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
@@ -1,745 +0,0 @@
|
|
1
|
-
/*
|
2
|
-
* C library of 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
|
-
#ifndef __LBFGS_H__
|
30
|
-
#define __LBFGS_H__
|
31
|
-
|
32
|
-
#ifdef __cplusplus
|
33
|
-
extern "C" {
|
34
|
-
#endif/*__cplusplus*/
|
35
|
-
|
36
|
-
/*
|
37
|
-
* The default precision of floating point values is 64bit (double).
|
38
|
-
*/
|
39
|
-
#ifndef LBFGS_FLOAT
|
40
|
-
#define LBFGS_FLOAT 64
|
41
|
-
#endif/*LBFGS_FLOAT*/
|
42
|
-
|
43
|
-
/*
|
44
|
-
* Activate optimization routines for IEEE754 floating point values.
|
45
|
-
*/
|
46
|
-
#ifndef LBFGS_IEEE_FLOAT
|
47
|
-
#define LBFGS_IEEE_FLOAT 1
|
48
|
-
#endif/*LBFGS_IEEE_FLOAT*/
|
49
|
-
|
50
|
-
#if LBFGS_FLOAT == 32
|
51
|
-
typedef float lbfgsfloatval_t;
|
52
|
-
|
53
|
-
#elif LBFGS_FLOAT == 64
|
54
|
-
typedef double lbfgsfloatval_t;
|
55
|
-
|
56
|
-
#else
|
57
|
-
#error "libLBFGS supports single (float; LBFGS_FLOAT = 32) or double (double; LBFGS_FLOAT=64) precision only."
|
58
|
-
|
59
|
-
#endif
|
60
|
-
|
61
|
-
|
62
|
-
/**
|
63
|
-
* \addtogroup liblbfgs_api libLBFGS API
|
64
|
-
* @{
|
65
|
-
*
|
66
|
-
* The libLBFGS API.
|
67
|
-
*/
|
68
|
-
|
69
|
-
/**
|
70
|
-
* Return values of lbfgs().
|
71
|
-
*
|
72
|
-
* Roughly speaking, a negative value indicates an error.
|
73
|
-
*/
|
74
|
-
enum {
|
75
|
-
/** L-BFGS reaches convergence. */
|
76
|
-
LBFGS_SUCCESS = 0,
|
77
|
-
LBFGS_CONVERGENCE = 0,
|
78
|
-
LBFGS_STOP,
|
79
|
-
/** The initial variables already minimize the objective function. */
|
80
|
-
LBFGS_ALREADY_MINIMIZED,
|
81
|
-
|
82
|
-
/** Unknown error. */
|
83
|
-
LBFGSERR_UNKNOWNERROR = -1024,
|
84
|
-
/** Logic error. */
|
85
|
-
LBFGSERR_LOGICERROR,
|
86
|
-
/** Insufficient memory. */
|
87
|
-
LBFGSERR_OUTOFMEMORY,
|
88
|
-
/** The minimization process has been canceled. */
|
89
|
-
LBFGSERR_CANCELED,
|
90
|
-
/** Invalid number of variables specified. */
|
91
|
-
LBFGSERR_INVALID_N,
|
92
|
-
/** Invalid number of variables (for SSE) specified. */
|
93
|
-
LBFGSERR_INVALID_N_SSE,
|
94
|
-
/** The array x must be aligned to 16 (for SSE). */
|
95
|
-
LBFGSERR_INVALID_X_SSE,
|
96
|
-
/** Invalid parameter lbfgs_parameter_t::epsilon specified. */
|
97
|
-
LBFGSERR_INVALID_EPSILON,
|
98
|
-
/** Invalid parameter lbfgs_parameter_t::past specified. */
|
99
|
-
LBFGSERR_INVALID_TESTPERIOD,
|
100
|
-
/** Invalid parameter lbfgs_parameter_t::delta specified. */
|
101
|
-
LBFGSERR_INVALID_DELTA,
|
102
|
-
/** Invalid parameter lbfgs_parameter_t::linesearch specified. */
|
103
|
-
LBFGSERR_INVALID_LINESEARCH,
|
104
|
-
/** Invalid parameter lbfgs_parameter_t::max_step specified. */
|
105
|
-
LBFGSERR_INVALID_MINSTEP,
|
106
|
-
/** Invalid parameter lbfgs_parameter_t::max_step specified. */
|
107
|
-
LBFGSERR_INVALID_MAXSTEP,
|
108
|
-
/** Invalid parameter lbfgs_parameter_t::ftol specified. */
|
109
|
-
LBFGSERR_INVALID_FTOL,
|
110
|
-
/** Invalid parameter lbfgs_parameter_t::wolfe specified. */
|
111
|
-
LBFGSERR_INVALID_WOLFE,
|
112
|
-
/** Invalid parameter lbfgs_parameter_t::gtol specified. */
|
113
|
-
LBFGSERR_INVALID_GTOL,
|
114
|
-
/** Invalid parameter lbfgs_parameter_t::xtol specified. */
|
115
|
-
LBFGSERR_INVALID_XTOL,
|
116
|
-
/** Invalid parameter lbfgs_parameter_t::max_linesearch specified. */
|
117
|
-
LBFGSERR_INVALID_MAXLINESEARCH,
|
118
|
-
/** Invalid parameter lbfgs_parameter_t::orthantwise_c specified. */
|
119
|
-
LBFGSERR_INVALID_ORTHANTWISE,
|
120
|
-
/** Invalid parameter lbfgs_parameter_t::orthantwise_start specified. */
|
121
|
-
LBFGSERR_INVALID_ORTHANTWISE_START,
|
122
|
-
/** Invalid parameter lbfgs_parameter_t::orthantwise_end specified. */
|
123
|
-
LBFGSERR_INVALID_ORTHANTWISE_END,
|
124
|
-
/** The line-search step went out of the interval of uncertainty. */
|
125
|
-
LBFGSERR_OUTOFINTERVAL,
|
126
|
-
/** A logic error occurred; alternatively, the interval of uncertainty
|
127
|
-
became too small. */
|
128
|
-
LBFGSERR_INCORRECT_TMINMAX,
|
129
|
-
/** A rounding error occurred; alternatively, no line-search step
|
130
|
-
satisfies the sufficient decrease and curvature conditions. */
|
131
|
-
LBFGSERR_ROUNDING_ERROR,
|
132
|
-
/** The line-search step became smaller than lbfgs_parameter_t::min_step. */
|
133
|
-
LBFGSERR_MINIMUMSTEP,
|
134
|
-
/** The line-search step became larger than lbfgs_parameter_t::max_step. */
|
135
|
-
LBFGSERR_MAXIMUMSTEP,
|
136
|
-
/** The line-search routine reaches the maximum number of evaluations. */
|
137
|
-
LBFGSERR_MAXIMUMLINESEARCH,
|
138
|
-
/** The algorithm routine reaches the maximum number of iterations. */
|
139
|
-
LBFGSERR_MAXIMUMITERATION,
|
140
|
-
/** Relative width of the interval of uncertainty is at most
|
141
|
-
lbfgs_parameter_t::xtol. */
|
142
|
-
LBFGSERR_WIDTHTOOSMALL,
|
143
|
-
/** A logic error (negative line-search step) occurred. */
|
144
|
-
LBFGSERR_INVALIDPARAMETERS,
|
145
|
-
/** The current search direction increases the objective function value. */
|
146
|
-
LBFGSERR_INCREASEGRADIENT,
|
147
|
-
};
|
148
|
-
|
149
|
-
/**
|
150
|
-
* Line search algorithms.
|
151
|
-
*/
|
152
|
-
enum {
|
153
|
-
/** The default algorithm (MoreThuente method). */
|
154
|
-
LBFGS_LINESEARCH_DEFAULT = 0,
|
155
|
-
/** MoreThuente method proposd by More and Thuente. */
|
156
|
-
LBFGS_LINESEARCH_MORETHUENTE = 0,
|
157
|
-
/**
|
158
|
-
* Backtracking method with the Armijo condition.
|
159
|
-
* The backtracking method finds the step length such that it satisfies
|
160
|
-
* the sufficient decrease (Armijo) condition,
|
161
|
-
* - f(x + a * d) <= f(x) + lbfgs_parameter_t::ftol * a * g(x)^T d,
|
162
|
-
*
|
163
|
-
* where x is the current point, d is the current search direction, and
|
164
|
-
* a is the step length.
|
165
|
-
*/
|
166
|
-
LBFGS_LINESEARCH_BACKTRACKING_ARMIJO = 1,
|
167
|
-
/** The backtracking method with the defualt (regular Wolfe) condition. */
|
168
|
-
LBFGS_LINESEARCH_BACKTRACKING = 2,
|
169
|
-
/**
|
170
|
-
* Backtracking method with regular Wolfe condition.
|
171
|
-
* The backtracking method finds the step length such that it satisfies
|
172
|
-
* both the Armijo condition (LBFGS_LINESEARCH_BACKTRACKING_ARMIJO)
|
173
|
-
* and the curvature condition,
|
174
|
-
* - g(x + a * d)^T d >= lbfgs_parameter_t::wolfe * g(x)^T d,
|
175
|
-
*
|
176
|
-
* where x is the current point, d is the current search direction, and
|
177
|
-
* a is the step length.
|
178
|
-
*/
|
179
|
-
LBFGS_LINESEARCH_BACKTRACKING_WOLFE = 2,
|
180
|
-
/**
|
181
|
-
* Backtracking method with strong Wolfe condition.
|
182
|
-
* The backtracking method finds the step length such that it satisfies
|
183
|
-
* both the Armijo condition (LBFGS_LINESEARCH_BACKTRACKING_ARMIJO)
|
184
|
-
* and the following condition,
|
185
|
-
* - |g(x + a * d)^T d| <= lbfgs_parameter_t::wolfe * |g(x)^T d|,
|
186
|
-
*
|
187
|
-
* where x is the current point, d is the current search direction, and
|
188
|
-
* a is the step length.
|
189
|
-
*/
|
190
|
-
LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 3,
|
191
|
-
};
|
192
|
-
|
193
|
-
/**
|
194
|
-
* L-BFGS optimization parameters.
|
195
|
-
* Call lbfgs_parameter_init() function to initialize parameters to the
|
196
|
-
* default values.
|
197
|
-
*/
|
198
|
-
typedef struct {
|
199
|
-
/**
|
200
|
-
* The number of corrections to approximate the inverse hessian matrix.
|
201
|
-
* The L-BFGS routine stores the computation results of previous \ref m
|
202
|
-
* iterations to approximate the inverse hessian matrix of the current
|
203
|
-
* iteration. This parameter controls the size of the limited memories
|
204
|
-
* (corrections). The default value is \c 6. Values less than \c 3 are
|
205
|
-
* not recommended. Large values will result in excessive computing time.
|
206
|
-
*/
|
207
|
-
int m;
|
208
|
-
|
209
|
-
/**
|
210
|
-
* Epsilon for convergence test.
|
211
|
-
* This parameter determines the accuracy with which the solution is to
|
212
|
-
* be found. A minimization terminates when
|
213
|
-
* ||g|| < \ref epsilon * max(1, ||x||),
|
214
|
-
* where ||.|| denotes the Euclidean (L2) norm. The default value is
|
215
|
-
* \c 1e-5.
|
216
|
-
*/
|
217
|
-
lbfgsfloatval_t epsilon;
|
218
|
-
|
219
|
-
/**
|
220
|
-
* Distance for delta-based convergence test.
|
221
|
-
* This parameter determines the distance, in iterations, to compute
|
222
|
-
* the rate of decrease of the objective function. If the value of this
|
223
|
-
* parameter is zero, the library does not perform the delta-based
|
224
|
-
* convergence test. The default value is \c 0.
|
225
|
-
*/
|
226
|
-
int past;
|
227
|
-
|
228
|
-
/**
|
229
|
-
* Delta for convergence test.
|
230
|
-
* This parameter determines the minimum rate of decrease of the
|
231
|
-
* objective function. The library stops iterations when the
|
232
|
-
* following condition is met:
|
233
|
-
* (f' - f) / f < \ref delta,
|
234
|
-
* where f' is the objective value of \ref past iterations ago, and f is
|
235
|
-
* the objective value of the current iteration.
|
236
|
-
* The default value is \c 0.
|
237
|
-
*/
|
238
|
-
lbfgsfloatval_t delta;
|
239
|
-
|
240
|
-
/**
|
241
|
-
* The maximum number of iterations.
|
242
|
-
* The lbfgs() function terminates an optimization process with
|
243
|
-
* ::LBFGSERR_MAXIMUMITERATION status code when the iteration count
|
244
|
-
* exceedes this parameter. Setting this parameter to zero continues an
|
245
|
-
* optimization process until a convergence or error. The default value
|
246
|
-
* is \c 0.
|
247
|
-
*/
|
248
|
-
int max_iterations;
|
249
|
-
|
250
|
-
/**
|
251
|
-
* The line search algorithm.
|
252
|
-
* This parameter specifies a line search algorithm to be used by the
|
253
|
-
* L-BFGS routine.
|
254
|
-
*/
|
255
|
-
int linesearch;
|
256
|
-
|
257
|
-
/**
|
258
|
-
* The maximum number of trials for the line search.
|
259
|
-
* This parameter controls the number of function and gradients evaluations
|
260
|
-
* per iteration for the line search routine. The default value is \c 20.
|
261
|
-
*/
|
262
|
-
int max_linesearch;
|
263
|
-
|
264
|
-
/**
|
265
|
-
* The minimum step of the line search routine.
|
266
|
-
* The default value is \c 1e-20. This value need not be modified unless
|
267
|
-
* the exponents are too large for the machine being used, or unless the
|
268
|
-
* problem is extremely badly scaled (in which case the exponents should
|
269
|
-
* be increased).
|
270
|
-
*/
|
271
|
-
lbfgsfloatval_t min_step;
|
272
|
-
|
273
|
-
/**
|
274
|
-
* The maximum step of the line search.
|
275
|
-
* The default value is \c 1e+20. This value need not be modified unless
|
276
|
-
* the exponents are too large for the machine being used, or unless the
|
277
|
-
* problem is extremely badly scaled (in which case the exponents should
|
278
|
-
* be increased).
|
279
|
-
*/
|
280
|
-
lbfgsfloatval_t max_step;
|
281
|
-
|
282
|
-
/**
|
283
|
-
* A parameter to control the accuracy of the line search routine.
|
284
|
-
* The default value is \c 1e-4. This parameter should be greater
|
285
|
-
* than zero and smaller than \c 0.5.
|
286
|
-
*/
|
287
|
-
lbfgsfloatval_t ftol;
|
288
|
-
|
289
|
-
/**
|
290
|
-
* A coefficient for the Wolfe condition.
|
291
|
-
* This parameter is valid only when the backtracking line-search
|
292
|
-
* algorithm is used with the Wolfe condition,
|
293
|
-
* ::LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE or
|
294
|
-
* ::LBFGS_LINESEARCH_BACKTRACKING_WOLFE .
|
295
|
-
* The default value is \c 0.9. This parameter should be greater
|
296
|
-
* the \ref ftol parameter and smaller than \c 1.0.
|
297
|
-
*/
|
298
|
-
lbfgsfloatval_t wolfe;
|
299
|
-
|
300
|
-
/**
|
301
|
-
* A parameter to control the accuracy of the line search routine.
|
302
|
-
* The default value is \c 0.9. If the function and gradient
|
303
|
-
* evaluations are inexpensive with respect to the cost of the
|
304
|
-
* iteration (which is sometimes the case when solving very large
|
305
|
-
* problems) it may be advantageous to set this parameter to a small
|
306
|
-
* value. A typical small value is \c 0.1. This parameter shuold be
|
307
|
-
* greater than the \ref ftol parameter (\c 1e-4) and smaller than
|
308
|
-
* \c 1.0.
|
309
|
-
*/
|
310
|
-
lbfgsfloatval_t gtol;
|
311
|
-
|
312
|
-
/**
|
313
|
-
* The machine precision for floating-point values.
|
314
|
-
* This parameter must be a positive value set by a client program to
|
315
|
-
* estimate the machine precision. The line search routine will terminate
|
316
|
-
* with the status code (::LBFGSERR_ROUNDING_ERROR) if the relative width
|
317
|
-
* of the interval of uncertainty is less than this parameter.
|
318
|
-
*/
|
319
|
-
lbfgsfloatval_t xtol;
|
320
|
-
|
321
|
-
/**
|
322
|
-
* Coeefficient for the L1 norm of variables.
|
323
|
-
* This parameter should be set to zero for standard minimization
|
324
|
-
* problems. Setting this parameter to a positive value activates
|
325
|
-
* Orthant-Wise Limited-memory Quasi-Newton (OWL-QN) method, which
|
326
|
-
* minimizes the objective function F(x) combined with the L1 norm |x|
|
327
|
-
* of the variables, {F(x) + C |x|}. This parameter is the coeefficient
|
328
|
-
* for the |x|, i.e., C. As the L1 norm |x| is not differentiable at
|
329
|
-
* zero, the library modifies function and gradient evaluations from
|
330
|
-
* a client program suitably; a client program thus have only to return
|
331
|
-
* the function value F(x) and gradients G(x) as usual. The default value
|
332
|
-
* is zero.
|
333
|
-
*/
|
334
|
-
lbfgsfloatval_t orthantwise_c;
|
335
|
-
|
336
|
-
/**
|
337
|
-
* Start index for computing L1 norm of the variables.
|
338
|
-
* This parameter is valid only for OWL-QN method
|
339
|
-
* (i.e., \ref orthantwise_c != 0). This parameter b (0 <= b < N)
|
340
|
-
* specifies the index number from which the library computes the
|
341
|
-
* L1 norm of the variables x,
|
342
|
-
* |x| := |x_{b}| + |x_{b+1}| + ... + |x_{N}| .
|
343
|
-
* In other words, variables x_1, ..., x_{b-1} are not used for
|
344
|
-
* computing the L1 norm. Setting b (0 < b < N), one can protect
|
345
|
-
* variables, x_1, ..., x_{b-1} (e.g., a bias term of logistic
|
346
|
-
* regression) from being regularized. The default value is zero.
|
347
|
-
*/
|
348
|
-
int orthantwise_start;
|
349
|
-
|
350
|
-
/**
|
351
|
-
* End index for computing L1 norm of the variables.
|
352
|
-
* This parameter is valid only for OWL-QN method
|
353
|
-
* (i.e., \ref orthantwise_c != 0). This parameter e (0 < e <= N)
|
354
|
-
* specifies the index number at which the library stops computing the
|
355
|
-
* L1 norm of the variables x,
|
356
|
-
*/
|
357
|
-
int orthantwise_end;
|
358
|
-
} lbfgs_parameter_t;
|
359
|
-
|
360
|
-
|
361
|
-
/**
|
362
|
-
* Callback interface to provide objective function and gradient evaluations.
|
363
|
-
*
|
364
|
-
* The lbfgs() function call this function to obtain the values of objective
|
365
|
-
* function and its gradients when needed. A client program must implement
|
366
|
-
* this function to evaluate the values of the objective function and its
|
367
|
-
* gradients, given current values of variables.
|
368
|
-
*
|
369
|
-
* @param instance The user data sent for lbfgs() function by the client.
|
370
|
-
* @param x The current values of variables.
|
371
|
-
* @param g The gradient vector. The callback function must compute
|
372
|
-
* the gradient values for the current variables.
|
373
|
-
* @param n The number of variables.
|
374
|
-
* @param step The current step of the line search routine.
|
375
|
-
* @retval lbfgsfloatval_t The value of the objective function for the current
|
376
|
-
* variables.
|
377
|
-
*/
|
378
|
-
typedef lbfgsfloatval_t (*lbfgs_evaluate_t)(
|
379
|
-
void *instance,
|
380
|
-
const lbfgsfloatval_t *x,
|
381
|
-
lbfgsfloatval_t *g,
|
382
|
-
const int n,
|
383
|
-
const lbfgsfloatval_t step
|
384
|
-
);
|
385
|
-
|
386
|
-
/**
|
387
|
-
* Callback interface to receive the progress of the optimization process.
|
388
|
-
*
|
389
|
-
* The lbfgs() function call this function for each iteration. Implementing
|
390
|
-
* this function, a client program can store or display the current progress
|
391
|
-
* of the optimization process.
|
392
|
-
*
|
393
|
-
* @param instance The user data sent for lbfgs() function by the client.
|
394
|
-
* @param x The current values of variables.
|
395
|
-
* @param g The current gradient values of variables.
|
396
|
-
* @param fx The current value of the objective function.
|
397
|
-
* @param xnorm The Euclidean norm of the variables.
|
398
|
-
* @param gnorm The Euclidean norm of the gradients.
|
399
|
-
* @param step The line-search step used for this iteration.
|
400
|
-
* @param n The number of variables.
|
401
|
-
* @param k The iteration count.
|
402
|
-
* @param ls The number of evaluations called for this iteration.
|
403
|
-
* @retval int Zero to continue the optimization process. Returning a
|
404
|
-
* non-zero value will cancel the optimization process.
|
405
|
-
*/
|
406
|
-
typedef int (*lbfgs_progress_t)(
|
407
|
-
void *instance,
|
408
|
-
const lbfgsfloatval_t *x,
|
409
|
-
const lbfgsfloatval_t *g,
|
410
|
-
const lbfgsfloatval_t fx,
|
411
|
-
const lbfgsfloatval_t xnorm,
|
412
|
-
const lbfgsfloatval_t gnorm,
|
413
|
-
const lbfgsfloatval_t step,
|
414
|
-
int n,
|
415
|
-
int k,
|
416
|
-
int ls
|
417
|
-
);
|
418
|
-
|
419
|
-
/*
|
420
|
-
A user must implement a function compatible with ::lbfgs_evaluate_t (evaluation
|
421
|
-
callback) and pass the pointer to the callback function to lbfgs() arguments.
|
422
|
-
Similarly, a user can implement a function compatible with ::lbfgs_progress_t
|
423
|
-
(progress callback) to obtain the current progress (e.g., variables, function
|
424
|
-
value, ||G||, etc) and to cancel the iteration process if necessary.
|
425
|
-
Implementation of a progress callback is optional: a user can pass \c NULL if
|
426
|
-
progress notification is not necessary.
|
427
|
-
|
428
|
-
In addition, a user must preserve two requirements:
|
429
|
-
- The number of variables must be multiples of 16 (this is not 4).
|
430
|
-
- The memory block of variable array ::x must be aligned to 16.
|
431
|
-
|
432
|
-
This algorithm terminates an optimization
|
433
|
-
when:
|
434
|
-
|
435
|
-
||G|| < \epsilon \cdot \max(1, ||x||) .
|
436
|
-
|
437
|
-
In this formula, ||.|| denotes the Euclidean norm.
|
438
|
-
*/
|
439
|
-
|
440
|
-
/**
|
441
|
-
* Start a L-BFGS optimization.
|
442
|
-
*
|
443
|
-
* @param n The number of variables.
|
444
|
-
* @param x The array of variables. A client program can set
|
445
|
-
* default values for the optimization and receive the
|
446
|
-
* optimization result through this array. This array
|
447
|
-
* must be allocated by ::lbfgs_malloc function
|
448
|
-
* for libLBFGS built with SSE/SSE2 optimization routine
|
449
|
-
* enabled. The library built without SSE/SSE2
|
450
|
-
* optimization does not have such a requirement.
|
451
|
-
* @param ptr_fx The pointer to the variable that receives the final
|
452
|
-
* value of the objective function for the variables.
|
453
|
-
* This argument can be set to \c NULL if the final
|
454
|
-
* value of the objective function is unnecessary.
|
455
|
-
* @param proc_evaluate The callback function to provide function and
|
456
|
-
* gradient evaluations given a current values of
|
457
|
-
* variables. A client program must implement a
|
458
|
-
* callback function compatible with \ref
|
459
|
-
* lbfgs_evaluate_t and pass the pointer to the
|
460
|
-
* callback function.
|
461
|
-
* @param proc_progress The callback function to receive the progress
|
462
|
-
* (the number of iterations, the current value of
|
463
|
-
* the objective function) of the minimization
|
464
|
-
* process. This argument can be set to \c NULL if
|
465
|
-
* a progress report is unnecessary.
|
466
|
-
* @param instance A user data for the client program. The callback
|
467
|
-
* functions will receive the value of this argument.
|
468
|
-
* @param param The pointer to a structure representing parameters for
|
469
|
-
* L-BFGS optimization. A client program can set this
|
470
|
-
* parameter to \c NULL to use the default parameters.
|
471
|
-
* Call lbfgs_parameter_init() function to fill a
|
472
|
-
* structure with the default values.
|
473
|
-
* @retval int The status code. This function returns zero if the
|
474
|
-
* minimization process terminates without an error. A
|
475
|
-
* non-zero value indicates an error.
|
476
|
-
*/
|
477
|
-
int lbfgs(
|
478
|
-
int n,
|
479
|
-
lbfgsfloatval_t *x,
|
480
|
-
lbfgsfloatval_t *ptr_fx,
|
481
|
-
lbfgs_evaluate_t proc_evaluate,
|
482
|
-
lbfgs_progress_t proc_progress,
|
483
|
-
void *instance,
|
484
|
-
lbfgs_parameter_t *param
|
485
|
-
);
|
486
|
-
|
487
|
-
/**
|
488
|
-
* Initialize L-BFGS parameters to the default values.
|
489
|
-
*
|
490
|
-
* Call this function to fill a parameter structure with the default values
|
491
|
-
* and overwrite parameter values if necessary.
|
492
|
-
*
|
493
|
-
* @param param The pointer to the parameter structure.
|
494
|
-
*/
|
495
|
-
void lbfgs_parameter_init(lbfgs_parameter_t *param);
|
496
|
-
|
497
|
-
/**
|
498
|
-
* Allocate an array for variables.
|
499
|
-
*
|
500
|
-
* This function allocates an array of variables for the convenience of
|
501
|
-
* ::lbfgs function; the function has a requreiemt for a variable array
|
502
|
-
* when libLBFGS is built with SSE/SSE2 optimization routines. A user does
|
503
|
-
* not have to use this function for libLBFGS built without SSE/SSE2
|
504
|
-
* optimization.
|
505
|
-
*
|
506
|
-
* @param n The number of variables.
|
507
|
-
*/
|
508
|
-
lbfgsfloatval_t* lbfgs_malloc(int n);
|
509
|
-
|
510
|
-
/**
|
511
|
-
* Free an array of variables.
|
512
|
-
*
|
513
|
-
* @param x The array of variables allocated by ::lbfgs_malloc
|
514
|
-
* function.
|
515
|
-
*/
|
516
|
-
void lbfgs_free(lbfgsfloatval_t *x);
|
517
|
-
|
518
|
-
/** @} */
|
519
|
-
|
520
|
-
#ifdef __cplusplus
|
521
|
-
}
|
522
|
-
#endif/*__cplusplus*/
|
523
|
-
|
524
|
-
|
525
|
-
|
526
|
-
/**
|
527
|
-
@mainpage libLBFGS: a library of Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS)
|
528
|
-
|
529
|
-
@section intro Introduction
|
530
|
-
|
531
|
-
This library is a C port of the implementation of Limited-memory
|
532
|
-
Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) method written by Jorge Nocedal.
|
533
|
-
The original FORTRAN source code is available at:
|
534
|
-
http://www.ece.northwestern.edu/~nocedal/lbfgs.html
|
535
|
-
|
536
|
-
The L-BFGS method solves the unconstrainted minimization problem,
|
537
|
-
|
538
|
-
<pre>
|
539
|
-
minimize F(x), x = (x1, x2, ..., xN),
|
540
|
-
</pre>
|
541
|
-
|
542
|
-
only if the objective function F(x) and its gradient G(x) are computable. The
|
543
|
-
well-known Newton's method requires computation of the inverse of the hessian
|
544
|
-
matrix of the objective function. However, the computational cost for the
|
545
|
-
inverse hessian matrix is expensive especially when the objective function
|
546
|
-
takes a large number of variables. The L-BFGS method iteratively finds a
|
547
|
-
minimizer by approximating the inverse hessian matrix by information from last
|
548
|
-
m iterations. This innovation saves the memory storage and computational time
|
549
|
-
drastically for large-scaled problems.
|
550
|
-
|
551
|
-
Among the various ports of L-BFGS, this library provides several features:
|
552
|
-
- <b>Optimization with L1-norm (Orthant-Wise Limited-memory Quasi-Newton
|
553
|
-
(OWL-QN) method)</b>:
|
554
|
-
In addition to standard minimization problems, the library can minimize
|
555
|
-
a function F(x) combined with L1-norm |x| of the variables,
|
556
|
-
{F(x) + C |x|}, where C is a constant scalar parameter. This feature is
|
557
|
-
useful for estimating parameters of sparse log-linear models (e.g.,
|
558
|
-
logistic regression and maximum entropy) with L1-regularization (or
|
559
|
-
Laplacian prior).
|
560
|
-
- <b>Clean C code</b>:
|
561
|
-
Unlike C codes generated automatically by f2c (Fortran 77 into C converter),
|
562
|
-
this port includes changes based on my interpretations, improvements,
|
563
|
-
optimizations, and clean-ups so that the ported code would be well-suited
|
564
|
-
for a C code. In addition to comments inherited from the original code,
|
565
|
-
a number of comments were added through my interpretations.
|
566
|
-
- <b>Callback interface</b>:
|
567
|
-
The library receives function and gradient values via a callback interface.
|
568
|
-
The library also notifies the progress of the optimization by invoking a
|
569
|
-
callback function. In the original implementation, a user had to set
|
570
|
-
function and gradient values every time the function returns for obtaining
|
571
|
-
updated values.
|
572
|
-
- <b>Thread safe</b>:
|
573
|
-
The library is thread-safe, which is the secondary gain from the callback
|
574
|
-
interface.
|
575
|
-
- <b>Cross platform.</b> The source code can be compiled on Microsoft Visual
|
576
|
-
Studio 2010, GNU C Compiler (gcc), etc.
|
577
|
-
- <b>Configurable precision</b>: A user can choose single-precision (float)
|
578
|
-
or double-precision (double) accuracy by changing ::LBFGS_FLOAT macro.
|
579
|
-
- <b>SSE/SSE2 optimization</b>:
|
580
|
-
This library includes SSE/SSE2 optimization (written in compiler intrinsics)
|
581
|
-
for vector arithmetic operations on Intel/AMD processors. The library uses
|
582
|
-
SSE for float values and SSE2 for double values. The SSE/SSE2 optimization
|
583
|
-
routine is disabled by default.
|
584
|
-
|
585
|
-
This library is used by:
|
586
|
-
- <a href="http://www.chokkan.org/software/crfsuite/">CRFsuite: A fast implementation of Conditional Random Fields (CRFs)</a>
|
587
|
-
- <a href="http://www.chokkan.org/software/classias/">Classias: A collection of machine-learning algorithms for classification</a>
|
588
|
-
- <a href="http://www.public.iastate.edu/~gdancik/mlegp/">mlegp: an R package for maximum likelihood estimates for Gaussian processes</a>
|
589
|
-
- <a href="http://infmath.uibk.ac.at/~matthiasf/imaging2/">imaging2: the imaging2 class library</a>
|
590
|
-
- <a href="http://search.cpan.org/~laye/Algorithm-LBFGS-0.16/">Algorithm::LBFGS - Perl extension for L-BFGS</a>
|
591
|
-
- <a href="http://www.cs.kuleuven.be/~bernd/yap-lbfgs/">YAP-LBFGS (an interface to call libLBFGS from YAP Prolog)</a>
|
592
|
-
|
593
|
-
@section download Download
|
594
|
-
|
595
|
-
- <a href="https://github.com/downloads/chokkan/liblbfgs/liblbfgs-1.10.tar.gz">Source code</a>
|
596
|
-
- <a href="https://github.com/chokkan/liblbfgs">GitHub repository</a>
|
597
|
-
|
598
|
-
libLBFGS is distributed under the term of the
|
599
|
-
<a href="http://opensource.org/licenses/mit-license.php">MIT license</a>.
|
600
|
-
|
601
|
-
@section changelog History
|
602
|
-
- Version 1.10 (2010-12-22):
|
603
|
-
- Fixed compiling errors on Mac OS X; this patch was kindly submitted by
|
604
|
-
Nic Schraudolph.
|
605
|
-
- Reduced compiling warnings on Mac OS X; this patch was kindly submitted
|
606
|
-
by Tamas Nepusz.
|
607
|
-
- Replaced memalign() with posix_memalign().
|
608
|
-
- Updated solution and project files for Microsoft Visual Studio 2010.
|
609
|
-
- Version 1.9 (2010-01-29):
|
610
|
-
- Fixed a mistake in checking the validity of the parameters "ftol" and
|
611
|
-
"wolfe"; this was discovered by Kevin S. Van Horn.
|
612
|
-
- Version 1.8 (2009-07-13):
|
613
|
-
- Accepted the patch submitted by Takashi Imamichi;
|
614
|
-
the backtracking method now has three criteria for choosing the step
|
615
|
-
length:
|
616
|
-
- ::LBFGS_LINESEARCH_BACKTRACKING_ARMIJO: sufficient decrease (Armijo)
|
617
|
-
condition only
|
618
|
-
- ::LBFGS_LINESEARCH_BACKTRACKING_WOLFE: regular Wolfe condition
|
619
|
-
(sufficient decrease condition + curvature condition)
|
620
|
-
- ::LBFGS_LINESEARCH_BACKTRACKING_STRONG_WOLFE: strong Wolfe condition
|
621
|
-
- Updated the documentation to explain the above three criteria.
|
622
|
-
- Version 1.7 (2009-02-28):
|
623
|
-
- Improved OWL-QN routines for stability.
|
624
|
-
- Removed the support of OWL-QN method in MoreThuente algorithm because
|
625
|
-
it accidentally fails in early stages of iterations for some objectives.
|
626
|
-
Because of this change, <b>the OW-LQN method must be used with the
|
627
|
-
backtracking algorithm (::LBFGS_LINESEARCH_BACKTRACKING)</b>, or the
|
628
|
-
library returns ::LBFGSERR_INVALID_LINESEARCH.
|
629
|
-
- Renamed line search algorithms as follows:
|
630
|
-
- ::LBFGS_LINESEARCH_BACKTRACKING: regular Wolfe condition.
|
631
|
-
- ::LBFGS_LINESEARCH_BACKTRACKING_LOOSE: regular Wolfe condition.
|
632
|
-
- ::LBFGS_LINESEARCH_BACKTRACKING_STRONG: strong Wolfe condition.
|
633
|
-
- Source code clean-up.
|
634
|
-
- Version 1.6 (2008-11-02):
|
635
|
-
- Improved line-search algorithm with strong Wolfe condition, which was
|
636
|
-
contributed by Takashi Imamichi. This routine is now default for
|
637
|
-
::LBFGS_LINESEARCH_BACKTRACKING. The previous line search algorithm
|
638
|
-
with regular Wolfe condition is still available as
|
639
|
-
::LBFGS_LINESEARCH_BACKTRACKING_LOOSE.
|
640
|
-
- Configurable stop index for L1-norm computation. A member variable
|
641
|
-
::lbfgs_parameter_t::orthantwise_end was added to specify the index
|
642
|
-
number at which the library stops computing the L1 norm of the
|
643
|
-
variables. This is useful to prevent some variables from being
|
644
|
-
regularized by the OW-LQN method.
|
645
|
-
- A sample program written in C++ (sample/sample.cpp).
|
646
|
-
- Version 1.5 (2008-07-10):
|
647
|
-
- Configurable starting index for L1-norm computation. A member variable
|
648
|
-
::lbfgs_parameter_t::orthantwise_start was added to specify the index
|
649
|
-
number from which the library computes the L1 norm of the variables.
|
650
|
-
This is useful to prevent some variables from being regularized by the
|
651
|
-
OWL-QN method.
|
652
|
-
- Fixed a zero-division error when the initial variables have already
|
653
|
-
been a minimizer (reported by Takashi Imamichi). In this case, the
|
654
|
-
library returns ::LBFGS_ALREADY_MINIMIZED status code.
|
655
|
-
- Defined ::LBFGS_SUCCESS status code as zero; removed unused constants,
|
656
|
-
LBFGSFALSE and LBFGSTRUE.
|
657
|
-
- Fixed a compile error in an implicit down-cast.
|
658
|
-
- Version 1.4 (2008-04-25):
|
659
|
-
- Configurable line search algorithms. A member variable
|
660
|
-
::lbfgs_parameter_t::linesearch was added to choose either MoreThuente
|
661
|
-
method (::LBFGS_LINESEARCH_MORETHUENTE) or backtracking algorithm
|
662
|
-
(::LBFGS_LINESEARCH_BACKTRACKING).
|
663
|
-
- Fixed a bug: the previous version did not compute psuedo-gradients
|
664
|
-
properly in the line search routines for OWL-QN. This bug might quit
|
665
|
-
an iteration process too early when the OWL-QN routine was activated
|
666
|
-
(0 < ::lbfgs_parameter_t::orthantwise_c).
|
667
|
-
- Configure script for POSIX environments.
|
668
|
-
- SSE/SSE2 optimizations with GCC.
|
669
|
-
- New functions ::lbfgs_malloc and ::lbfgs_free to use SSE/SSE2 routines
|
670
|
-
transparently. It is uncessary to use these functions for libLBFGS built
|
671
|
-
without SSE/SSE2 routines; you can still use any memory allocators if
|
672
|
-
SSE/SSE2 routines are disabled in libLBFGS.
|
673
|
-
- Version 1.3 (2007-12-16):
|
674
|
-
- An API change. An argument was added to lbfgs() function to receive the
|
675
|
-
final value of the objective function. This argument can be set to
|
676
|
-
\c NULL if the final value is unnecessary.
|
677
|
-
- Fixed a null-pointer bug in the sample code (reported by Takashi Imamichi).
|
678
|
-
- Added build scripts for Microsoft Visual Studio 2005 and GCC.
|
679
|
-
- Added README file.
|
680
|
-
- Version 1.2 (2007-12-13):
|
681
|
-
- Fixed a serious bug in orthant-wise L-BFGS.
|
682
|
-
An important variable was used without initialization.
|
683
|
-
- Version 1.1 (2007-12-01):
|
684
|
-
- Implemented orthant-wise L-BFGS.
|
685
|
-
- Implemented lbfgs_parameter_init() function.
|
686
|
-
- Fixed several bugs.
|
687
|
-
- API documentation.
|
688
|
-
- Version 1.0 (2007-09-20):
|
689
|
-
- Initial release.
|
690
|
-
|
691
|
-
@section api Documentation
|
692
|
-
|
693
|
-
- @ref liblbfgs_api "libLBFGS API"
|
694
|
-
|
695
|
-
@section sample Sample code
|
696
|
-
|
697
|
-
@include sample.c
|
698
|
-
|
699
|
-
@section ack Acknowledgements
|
700
|
-
|
701
|
-
The L-BFGS algorithm is described in:
|
702
|
-
- Jorge Nocedal.
|
703
|
-
Updating Quasi-Newton Matrices with Limited Storage.
|
704
|
-
<i>Mathematics of Computation</i>, Vol. 35, No. 151, pp. 773--782, 1980.
|
705
|
-
- Dong C. Liu and Jorge Nocedal.
|
706
|
-
On the limited memory BFGS method for large scale optimization.
|
707
|
-
<i>Mathematical Programming</i> B, Vol. 45, No. 3, pp. 503-528, 1989.
|
708
|
-
|
709
|
-
The line search algorithms used in this implementation are described in:
|
710
|
-
- John E. Dennis and Robert B. Schnabel.
|
711
|
-
<i>Numerical Methods for Unconstrained Optimization and Nonlinear
|
712
|
-
Equations</i>, Englewood Cliffs, 1983.
|
713
|
-
- Jorge J. More and David J. Thuente.
|
714
|
-
Line search algorithm with guaranteed sufficient decrease.
|
715
|
-
<i>ACM Transactions on Mathematical Software (TOMS)</i>, Vol. 20, No. 3,
|
716
|
-
pp. 286-307, 1994.
|
717
|
-
|
718
|
-
This library also implements Orthant-Wise Limited-memory Quasi-Newton (OWL-QN)
|
719
|
-
method presented in:
|
720
|
-
- Galen Andrew and Jianfeng Gao.
|
721
|
-
Scalable training of L1-regularized log-linear models.
|
722
|
-
In <i>Proceedings of the 24th International Conference on Machine
|
723
|
-
Learning (ICML 2007)</i>, pp. 33-40, 2007.
|
724
|
-
|
725
|
-
Special thanks go to:
|
726
|
-
- Yoshimasa Tsuruoka and Daisuke Okanohara for technical information about
|
727
|
-
OWL-QN
|
728
|
-
- Takashi Imamichi for the useful enhancements of the backtracking method
|
729
|
-
- Kevin S. Van Horn, Nic Schraudolph, and Tamas Nepusz for bug fixes
|
730
|
-
|
731
|
-
Finally I would like to thank the original author, Jorge Nocedal, who has been
|
732
|
-
distributing the effieicnt and explanatory implementation in an open source
|
733
|
-
licence.
|
734
|
-
|
735
|
-
@section reference Reference
|
736
|
-
|
737
|
-
- <a href="http://www.ece.northwestern.edu/~nocedal/lbfgs.html">L-BFGS</a> by Jorge Nocedal.
|
738
|
-
- <a href="http://research.microsoft.com/en-us/downloads/b1eb1016-1738-4bd5-83a9-370c9d498a03/default.aspx">Orthant-Wise Limited-memory Quasi-Newton Optimizer for L1-regularized Objectives</a> by Galen Andrew.
|
739
|
-
- <a href="http://chasen.org/~taku/software/misc/lbfgs/">C port (via f2c)</a> by Taku Kudo.
|
740
|
-
- <a href="http://www.alglib.net/optimization/lbfgs.php">C#/C++/Delphi/VisualBasic6 port</a> in ALGLIB.
|
741
|
-
- <a href="http://cctbx.sourceforge.net/">Computational Crystallography Toolbox</a> includes
|
742
|
-
<a href="http://cctbx.sourceforge.net/current_cvs/c_plus_plus/namespacescitbx_1_1lbfgs.html">scitbx::lbfgs</a>.
|
743
|
-
*/
|
744
|
-
|
745
|
-
#endif/*__LBFGS_H__*/
|