
354
|
-
linesearch = line_search_morethuente;
|
355
|
-
break;
|
356
|
-
case LBFGS_LINESEARCH_BACKTRACKING_ARMIJO:
|
357
|
-
case LBFGS_LINESEARCH_BACKTRACKING_WOLFE:
|
358
|
-
case LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE:
|
359
|
-
linesearch = line_search_backtracking;
|
360
|
-
break;
|
361
|
-
default:
|
362
|
-
return LBFGSERR_INVALID_LINESEARCH;
|
363
|
-
}
|
364
|
-
}
|
365
|
-
|
366
|
-
/* Allocate working space. */
|
367
|
-
xp = (lbfgsfloatval_t*)vecalloc(n * sizeof(lbfgsfloatval_t));
|
368
|
-
g = (lbfgsfloatval_t*)vecalloc(n * sizeof(lbfgsfloatval_t));
|
369
|
-
gp = (lbfgsfloatval_t*)vecalloc(n * sizeof(lbfgsfloatval_t));
|
370
|
-
d = (lbfgsfloatval_t*)vecalloc(n * sizeof(lbfgsfloatval_t));
|
371
|
-
w = (lbfgsfloatval_t*)vecalloc(n * sizeof(lbfgsfloatval_t));
|
372
|
-
if (xp == NULL || g == NULL || gp == NULL || d == NULL || w == NULL) {
|
373
|
-
ret = LBFGSERR_OUTOFMEMORY;
|
374
|
-
goto lbfgs_exit;
|
375
|
-
}
|
376
|
-
|
377
|
-
if (param.orthantwise_c != 0.) {
|
378
|
-
/* Allocate working space for OW-LQN. */
|
379
|
-
pg = (lbfgsfloatval_t*)vecalloc(n * sizeof(lbfgsfloatval_t));
|
380
|
-
if (pg == NULL) {
|
381
|
-
ret = LBFGSERR_OUTOFMEMORY;
|
382
|
-
goto lbfgs_exit;
|
383
|
-
}
|
384
|
-
}
|
385
|
-
|
386
|
-
/* Allocate limited memory storage. */
|
387
|
-
lm = (iteration_data_t*)vecalloc(m * sizeof(iteration_data_t));
|
388
|
-
if (lm == NULL) {
|
389
|
-
ret = LBFGSERR_OUTOFMEMORY;
|
390
|
-
goto lbfgs_exit;
|
391
|
-
}
|
392
|
-
|
393
|
-
/* Initialize the limited memory. */
|
394
|
-
for (i = 0;i < m;++i) {
|
395
|
-
it = &lm[i];
|
396
|
-
it->alpha = 0;
|
397
|
-
it->ys = 0;
|
398
|
-
it->s = (lbfgsfloatval_t*)vecalloc(n * sizeof(lbfgsfloatval_t));
|
399
|
-
it->y = (lbfgsfloatval_t*)vecalloc(n * sizeof(lbfgsfloatval_t));
|
400
|
-
if (it->s == NULL || it->y == NULL) {
|
401
|
-
ret = LBFGSERR_OUTOFMEMORY;
|
402
|
-
goto lbfgs_exit;
|
403
|
-
}
|
404
|
-
}
|
405
|
-
|
406
|
-
/* Allocate an array for storing previous values of the objective function. */
|
407
|
-
if (0 < param.past) {
|
408
|
-
pf = (lbfgsfloatval_t*)vecalloc(param.past * sizeof(lbfgsfloatval_t));
|
409
|
-
}
|
410
|
-
|
411
|
-
/* Evaluate the function value and its gradient. */
|
412
|
-
fx = cd.proc_evaluate(cd.instance, x, g, cd.n, 0);
|
413
|
-
if (0. != param.orthantwise_c) {
|
414
|
-
/* Compute the L1 norm of the variable and add it to the object value. */
|
415
|
-
xnorm = owlqn_x1norm(x, param.orthantwise_start, param.orthantwise_end);
|
416
|
-
fx += xnorm * param.orthantwise_c;
|
417
|
-
owlqn_pseudo_gradient(
|
418
|
-
pg, x, g, n,
|
419
|
-
param.orthantwise_c, param.orthantwise_start, param.orthantwise_end
|
420
|
-
);
|
421
|
-
}
|
422
|
-
|
423
|
-
/* Store the initial value of the objective function. */
|
424
|
-
if (pf != NULL) {
|
425
|
-
pf[0] = fx;
|
426
|
-
}
|
427
|
-
|
428
|
-
/*
|
429
|
-
Compute the direction;
|
430
|
-
we assume the initial hessian matrix H_0 as the identity matrix.
|
431
|
-
*/
|
432
|
-
if (param.orthantwise_c == 0.) {
|
433
|
-
vecncpy(d, g, n);
|
434
|
-
} else {
|
435
|
-
vecncpy(d, pg, n);
|
436
|
-
}
|
437
|
-
|
438
|
-
/*
|
439
|
-
Make sure that the initial variables are not a minimizer.
|
440
|
-
*/
|
441
|
-
vec2norm(&xnorm, x, n);
|
442
|
-
if (param.orthantwise_c == 0.) {
|
443
|
-
vec2norm(&gnorm, g, n);
|
444
|
-
} else {
|
445
|
-
vec2norm(&gnorm, pg, n);
|
446
|
-
}
|
447
|
-
if (xnorm < 1.0) xnorm = 1.0;
|
448
|
-
if (gnorm / xnorm <= param.epsilon) {
|
449
|
-
ret = LBFGS_ALREADY_MINIMIZED;
|
450
|
-
goto lbfgs_exit;
|
451
|
-
}
|
452
|
-
|
453
|
-
/* Compute the initial step:
|
454
|
-
step = 1.0 / sqrt(vecdot(d, d, n))
|
455
|
-
*/
|
456
|
-
vec2norminv(&step, d, n);
|
457
|
-
|
458
|
-
k = 1;
|
459
|
-
end = 0;
|
460
|
-
for (;;) {
|
461
|
-
/* Store the current position and gradient vectors. */
|
462
|
-
veccpy(xp, x, n);
|
463
|
-
veccpy(gp, g, n);
|
464
|
-
|
465
|
-
/* Search for an optimal step. */
|
466
|
-
if (param.orthantwise_c == 0.) {
|
467
|
-
ls = linesearch(n, x, &fx, g, d, &step, xp, gp, w, &cd, ¶m);
|
468
|
-
} else {
|
469
|
-
ls = linesearch(n, x, &fx, g, d, &step, xp, pg, w, &cd, ¶m);
|
470
|
-
owlqn_pseudo_gradient(
|
471
|
-
pg, x, g, n,
|
472
|
-
param.orthantwise_c, param.orthantwise_start, param.orthantwise_end
|
473
|
-
);
|
474
|
-
}
|
475
|
-
if (ls < 0) {
|
476
|
-
/* Revert to the previous point. */
|
477
|
-
veccpy(x, xp, n);
|
478
|
-
veccpy(g, gp, n);
|
479
|
-
ret = ls;
|
480
|
-
goto lbfgs_exit;
|
481
|
-
}
|
482
|
-
|
483
|
-
/* Compute x and g norms. */
|
484
|
-
vec2norm(&xnorm, x, n);
|
485
|
-
if (param.orthantwise_c == 0.) {
|
486
|
-
vec2norm(&gnorm, g, n);
|
487
|
-
} else {
|
488
|
-
vec2norm(&gnorm, pg, n);
|
489
|
-
}
|
490
|
-
|
491
|
-
/* Report the progress. */
|
492
|
-
if (cd.proc_progress) {
|
493
|
-
if ((ret = cd.proc_progress(cd.instance, x, g, fx, xnorm, gnorm, step, cd.n, k, ls))) {
|
494
|
-
goto lbfgs_exit;
|
495
|
-
}
|
496
|
-
}
|
497
|
-
|
498
|
-
/*
|
499
|
-
Convergence test.
|
500
|
-
The criterion is given by the following formula:
|
501
|
-
|g(x)| / \max(1, |x|) < \epsilon
|
502
|
-
*/
|
503
|
-
if (xnorm < 1.0) xnorm = 1.0;
|
504
|
-
if (gnorm / xnorm <= param.epsilon) {
|
505
|
-
/* Convergence. */
|
506
|
-
ret = LBFGS_SUCCESS;
|
507
|
-
break;
|
508
|
-
}
|
509
|
-
|
510
|
-
/*
|
511
|
-
Test for stopping criterion.
|
512
|
-
The criterion is given by the following formula:
|
513
|
-
(f(past_x) - f(x)) / f(x) < \delta
|
514
|
-
*/
|
515
|
-
if (pf != NULL) {
|
516
|
-
/* We don't test the stopping criterion while k < past. */
|
517
|
-
if (param.past <= k) {
|
518
|
-
/* Compute the relative improvement from the past. */
|
519
|
-
rate = (pf[k % param.past] - fx) / fx;
|
520
|
-
|
521
|
-
/* The stopping criterion. */
|
522
|
-
if (rate < param.delta) {
|
523
|
-
ret = LBFGS_STOP;
|
524
|
-
break;
|
525
|
-
}
|
526
|
-
}
|
527
|
-
|
528
|
-
/* Store the current value of the objective function. */
|
529
|
-
pf[k % param.past] = fx;
|
530
|
-
}
|
531
|
-
|
532
|
-
if (param.max_iterations != 0 && param.max_iterations < k+1) {
|
533
|
-
/* Maximum number of iterations. */
|
534
|
-
ret = LBFGSERR_MAXIMUMITERATION;
|
535
|
-
break;
|
536
|
-
}
|
537
|
-
|
538
|
-
/*
|
539
|
-
Update vectors s and y:
|
540
|
-
s_{k+1} = x_{k+1} - x_{k} = \step * d_{k}.
|
541
|
-
y_{k+1} = g_{k+1} - g_{k}.
|
542
|
-
*/
|
543
|
-
it = &lm[end];
|
544
|
-
vecdiff(it->s, x, xp, n);
|
545
|
-
vecdiff(it->y, g, gp, n);
|
546
|
-
|
547
|
-
/*
|
548
|
-
Compute scalars ys and yy:
|
549
|
-
ys = y^t \cdot s = 1 / \rho.
|
550
|
-
yy = y^t \cdot y.
|
551
|
-
Notice that yy is used for scaling the hessian matrix H_0 (Cholesky factor).
|
552
|
-
*/
|
553
|
-
vecdot(&ys, it->y, it->s, n);
|
554
|
-
vecdot(&yy, it->y, it->y, n);
|
555
|
-
it->ys = ys;
|
556
|
-
|
557
|
-
/*
|
558
|
-
Recursive formula to compute dir = -(H \cdot g).
|
559
|
-
This is described in page 779 of:
|
560
|
-
Jorge Nocedal.
|
561
|
-
Updating Quasi-Newton Matrices with Limited Storage.
|
562
|
-
Mathematics of Computation, Vol. 35, No. 151,
|
563
|
-
pp. 773--782, 1980.
|
564
|
-
*/
|
565
|
-
bound = (m <= k) ? m : k;
|
566
|
-
++k;
|
567
|
-
end = (end + 1) % m;
|
568
|
-
|
569
|
-
/* Compute the steepest direction. */
|
570
|
-
if (param.orthantwise_c == 0.) {
|
571
|
-
/* Compute the negative of gradients. */
|
572
|
-
vecncpy(d, g, n);
|
573
|
-
} else {
|
574
|
-
vecncpy(d, pg, n);
|
575
|
-
}
|
576
|
-
|
577
|
-
j = end;
|
578
|
-
for (i = 0;i < bound;++i) {
|
579
|
-
j = (j + m - 1) % m; /* if (--j == -1) j = m-1; */
|
580
|
-
it = &lm[j];
|
581
|
-
/* \alpha_{j} = \rho_{j} s^{t}_{j} \cdot q_{k+1}. */
|
582
|
-
vecdot(&it->alpha, it->s, d, n);
|
583
|
-
it->alpha /= it->ys;
|
584
|
-
/* q_{i} = q_{i+1} - \alpha_{i} y_{i}. */
|
585
|
-
vecadd(d, it->y, -it->alpha, n);
|
586
|
-
}
|
587
|
-
|
588
|
-
vecscale(d, ys / yy, n);
|
589
|
-
|
590
|
-
for (i = 0;i < bound;++i) {
|
591
|
-
it = &lm[j];
|
592
|
-
/* \beta_{j} = \rho_{j} y^t_{j} \cdot \gamma_{i}. */
|
593
|
-
vecdot(&beta, it->y, d, n);
|
594
|
-
beta /= it->ys;
|
595
|
-
/* \gamma_{i+1} = \gamma_{i} + (\alpha_{j} - \beta_{j}) s_{j}. */
|
596
|
-
vecadd(d, it->s, it->alpha - beta, n);
|
597
|
-
j = (j + 1) % m; /* if (++j == m) j = 0; */
|
598
|
-
}
|
599
|
-
|
600
|
-
/*
|
601
|
-
Constrain the search direction for orthant-wise updates.
|
602
|
-
*/
|
603
|
-
if (param.orthantwise_c != 0.) {
|
604
|
-
for (i = param.orthantwise_start;i < param.orthantwise_end;++i) {
|
605
|
-
if (d[i] * pg[i] >= 0) {
|
606
|
-
d[i] = 0;
|
607
|
-
}
|
608
|
-
}
|
609
|
-
}
|
610
|
-
|
611
|
-
/*
|
612
|
-
Now the search direction d is ready. We try step = 1 first.
|
613
|
-
*/
|
614
|
-
step = 1.0;
|
615
|
-
}
|
616
|
-
|
617
|
-
lbfgs_exit:
|
618
|
-
/* Return the final value of the objective function. */
|
619
|
-
if (ptr_fx != NULL) {
|
620
|
-
*ptr_fx = fx;
|
621
|
-
}
|
622
|
-
|
623
|
-
vecfree(pf);
|
624
|
-
|
625
|
-
/* Free memory blocks used by this function. */
|
626
|
-
if (lm != NULL) {
|
627
|
-
for (i = 0;i < m;++i) {
|
628
|
-
vecfree(lm[i].s);
|
629
|
-
vecfree(lm[i].y);
|
630
|
-
}
|
631
|
-
vecfree(lm);
|
632
|
-
}
|
633
|
-
vecfree(pg);
|
634
|
-
vecfree(w);
|
635
|
-
vecfree(d);
|
636
|
-
vecfree(gp);
|
637
|
-
vecfree(g);
|
638
|
-
vecfree(xp);
|
639
|
-
|
640
|
-
return ret;
|
641
|
-
}
|
642
|
-
|
643
|
-
|
644
|
-
|
645
|
-
static int line_search_backtracking(
|
646
|
-
int n,
|
647
|
-
lbfgsfloatval_t *x,
|
648
|
-
lbfgsfloatval_t *f,
|
649
|
-
lbfgsfloatval_t *g,
|
650
|
-
lbfgsfloatval_t *s,
|
651
|
-
lbfgsfloatval_t *stp,
|
652
|
-
const lbfgsfloatval_t* xp,
|
653
|
-
const lbfgsfloatval_t* gp,
|
654
|
-
lbfgsfloatval_t *wp,
|
655
|
-
callback_data_t *cd,
|
656
|
-
const lbfgs_parameter_t *param
|
657
|
-
)
|
658
|
-
{
|
659
|
-
int count = 0;
|
660
|
-
lbfgsfloatval_t width, dg;
|
661
|
-
lbfgsfloatval_t finit, dginit = 0., dgtest;
|
662
|
-
const lbfgsfloatval_t dec = 0.5, inc = 2.1;
|
663
|
-
|
664
|
-
/* Check the input parameters for errors. */
|
665
|
-
if (*stp <= 0.) {
|
666
|
-
return LBFGSERR_INVALIDPARAMETERS;
|
667
|
-
}
|
668
|
-
|
669
|
-
/* Compute the initial gradient in the search direction. */
|
670
|
-
vecdot(&dginit, g, s, n);
|
671
|
-
|
672
|
-
/* Make sure that s points to a descent direction. */
|
673
|
-
if (0 < dginit) {
|
674
|
-
return LBFGSERR_INCREASEGRADIENT;
|
675
|
-
}
|
676
|
-
|
677
|
-
/* The initial value of the objective function. */
|
678
|
-
finit = *f;
|
679
|
-
dgtest = param->ftol * dginit;
|
680
|
-
|
681
|
-
for (;;) {
|
682
|
-
veccpy(x, xp, n);
|
683
|
-
vecadd(x, s, *stp, n);
|
684
|
-
|
685
|
-
/* Evaluate the function and gradient values. */
|
686
|
-
*f = cd->proc_evaluate(cd->instance, x, g, cd->n, *stp);
|
687
|
-
|
688
|
-
++count;
|
689
|
-
|
690
|
-
if (*f > finit + *stp * dgtest) {
|
691
|
-
width = dec;
|
692
|
-
} else {
|
693
|
-
/* The sufficient decrease condition (Armijo condition). */
|
694
|
-
if (param->linesearch == LBFGS_LINESEARCH_BACKTRACKING_ARMIJO) {
|
695
|
-
/* Exit with the Armijo condition. */
|
696
|
-
return count;
|
697
|
-
}
|
698
|
-
|
699
|
-
/* Check the Wolfe condition. */
|
700
|
-
vecdot(&dg, g, s, n);
|
701
|
-
if (dg < param->wolfe * dginit) {
|
702
|
-
width = inc;
|
703
|
-
} else {
|
704
|
-
if(param->linesearch == LBFGS_LINESEARCH_BACKTRACKING_WOLFE) {
|
705
|
-
/* Exit with the regular Wolfe condition. */
|
706
|
-
return count;
|
707
|
-
}
|
708
|
-
|
709
|
-
/* Check the strong Wolfe condition. */
|
710
|
-
if(dg > -param->wolfe * dginit) {
|
711
|
-
width = dec;
|
712
|
-
} else {
|
713
|
-
/* Exit with the strong Wolfe condition. */
|
714
|
-
return count;
|
715
|
-
}
|
716
|
-
}
|
717
|
-
}
|
718
|
-
|
719
|
-
if (*stp < param->min_step) {
|
720
|
-
/* The step is the minimum value. */
|
721
|
-
return LBFGSERR_MINIMUMSTEP;
|
722
|
-
}
|
723
|
-
if (*stp > param->max_step) {
|
724
|
-
/* The step is the maximum value. */
|
725
|
-
return LBFGSERR_MAXIMUMSTEP;
|
726
|
-
}
|
727
|
-
if (param->max_linesearch <= count) {
|
728
|
-
/* Maximum number of iteration. */
|
729
|
-
return LBFGSERR_MAXIMUMLINESEARCH;
|
730
|
-
}
|
731
|
-
|
732
|
-
(*stp) *= width;
|
733
|
-
}
|
734
|
-
}
|
735
|
-
|
736
|
-
|
737
|
-
|
738
|
-
static int line_search_backtracking_owlqn(
|
739
|
-
int n,
|
740
|
-
lbfgsfloatval_t *x,
|
741
|
-
lbfgsfloatval_t *f,
|
742
|
-
lbfgsfloatval_t *g,
|
743
|
-
lbfgsfloatval_t *s,
|
744
|
-
lbfgsfloatval_t *stp,
|
745
|
-
const lbfgsfloatval_t* xp,
|
746
|
-
const lbfgsfloatval_t* gp,
|
747
|
-
lbfgsfloatval_t *wp,
|
748
|
-
callback_data_t *cd,
|
749
|
-
const lbfgs_parameter_t *param
|
750
|
-
)
|
751
|
-
{
|
752
|
-
int i, count = 0;
|
753
|
-
lbfgsfloatval_t width = 0.5, norm = 0.;
|
754
|
-
lbfgsfloatval_t finit = *f, dgtest;
|
755
|
-
|
756
|
-
/* Check the input parameters for errors. */
|
757
|
-
if (*stp <= 0.) {
|
758
|
-
return LBFGSERR_INVALIDPARAMETERS;
|
759
|
-
}
|
760
|
-
|
761
|
-
/* Choose the orthant for the new point. */
|
762
|
-
for (i = 0;i < n;++i) {
|
763
|
-
wp[i] = (xp[i] == 0.) ? -gp[i] : xp[i];
|
764
|
-
}
|
765
|
-
|
766
|
-
for (;;) {
|
767
|
-
/* Update the current point. */
|
768
|
-
veccpy(x, xp, n);
|
769
|
-
vecadd(x, s, *stp, n);
|
770
|
-
|
771
|
-
/* The current point is projected onto the orthant. */
|
772
|
-
owlqn_project(x, wp, param->orthantwise_start, param->orthantwise_end);
|
773
|
-
|
774
|
-
/* Evaluate the function and gradient values. */
|
775
|
-
*f = cd->proc_evaluate(cd->instance, x, g, cd->n, *stp);
|
776
|
-
|
777
|
-
/* Compute the L1 norm of the variables and add it to the object value. */
|
778
|
-
norm = owlqn_x1norm(x, param->orthantwise_start, param->orthantwise_end);
|
779
|
-
*f += norm * param->orthantwise_c;
|
780
|
-
|
781
|
-
++count;
|
782
|
-
|
783
|
-
dgtest = 0.;
|
784
|
-
for (i = 0;i < n;++i) {
|
785
|
-
dgtest += (x[i] - xp[i]) * gp[i];
|
786
|
-
}
|
787
|
-
|
788
|
-
if (*f <= finit + param->ftol * dgtest) {
|
789
|
-
/* The sufficient decrease condition. */
|
790
|
-
return count;
|
791
|
-
}
|
792
|
-
|
793
|
-
if (*stp < param->min_step) {
|
794
|
-
/* The step is the minimum value. */
|
795
|
-
return LBFGSERR_MINIMUMSTEP;
|
796
|
-
}
|
797
|
-
if (*stp > param->max_step) {
|
798
|
-
/* The step is the maximum value. */
|
799
|
-
return LBFGSERR_MAXIMUMSTEP;
|
800
|
-
}
|
801
|
-
if (param->max_linesearch <= count) {
|
802
|
-
/* Maximum number of iteration. */
|
803
|
-
return LBFGSERR_MAXIMUMLINESEARCH;
|
804
|
-
}
|
805
|
-
|
806
|
-
(*stp) *= width;
|
807
|
-
}
|
808
|
-
}
|
809
|
-
|
810
|
-
|
811
|
-
|
812
|
-
static int line_search_morethuente(
|
813
|
-
int n,
|
814
|
-
lbfgsfloatval_t *x,
|
815
|
-
lbfgsfloatval_t *f,
|
816
|
-
lbfgsfloatval_t *g,
|
817
|
-
lbfgsfloatval_t *s,
|
818
|
-
lbfgsfloatval_t *stp,
|
819
|
-
const lbfgsfloatval_t* xp,
|
820
|
-
const lbfgsfloatval_t* gp,
|
821
|
-
lbfgsfloatval_t *wa,
|
822
|
-
callback_data_t *cd,
|
823
|
-
const lbfgs_parameter_t *param
|
824
|
-
)
|
825
|
-
{
|
826
|
-
int count = 0;
|
827
|
-
int brackt, stage1, uinfo = 0;
|
828
|
-
lbfgsfloatval_t dg;
|
829
|
-
lbfgsfloatval_t stx, fx, dgx;
|
830
|
-
lbfgsfloatval_t sty, fy, dgy;
|
831
|
-
lbfgsfloatval_t fxm, dgxm, fym, dgym, fm, dgm;
|
832
|
-
lbfgsfloatval_t finit, ftest1, dginit, dgtest;
|
833
|
-
lbfgsfloatval_t width, prev_width;
|
834
|
-
lbfgsfloatval_t stmin, stmax;
|
835
|
-
|
836
|
-
/* Check the input parameters for errors. */
|
837
|
-
if (*stp <= 0.) {
|
838
|
-
return LBFGSERR_INVALIDPARAMETERS;
|
839
|
-
}
|
840
|
-
|
841
|
-
/* Compute the initial gradient in the search direction. */
|
842
|
-
vecdot(&dginit, g, s, n);
|
843
|
-
|
844
|
-
/* Make sure that s points to a descent direction. */
|
845
|
-
if (0 < dginit) {
|
846
|
-
return LBFGSERR_INCREASEGRADIENT;
|
847
|
-
}
|
848
|
-
|
849
|
-
/* Initialize local variables. */
|
850
|
-
brackt = 0;
|
851
|
-
stage1 = 1;
|
852
|
-
finit = *f;
|
853
|
-
dgtest = param->ftol * dginit;
|
854
|
-
width = param->max_step - param->min_step;
|
855
|
-
prev_width = 2.0 * width;
|
856
|
-
|
857
|
-
/*
|
858
|
-
The variables stx, fx, dgx contain the values of the step,
|
859
|
-
function, and directional derivative at the best step.
|
860
|
-
The variables sty, fy, dgy contain the value of the step,
|
861
|
-
function, and derivative at the other endpoint of
|
862
|
-
the interval of uncertainty.
|
863
|
-
The variables stp, f, dg contain the values of the step,
|
864
|
-
function, and derivative at the current step.
|
865
|
-
*/
|
866
|
-
stx = sty = 0.;
|
867
|
-
fx = fy = finit;
|
868
|
-
dgx = dgy = dginit;
|
869
|
-
|
870
|
-
for (;;) {
|
871
|
-
/*
|
872
|
-
Set the minimum and maximum steps to correspond to the
|
873
|
-
present interval of uncertainty.
|
874
|
-
*/
|
875
|
-
if (brackt) {
|
876
|
-
stmin = min2(stx, sty);
|
877
|
-
stmax = max2(stx, sty);
|
878
|
-
} else {
|
879
|
-
stmin = stx;
|
880
|
-
stmax = *stp + 4.0 * (*stp - stx);
|
881
|
-
}
|
882
|
-
|
883
|
-
/* Clip the step in the range of [stpmin, stpmax]. */
|
884
|
-
if (*stp < param->min_step) *stp = param->min_step;
|
885
|
-
if (param->max_step < *stp) *stp = param->max_step;
|
886
|
-
|
887
|
-
/*
|
888
|
-
If an unusual termination is to occur then let
|
889
|
-
stp be the lowest point obtained so far.
|
890
|
-
*/
|
891
|
-
if ((brackt && ((*stp <= stmin || stmax <= *stp) || param->max_linesearch <= count + 1 || uinfo != 0)) || (brackt && (stmax - stmin <= param->xtol * stmax))) {
|
892
|
-
*stp = stx;
|
893
|
-
}
|
894
|
-
|
895
|
-
/*
|
896
|
-
Compute the current value of x:
|
897
|
-
x <- x + (*stp) * s.
|
898
|
-
*/
|
899
|
-
veccpy(x, xp, n);
|
900
|
-
vecadd(x, s, *stp, n);
|
901
|
-
|
902
|
-
/* Evaluate the function and gradient values. */
|
903
|
-
*f = cd->proc_evaluate(cd->instance, x, g, cd->n, *stp);
|
904
|
-
vecdot(&dg, g, s, n);
|
905
|
-
|
906
|
-
ftest1 = finit + *stp * dgtest;
|
907
|
-
++count;
|
908
|
-
|
909
|
-
/* Test for errors and convergence. */
|
910
|
-
if (brackt && ((*stp <= stmin || stmax <= *stp) || uinfo != 0)) {
|
911
|
-
/* Rounding errors prevent further progress. */
|
912
|
-
return LBFGSERR_ROUNDING_ERROR;
|
913
|
-
}
|
914
|
-
if (*stp == param->max_step && *f <= ftest1 && dg <= dgtest) {
|
915
|
-
/* The step is the maximum value. */
|
916
|
-
return LBFGSERR_MAXIMUMSTEP;
|
917
|
-
}
|
918
|
-
if (*stp == param->min_step && (ftest1 < *f || dgtest <= dg)) {
|
919
|
-
/* The step is the minimum value. */
|
920
|
-
return LBFGSERR_MINIMUMSTEP;
|
921
|
-
}
|
922
|
-
if (brackt && (stmax - stmin) <= param->xtol * stmax) {
|
923
|
-
/* Relative width of the interval of uncertainty is at most xtol. */
|
924
|
-
return LBFGSERR_WIDTHTOOSMALL;
|
925
|
-
}
|
926
|
-
if (param->max_linesearch <= count) {
|
927
|
-
/* Maximum number of iteration. */
|
928
|
-
return LBFGSERR_MAXIMUMLINESEARCH;
|
929
|
-
}
|
930
|
-
if (*f <= ftest1 && fabs(dg) <= param->gtol * (-dginit)) {
|
931
|
-
/* The sufficient decrease condition and the directional derivative condition hold. */
|
932
|
-
return count;
|
933
|
-
}
|
934
|
-
|
935
|
-
/*
|
936
|
-
In the first stage we seek a step for which the modified
|
937
|
-
function has a nonpositive value and nonnegative derivative.
|
938
|
-
*/
|
939
|
-
if (stage1 && *f <= ftest1 && min2(param->ftol, param->gtol) * dginit <= dg) {
|
940
|
-
stage1 = 0;
|
941
|
-
}
|
942
|
-
|
943
|
-
/*
|
944
|
-
A modified function is used to predict the step only if
|
945
|
-
we have not obtained a step for which the modified
|
946
|
-
function has a nonpositive function value and nonnegative
|
947
|
-
derivative, and if a lower function value has been
|
948
|
-
obtained but the decrease is not sufficient.
|
949
|
-
*/
|
950
|
-
if (stage1 && ftest1 < *f && *f <= fx) {
|
951
|
-
/* Define the modified function and derivative values. */
|
952
|
-
fm = *f - *stp * dgtest;
|
953
|
-
fxm = fx - stx * dgtest;
|
954
|
-
fym = fy - sty * dgtest;
|
955
|
-
dgm = dg - dgtest;
|
956
|
-
dgxm = dgx - dgtest;
|
957
|
-
dgym = dgy - dgtest;
|
958
|
-
|
959
|
-
/*
|
960
|
-
Call update_trial_interval() to update the interval of
|
961
|
-
uncertainty and to compute the new step.
|
962
|
-
*/
|
963
|
-
uinfo = update_trial_interval(
|
964
|
-
&stx, &fxm, &dgxm,
|
965
|
-
&sty, &fym, &dgym,
|
966
|
-
stp, &fm, &dgm,
|
967
|
-
stmin, stmax, &brackt
|
968
|
-
);
|
969
|
-
|
970
|
-
/* Reset the function and gradient values for f. */
|
971
|
-
fx = fxm + stx * dgtest;
|
972
|
-
fy = fym + sty * dgtest;
|
973
|
-
dgx = dgxm + dgtest;
|
974
|
-
dgy = dgym + dgtest;
|
975
|
-
} else {
|
976
|
-
/*
|
977
|
-
Call update_trial_interval() to update the interval of
|
978
|
-
uncertainty and to compute the new step.
|
979
|
-
*/
|
980
|
-
uinfo = update_trial_interval(
|
981
|
-
&stx, &fx, &dgx,
|
982
|
-
&sty, &fy, &dgy,
|
983
|
-
stp, f, &dg,
|
984
|
-
stmin, stmax, &brackt
|
985
|
-
);
|
986
|
-
}
|
987
|
-
|
988
|
-
/*
|
989
|
-
Force a sufficient decrease in the interval of uncertainty.
|
990
|
-
*/
|
991
|
-
if (brackt) {
|
992
|
-
if (0.66 * prev_width <= fabs(sty - stx)) {
|
993
|
-
*stp = stx + 0.5 * (sty - stx);
|
994
|
-
}
|
995
|
-
prev_width = width;
|
996
|
-
width = fabs(sty - stx);
|
997
|
-
}
|
998
|
-
}
|
999
|
-
|
1000
|
-
return LBFGSERR_LOGICERROR;
|
1001
|
-
}
|
1002
|
-
|
1003
|
-
|
1004
|
-
|
1005
|
-
/**
|
1006
|
-
* Define the local variables for computing minimizers.
|
1007
|
-
*/
|
1008
|
-
#define USES_MINIMIZER \
|
1009
|
-
lbfgsfloatval_t a, d, gamma, theta, p, q, r, s;
|
1010
|
-
|
1011
|
-
/**
|
1012
|
-
* Find a minimizer of an interpolated cubic function.
|
1013
|
-
* @param cm The minimizer of the interpolated cubic.
|
1014
|
-
* @param u The value of one point, u.
|
1015
|
-
* @param fu The value of f(u).
|
1016
|
-
* @param du The value of f'(u).
|
1017
|
-
* @param v The value of another point, v.
|
1018
|
-
* @param fv The value of f(v).
|
1019
|
-
* @param du The value of f'(v).
|
1020
|
-
*/
|
1021
|
-
#define CUBIC_MINIMIZER(cm, u, fu, du, v, fv, dv) \
|
1022
|
-
d = (v) - (u); \
|
1023
|
-
theta = ((fu) - (fv)) * 3 / d + (du) + (dv); \
|
1024
|
-
p = fabs(theta); \
|
1025
|
-
q = fabs(du); \
|
1026
|
-
r = fabs(dv); \
|
1027
|
-
s = max3(p, q, r); \
|
1028
|
-
/* gamma = s*sqrt((theta/s)**2 - (du/s) * (dv/s)) */ \
|
1029
|
-
a = theta / s; \
|
1030
|
-
gamma = s * sqrt(a * a - ((du) / s) * ((dv) / s)); \
|
1031
|
-
if ((v) < (u)) gamma = -gamma; \
|
1032
|
-
p = gamma - (du) + theta; \
|
1033
|
-
q = gamma - (du) + gamma + (dv); \
|
1034
|
-
r = p / q; \
|
1035
|
-
(cm) = (u) + r * d;
|
1036
|
-
|
1037
|
-
/**
|
1038
|
-
* Find a minimizer of an interpolated cubic function.
|
1039
|
-
* @param cm The minimizer of the interpolated cubic.
|
1040
|
-
* @param u The value of one point, u.
|
1041
|
-
* @param fu The value of f(u).
|
1042
|
-
* @param du The value of f'(u).
|
1043
|
-
* @param v The value of another point, v.
|
1044
|
-
* @param fv The value of f(v).
|
1045
|
-
* @param du The value of f'(v).
|
1046
|
-
* @param xmin The maximum value.
|
1047
|
-
* @param xmin The minimum value.
|
1048
|
-
*/
|
1049
|
-
#define CUBIC_MINIMIZER2(cm, u, fu, du, v, fv, dv, xmin, xmax) \
|
1050
|
-
d = (v) - (u); \
|
1051
|
-
theta = ((fu) - (fv)) * 3 / d + (du) + (dv); \
|
1052
|
-
p = fabs(theta); \
|
1053
|
-
q = fabs(du); \
|
1054
|
-
r = fabs(dv); \
|
1055
|
-
s = max3(p, q, r); \
|
1056
|
-
/* gamma = s*sqrt((theta/s)**2 - (du/s) * (dv/s)) */ \
|
1057
|
-
a = theta / s; \
|
1058
|
-
gamma = s * sqrt(max2(0, a * a - ((du) / s) * ((dv) / s))); \
|
1059
|
-
if ((u) < (v)) gamma = -gamma; \
|
1060
|
-
p = gamma - (dv) + theta; \
|
1061
|
-
q = gamma - (dv) + gamma + (du); \
|
1062
|
-
r = p / q; \
|
1063
|
-
if (r < 0. && gamma != 0.) { \
|
1064
|
-
(cm) = (v) - r * d; \
|
1065
|
-
} else if (a < 0) { \
|
1066
|
-
(cm) = (xmax); \
|
1067
|
-
} else { \
|
1068
|
-
(cm) = (xmin); \
|
1069
|
-
}
|
1070
|
-
|
1071
|
-
/**
|
1072
|
-
* Find a minimizer of an interpolated quadratic function.
|
1073
|
-
* @param qm The minimizer of the interpolated quadratic.
|
1074
|
-
* @param u The value of one point, u.
|
1075
|
-
* @param fu The value of f(u).
|
1076
|
-
* @param du The value of f'(u).
|
1077
|
-
* @param v The value of another point, v.
|
1078
|
-
* @param fv The value of f(v).
|
1079
|
-
*/
|
1080
|
-
#define QUARD_MINIMIZER(qm, u, fu, du, v, fv) \
|
1081
|
-
a = (v) - (u); \
|
1082
|
-
(qm) = (u) + (du) / (((fu) - (fv)) / a + (du)) / 2 * a;
|
1083
|
-
|
1084
|
-
/**
|
1085
|
-
* Find a minimizer of an interpolated quadratic function.
|
1086
|
-
* @param qm The minimizer of the interpolated quadratic.
|
1087
|
-
* @param u The value of one point, u.
|
1088
|
-
* @param du The value of f'(u).
|
1089
|
-
* @param v The value of another point, v.
|
1090
|
-
* @param dv The value of f'(v).
|
1091
|
-
*/
|
1092
|
-
#define QUARD_MINIMIZER2(qm, u, du, v, dv) \
|
1093
|
-
a = (u) - (v); \
|
1094
|
-
(qm) = (v) + (dv) / ((dv) - (du)) * a;
|
1095
|
-
|
1096
|
-
/**
|
1097
|
-
* Update a safeguarded trial value and interval for line search.
|
1098
|
-
*
|
1099
|
-
* The parameter x represents the step with the least function value.
|
1100
|
-
* The parameter t represents the current step. This function assumes
|
1101
|
-
* that the derivative at the point of x in the direction of the step.
|
1102
|
-
* If the bracket is set to true, the minimizer has been bracketed in
|
1103
|
-
* an interval of uncertainty with endpoints between x and y.
|
1104
|
-
*
|
1105
|
-
* @param x The pointer to the value of one endpoint.
|
1106
|
-
* @param fx The pointer to the value of f(x).
|
1107
|
-
* @param dx The pointer to the value of f'(x).
|
1108
|
-
* @param y The pointer to the value of another endpoint.
|
1109
|
-
* @param fy The pointer to the value of f(y).
|
1110
|
-
* @param dy The pointer to the value of f'(y).
|
1111
|
-
* @param t The pointer to the value of the trial value, t.
|
1112
|
-
* @param ft The pointer to the value of f(t).
|
1113
|
-
* @param dt The pointer to the value of f'(t).
|
1114
|
-
* @param tmin The minimum value for the trial value, t.
|
1115
|
-
* @param tmax The maximum value for the trial value, t.
|
1116
|
-
* @param brackt The pointer to the predicate if the trial value is
|
1117
|
-
* bracketed.
|
1118
|
-
* @retval int Status value. Zero indicates a normal termination.
|
1119
|
-
*
|
1120
|
-
* @see
|
1121
|
-
* Jorge J. More and David J. Thuente. Line search algorithm with
|
1122
|
-
* guaranteed sufficient decrease. ACM Transactions on Mathematical
|
1123
|
-
* Software (TOMS), Vol 20, No 3, pp. 286-307, 1994.
|
1124
|
-
*/
|
1125
|
-
static int update_trial_interval(
|
1126
|
-
lbfgsfloatval_t *x,
|
1127
|
-
lbfgsfloatval_t *fx,
|
1128
|
-
lbfgsfloatval_t *dx,
|
1129
|
-
lbfgsfloatval_t *y,
|
1130
|
-
lbfgsfloatval_t *fy,
|
1131
|
-
lbfgsfloatval_t *dy,
|
1132
|
-
lbfgsfloatval_t *t,
|
1133
|
-
lbfgsfloatval_t *ft,
|
1134
|
-
lbfgsfloatval_t *dt,
|
1135
|
-
const lbfgsfloatval_t tmin,
|
1136
|
-
const lbfgsfloatval_t tmax,
|
1137
|
-
int *brackt
|
1138
|
-
)
|
1139
|
-
{
|
1140
|
-
int bound;
|
1141
|
-
int dsign = fsigndiff(dt, dx);
|
1142
|
-
lbfgsfloatval_t mc; /* minimizer of an interpolated cubic. */
|
1143
|
-
lbfgsfloatval_t mq; /* minimizer of an interpolated quadratic. */
|
1144
|
-
lbfgsfloatval_t newt; /* new trial value. */
|
1145
|
-
USES_MINIMIZER; /* for CUBIC_MINIMIZER and QUARD_MINIMIZER. */
|
1146
|
-
|
1147
|
-
/* Check the input parameters for errors. */
|
1148
|
-
if (*brackt) {
|
1149
|
-
if (*t <= min2(*x, *y) || max2(*x, *y) <= *t) {
|
1150
|
-
/* The trival value t is out of the interval. */
|
1151
|
-
return LBFGSERR_OUTOFINTERVAL;
|
1152
|
-
}
|
1153
|
-
if (0. <= *dx * (*t - *x)) {
|
1154
|
-
/* The function must decrease from x. */
|
1155
|
-
return LBFGSERR_INCREASEGRADIENT;
|
1156
|
-
}
|
1157
|
-
if (tmax < tmin) {
|
1158
|
-
/* Incorrect tmin and tmax specified. */
|
1159
|
-
return LBFGSERR_INCORRECT_TMINMAX;
|
1160
|
-
}
|
1161
|
-
}
|
1162
|
-
|
1163
|
-
/*
|
1164
|
-
Trial value selection.
|
1165
|
-
*/
|
1166
|
-
if (*fx < *ft) {
|
1167
|
-
/*
|
1168
|
-
Case 1: a higher function value.
|
1169
|
-
The minimum is brackt. If the cubic minimizer is closer
|
1170
|
-
to x than the quadratic one, the cubic one is taken, else
|
1171
|
-
the average of the minimizers is taken.
|
1172
|
-
*/
|
1173
|
-
*brackt = 1;
|
1174
|
-
bound = 1;
|
1175
|
-
CUBIC_MINIMIZER(mc, *x, *fx, *dx, *t, *ft, *dt);
|
1176
|
-
QUARD_MINIMIZER(mq, *x, *fx, *dx, *t, *ft);
|
1177
|
-
if (fabs(mc - *x) < fabs(mq - *x)) {
|
1178
|
-
newt = mc;
|
1179
|
-
} else {
|
1180
|
-
newt = mc + 0.5 * (mq - mc);
|
1181
|
-
}
|
1182
|
-
} else if (dsign) {
|
1183
|
-
/*
|
1184
|
-
Case 2: a lower function value and derivatives of
|
1185
|
-
opposite sign. The minimum is brackt. If the cubic
|
1186
|
-
minimizer is closer to x than the quadratic (secant) one,
|
1187
|
-
the cubic one is taken, else the quadratic one is taken.
|
1188
|
-
*/
|
1189
|
-
*brackt = 1;
|
1190
|
-
bound = 0;
|
1191
|
-
CUBIC_MINIMIZER(mc, *x, *fx, *dx, *t, *ft, *dt);
|
1192
|
-
QUARD_MINIMIZER2(mq, *x, *dx, *t, *dt);
|
1193
|
-
if (fabs(mc - *t) > fabs(mq - *t)) {
|
1194
|
-
newt = mc;
|
1195
|
-
} else {
|
1196
|
-
newt = mq;
|
1197
|
-
}
|
1198
|
-
} else if (fabs(*dt) < fabs(*dx)) {
|
1199
|
-
/*
|
1200
|
-
Case 3: a lower function value, derivatives of the
|
1201
|
-
same sign, and the magnitude of the derivative decreases.
|
1202
|
-
The cubic minimizer is only used if the cubic tends to
|
1203
|
-
infinity in the direction of the minimizer or if the minimum
|
1204
|
-
of the cubic is beyond t. Otherwise the cubic minimizer is
|
1205
|
-
defined to be either tmin or tmax. The quadratic (secant)
|
1206
|
-
minimizer is also computed and if the minimum is brackt
|
1207
|
-
then the the minimizer closest to x is taken, else the one
|
1208
|
-
farthest away is taken.
|
1209
|
-
*/
|
1210
|
-
bound = 1;
|
1211
|
-
CUBIC_MINIMIZER2(mc, *x, *fx, *dx, *t, *ft, *dt, tmin, tmax);
|
1212
|
-
QUARD_MINIMIZER2(mq, *x, *dx, *t, *dt);
|
1213
|
-
if (*brackt) {
|
1214
|
-
if (fabs(*t - mc) < fabs(*t - mq)) {
|
1215
|
-
newt = mc;
|
1216
|
-
} else {
|
1217
|
-
newt = mq;
|
1218
|
-
}
|
1219
|
-
} else {
|
1220
|
-
if (fabs(*t - mc) > fabs(*t - mq)) {
|
1221
|
-
newt = mc;
|
1222
|
-
} else {
|
1223
|
-
newt = mq;
|
1224
|
-
}
|
1225
|
-
}
|
1226
|
-
} else {
|
1227
|
-
/*
|
1228
|
-
Case 4: a lower function value, derivatives of the
|
1229
|
-
same sign, and the magnitude of the derivative does
|
1230
|
-
not decrease. If the minimum is not brackt, the step
|
1231
|
-
is either tmin or tmax, else the cubic minimizer is taken.
|
1232
|
-
*/
|
1233
|
-
bound = 0;
|
1234
|
-
if (*brackt) {
|
1235
|
-
CUBIC_MINIMIZER(newt, *t, *ft, *dt, *y, *fy, *dy);
|
1236
|
-
} else if (*x < *t) {
|
1237
|
-
newt = tmax;
|
1238
|
-
} else {
|
1239
|
-
newt = tmin;
|
1240
|
-
}
|
1241
|
-
}
|
1242
|
-
|
1243
|
-
/*
|
1244
|
-
Update the interval of uncertainty. This update does not
|
1245
|
-
depend on the new step or the case analysis above.
|
1246
|
-
|
1247
|
-
- Case a: if f(x) < f(t),
|
1248
|
-
x <- x, y <- t.
|
1249
|
-
- Case b: if f(t) <= f(x) && f'(t)*f'(x) > 0,
|
1250
|
-
x <- t, y <- y.
|
1251
|
-
- Case c: if f(t) <= f(x) && f'(t)*f'(x) < 0,
|
1252
|
-
x <- t, y <- x.
|
1253
|
-
*/
|
1254
|
-
if (*fx < *ft) {
|
1255
|
-
/* Case a */
|
1256
|
-
*y = *t;
|
1257
|
-
*fy = *ft;
|
1258
|
-
*dy = *dt;
|
1259
|
-
} else {
|
1260
|
-
/* Case c */
|
1261
|
-
if (dsign) {
|
1262
|
-
*y = *x;
|
1263
|
-
*fy = *fx;
|
1264
|
-
*dy = *dx;
|
1265
|
-
}
|
1266
|
-
/* Cases b and c */
|
1267
|
-
*x = *t;
|
1268
|
-
*fx = *ft;
|
1269
|
-
*dx = *dt;
|
1270
|
-
}
|
1271
|
-
|
1272
|
-
/* Clip the new trial value in [tmin, tmax]. */
|
1273
|
-
if (tmax < newt) newt = tmax;
|
1274
|
-
if (newt < tmin) newt = tmin;
|
1275
|
-
|
1276
|
-
/*
|
1277
|
-
Redefine the new trial value if it is close to the upper bound
|
1278
|
-
of the interval.
|
1279
|
-
*/
|
1280
|
-
if (*brackt && bound) {
|
1281
|
-
mq = *x + 0.66 * (*y - *x);
|
1282
|
-
if (*x < *y) {
|
1283
|
-
if (mq < newt) newt = mq;
|
1284
|
-
} else {
|
1285
|
-
if (newt < mq) newt = mq;
|
1286
|
-
}
|
1287
|
-
}
|
1288
|
-
|
1289
|
-
/* Return the new trial value. */
|
1290
|
-
*t = newt;
|
1291
|
-
return 0;
|
1292
|
-
}
|
1293
|
-
|
1294
|
-
|
1295
|
-
|
1296
|
-
|
1297
|
-
|
1298
|
-
static lbfgsfloatval_t owlqn_x1norm(
|
1299
|
-
const lbfgsfloatval_t* x,
|
1300
|
-
const int start,
|
1301
|
-
const int n
|
1302
|
-
)
|
1303
|
-
{
|
1304
|
-
int i;
|
1305
|
-
lbfgsfloatval_t norm = 0.;
|
1306
|
-
|
1307
|
-
for (i = start;i < n;++i) {
|
1308
|
-
norm += fabs(x[i]);
|
1309
|
-
}
|
1310
|
-
|
1311
|
-
return norm;
|
1312
|
-
}
|
1313
|
-
|
1314
|
-
static void owlqn_pseudo_gradient(
|
1315
|
-
lbfgsfloatval_t* pg,
|
1316
|
-
const lbfgsfloatval_t* x,
|
1317
|
-
const lbfgsfloatval_t* g,
|
1318
|
-
const int n,
|
1319
|
-
const lbfgsfloatval_t c,
|
1320
|
-
const int start,
|
1321
|
-
const int end
|
1322
|
-
)
|
1323
|
-
{
|
1324
|
-
int i;
|
1325
|
-
|
1326
|
-
/* Compute the negative of gradients. */
|
1327
|
-
for (i = 0;i < start;++i) {
|
1328
|
-
pg[i] = g[i];
|
1329
|
-
}
|
1330
|
-
|
1331
|
-
/* Compute the psuedo-gradients. */
|
1332
|
-
for (i = start;i < end;++i) {
|
1333
|
-
if (x[i] < 0.) {
|
1334
|
-
/* Differentiable. */
|
1335
|
-
pg[i] = g[i] - c;
|
1336
|
-
} else if (0. < x[i]) {
|
1337
|
-
/* Differentiable. */
|
1338
|
-
pg[i] = g[i] + c;
|
1339
|
-
} else {
|
1340
|
-
if (g[i] < -c) {
|
1341
|
-
/* Take the right partial derivative. */
|
1342
|
-
pg[i] = g[i] + c;
|
1343
|
-
} else if (c < g[i]) {
|
1344
|
-
/* Take the left partial derivative. */
|
1345
|
-
pg[i] = g[i] - c;
|
1346
|
-
} else {
|
1347
|
-
pg[i] = 0.;
|
1348
|
-
}
|
1349
|
-
}
|
1350
|
-
}
|
1351
|
-
|
1352
|
-
for (i = end;i < n;++i) {
|
1353
|
-
pg[i] = g[i];
|
1354
|
-
}
|
1355
|
-
}
|
1356
|
-
|
1357
|
-
static void owlqn_project(
|
1358
|
-
lbfgsfloatval_t* d,
|
1359
|
-
const lbfgsfloatval_t* sign,
|
1360
|
-
const int start,
|
1361
|
-
const int end
|
1362
|
-
)
|
1363
|
-
{
|
1364
|
-
int i;
|
1365
|
-
|
1366
|
-
for (i = start;i < end;++i) {
|
1367
|
-
if (d[i] * sign[i] <= 0) {
|
1368
|
-
d[i] = 0;
|
1369
|
-
}
|
1370
|
-
}
|
1371
|
-
}
|