1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
// Copyright (C) 2007 Davis E. King (davis@dlib.net)
// License: Boost Software License See LICENSE.txt for the full license.
#ifndef DLIB_SVm_C_TRAINER_Hh_
#define DLIB_SVm_C_TRAINER_Hh_
//#include "local/make_label_kernel_matrix.h"
#include "svm_c_trainer_abstract.h"
#include <cmath>
#include <limits>
#include <sstream>
#include "../matrix.h"
#include "../algs.h"
#include "function.h"
#include "kernel.h"
#include "../optimization/optimization_solve_qp3_using_smo.h"
namespace dlib
{
// ----------------------------------------------------------------------------------------
template <
typename K
>
class svm_c_trainer
{
public:
typedef K kernel_type;
typedef typename kernel_type::scalar_type scalar_type;
typedef typename kernel_type::sample_type sample_type;
typedef typename kernel_type::mem_manager_type mem_manager_type;
typedef decision_function<kernel_type> trained_function_type;
svm_c_trainer (
) :
Cpos(1),
Cneg(1),
cache_size(200),
eps(0.001)
{
}
svm_c_trainer (
const kernel_type& kernel_,
const scalar_type& C_
) :
kernel_function(kernel_),
Cpos(C_),
Cneg(C_),
cache_size(200),
eps(0.001)
{
// make sure requires clause is not broken
DLIB_ASSERT(0 < C_,
"\tsvm_c_trainer::svm_c_trainer(kernel,C)"
<< "\n\t invalid inputs were given to this function"
<< "\n\t C_: " << C_
);
}
void set_cache_size (
long cache_size_
)
{
// make sure requires clause is not broken
DLIB_ASSERT(cache_size_ > 0,
"\tvoid svm_c_trainer::set_cache_size(cache_size_)"
<< "\n\t invalid inputs were given to this function"
<< "\n\t cache_size: " << cache_size_
);
cache_size = cache_size_;
}
long get_cache_size (
) const
{
return cache_size;
}
void set_epsilon (
scalar_type eps_
)
{
// make sure requires clause is not broken
DLIB_ASSERT(eps_ > 0,
"\tvoid svm_c_trainer::set_epsilon(eps_)"
<< "\n\t invalid inputs were given to this function"
<< "\n\t eps_: " << eps_
);
eps = eps_;
}
const scalar_type get_epsilon (
) const
{
return eps;
}
void set_kernel (
const kernel_type& k
)
{
kernel_function = k;
}
const kernel_type& get_kernel (
) const
{
return kernel_function;
}
void set_c (
scalar_type C
)
{
// make sure requires clause is not broken
DLIB_ASSERT(C > 0,
"\t void svm_c_trainer::set_c()"
<< "\n\t C must be greater than 0"
<< "\n\t C: " << C
<< "\n\t this: " << this
);
Cpos = C;
Cneg = C;
}
const scalar_type get_c_class1 (
) const
{
return Cpos;
}
const scalar_type get_c_class2 (
) const
{
return Cneg;
}
void set_c_class1 (
scalar_type C
)
{
// make sure requires clause is not broken
DLIB_ASSERT(C > 0,
"\t void svm_c_trainer::set_c_class1()"
<< "\n\t C must be greater than 0"
<< "\n\t C: " << C
<< "\n\t this: " << this
);
Cpos = C;
}
void set_c_class2 (
scalar_type C
)
{
// make sure requires clause is not broken
DLIB_ASSERT(C > 0,
"\t void svm_c_trainer::set_c_class2()"
<< "\n\t C must be greater than 0"
<< "\n\t C: " << C
<< "\n\t this: " << this
);
Cneg = C;
}
template <
typename in_sample_vector_type,
typename in_scalar_vector_type
>
const decision_function<kernel_type> train (
const in_sample_vector_type& x,
const in_scalar_vector_type& y
) const
{
return do_train(mat(x), mat(y));
}
void swap (
svm_c_trainer& item
)
{
exchange(kernel_function, item.kernel_function);
exchange(Cpos, item.Cpos);
exchange(Cneg, item.Cneg);
exchange(cache_size, item.cache_size);
exchange(eps, item.eps);
}
private:
// ------------------------------------------------------------------------------------
template <
typename in_sample_vector_type,
typename in_scalar_vector_type
>
const decision_function<kernel_type> do_train (
const in_sample_vector_type& x,
const in_scalar_vector_type& y
) const
{
typedef typename K::scalar_type scalar_type;
typedef typename decision_function<K>::sample_vector_type sample_vector_type;
typedef typename decision_function<K>::scalar_vector_type scalar_vector_type;
// make sure requires clause is not broken
DLIB_ASSERT(is_binary_classification_problem(x,y) == true,
"\tdecision_function svm_c_trainer::train(x,y)"
<< "\n\t invalid inputs were given to this function"
<< "\n\t x.nr(): " << x.nr()
<< "\n\t y.nr(): " << y.nr()
<< "\n\t x.nc(): " << x.nc()
<< "\n\t y.nc(): " << y.nc()
<< "\n\t is_binary_classification_problem(x,y): " << is_binary_classification_problem(x,y)
);
scalar_vector_type alpha;
solve_qp3_using_smo<scalar_vector_type> solver;
solver(symmetric_matrix_cache<float>((diagm(y)*kernel_matrix(kernel_function,x)*diagm(y)), cache_size),
//solver(symmetric_matrix_cache<float>(make_label_kernel_matrix(kernel_matrix(kernel_function,x),y), cache_size),
uniform_matrix<scalar_type>(y.size(),1,-1),
y,
0,
Cpos,
Cneg,
alpha,
eps);
scalar_type b;
calculate_b(y,alpha,solver.get_gradient(),Cpos,Cneg,b);
alpha = pointwise_multiply(alpha,y);
// count the number of support vectors
const long sv_count = (long)sum(alpha != 0);
scalar_vector_type sv_alpha;
sample_vector_type support_vectors;
// size these column vectors so that they have an entry for each support vector
sv_alpha.set_size(sv_count);
support_vectors.set_size(sv_count);
// load the support vectors and their alpha values into these new column matrices
long idx = 0;
for (long i = 0; i < alpha.nr(); ++i)
{
if (alpha(i) != 0)
{
sv_alpha(idx) = alpha(i);
support_vectors(idx) = x(i);
++idx;
}
}
// now return the decision function
return decision_function<K> (sv_alpha, b, kernel_function, support_vectors);
}
// ------------------------------------------------------------------------------------
template <
typename scalar_vector_type,
typename scalar_vector_type2
>
void calculate_b(
const scalar_vector_type2& y,
const scalar_vector_type& alpha,
const scalar_vector_type& df,
const scalar_type& Cpos,
const scalar_type& Cneg,
scalar_type& b
) const
{
using namespace std;
long num_free = 0;
scalar_type sum_free = 0;
scalar_type upper_bound = -numeric_limits<scalar_type>::infinity();
scalar_type lower_bound = numeric_limits<scalar_type>::infinity();
for(long i = 0; i < alpha.nr(); ++i)
{
if(y(i) == 1)
{
if(alpha(i) == Cpos)
{
if (df(i) > upper_bound)
upper_bound = df(i);
}
else if(alpha(i) == 0)
{
if (df(i) < lower_bound)
lower_bound = df(i);
}
else
{
++num_free;
sum_free += df(i);
}
}
else
{
if(alpha(i) == Cneg)
{
if (-df(i) < lower_bound)
lower_bound = -df(i);
}
else if(alpha(i) == 0)
{
if (-df(i) > upper_bound)
upper_bound = -df(i);
}
else
{
++num_free;
sum_free -= df(i);
}
}
}
if(num_free > 0)
b = sum_free/num_free;
else
b = (upper_bound+lower_bound)/2;
}
// ------------------------------------------------------------------------------------
kernel_type kernel_function;
scalar_type Cpos;
scalar_type Cneg;
long cache_size;
scalar_type eps;
}; // end of class svm_c_trainer
// ----------------------------------------------------------------------------------------
template <typename K>
void swap (
svm_c_trainer<K>& a,
svm_c_trainer<K>& b
) { a.swap(b); }
// ----------------------------------------------------------------------------------------
}
#endif // DLIB_SVm_C_TRAINER_Hh_