File: pegasos.h

package info (click to toggle)
mldemos 0.5.1-3
  • links: PTS, VCS
  • area: main
  • in suites: jessie, jessie-kfreebsd
  • size: 32,224 kB
  • ctags: 46,525
  • sloc: cpp: 306,887; ansic: 167,718; ml: 126; sh: 109; makefile: 2
file content (714 lines) | stat: -rw-r--r-- 23,628 bytes parent folder | download | duplicates (4)
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
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
// Copyright (C) 2009  Davis E. King (davis@dlib.net)
// License: Boost Software License   See LICENSE.txt for the full license.
#ifndef DLIB_PEGASoS_
#define DLIB_PEGASoS_

#include "pegasos_abstract.h"
#include <cmath>
#include "../algs.h"
#include "function.h"
#include "kernel.h"
#include "kcentroid.h"
#include <iostream>
#include "../smart_pointers.h"

namespace dlib
{

// ----------------------------------------------------------------------------------------

    template <
        typename K 
        >
    class svm_pegasos
    {
        typedef kcentroid<offset_kernel<K> > kc_type;

    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;

        template <typename K_>
        struct rebind {
            typedef svm_pegasos<K_> other;
        };

        svm_pegasos (
        ) :
            max_sv(40),
            lambda_c1(0.0001),
            lambda_c2(0.0001),
            tau(0.01),
            tolerance(0.01),
            train_count(0),
            w(offset_kernel<kernel_type>(kernel,tau),tolerance, max_sv, false)
        {
        }

        svm_pegasos (
            const kernel_type& kernel_, 
            const scalar_type& lambda_,
            const scalar_type& tolerance_,
            unsigned long max_num_sv
        ) :
            max_sv(max_num_sv),
            kernel(kernel_),
            lambda_c1(lambda_),
            lambda_c2(lambda_),
            tau(0.01),
            tolerance(tolerance_),
            train_count(0),
            w(offset_kernel<kernel_type>(kernel,tau),tolerance, max_sv, false)
        {
            // make sure requires clause is not broken
            DLIB_ASSERT(lambda_ > 0 && tolerance > 0 && max_num_sv > 0,
                        "\tsvm_pegasos::svm_pegasos(kernel,lambda,tolerance)"
                        << "\n\t invalid inputs were given to this function"
                        << "\n\t lambda_: " << lambda_ 
                        << "\n\t max_num_sv: " << max_num_sv 
            );
        }

        void clear (
        )
        {
            // reset the w vector back to its initial state
            w = kc_type(offset_kernel<kernel_type>(kernel,tau),tolerance, max_sv, false);
            train_count = 0;
        }

        void set_kernel (
            kernel_type k
        )
        {
            kernel = k;
            clear();
        }

        void set_max_num_sv (
            unsigned long max_num_sv
        )
        {
            // make sure requires clause is not broken
            DLIB_ASSERT(max_num_sv > 0,
                        "\tvoid svm_pegasos::set_max_num_sv(max_num_sv)"
                        << "\n\t invalid inputs were given to this function"
                        << "\n\t max_num_sv: " << max_num_sv 
            );
            max_sv = max_num_sv; 
            clear();
        }

        unsigned long get_max_num_sv (
        ) const
        {
            return max_sv;
        }

        void set_tolerance (
            double tol
        )
        {
            // make sure requires clause is not broken
            DLIB_ASSERT(0 < tol,
                        "\tvoid svm_pegasos::set_tolerance(tol)"
                        << "\n\t invalid inputs were given to this function"
                        << "\n\t tol: " << tol 
            );
            tolerance = tol;
            clear();
        }

        void set_lambda (
            scalar_type lambda_
        )
        {
            // make sure requires clause is not broken
            DLIB_ASSERT(0 < lambda_,
                        "\tvoid svm_pegasos::set_lambda(lambda_)"
                        << "\n\t invalid inputs were given to this function"
                        << "\n\t lambda_: " << lambda_ 
            );
            lambda_c1 = lambda_;
            lambda_c2 = lambda_;

            max_wnorm = 1/std::sqrt(std::min(lambda_c1, lambda_c2));
            clear();
        }

        void set_lambda_class1 (
            scalar_type lambda_
        )
        {
            // make sure requires clause is not broken
            DLIB_ASSERT(0 < lambda_,
                        "\tvoid svm_pegasos::set_lambda_class1(lambda_)"
                        << "\n\t invalid inputs were given to this function"
                        << "\n\t lambda_: " << lambda_ 
            );
            lambda_c1 = lambda_;
            max_wnorm = 1/std::sqrt(std::min(lambda_c1, lambda_c2));
            clear();
        }

        void set_lambda_class2 (
            scalar_type lambda_
        )
        {
            // make sure requires clause is not broken
            DLIB_ASSERT(0 < lambda_,
                        "\tvoid svm_pegasos::set_lambda_class2(lambda_)"
                        << "\n\t invalid inputs were given to this function"
                        << "\n\t lambda_: " << lambda_ 
            );
            lambda_c2 = lambda_;
            max_wnorm = 1/std::sqrt(std::min(lambda_c1, lambda_c2));
            clear();
        }

        const scalar_type get_lambda_class1 (
        ) const
        {
            return lambda_c1;
        }

        const scalar_type get_lambda_class2 (
        ) const
        {
            return lambda_c2;
        }

        const scalar_type get_tolerance (
        ) const
        {
            return tolerance;
        }

        const kernel_type get_kernel (
        ) const
        {
            return kernel;
        }

        unsigned long get_train_count (
        ) const
        {
            return static_cast<unsigned long>(train_count);
        }

        scalar_type train (
            const sample_type& x,
            const scalar_type& y
        ) 
        {
            // make sure requires clause is not broken
            DLIB_ASSERT(y == -1 || y == 1,
                        "\tscalar_type svm_pegasos::train(x,y)"
                        << "\n\t invalid inputs were given to this function"
                        << "\n\t y: " << y
            );

            const double lambda = (y==+1)? lambda_c1 : lambda_c2;

            ++train_count;
            const scalar_type learning_rate = 1/(lambda*train_count);

            // if this sample point is within the margin of the current hyperplane
            if (y*w.inner_product(x) < 1)
            {

                // compute: w = (1-learning_rate*lambda)*w + y*learning_rate*x
                w.train(x,  1 - learning_rate*lambda,  y*learning_rate);

                scalar_type wnorm = std::sqrt(w.squared_norm());
                scalar_type temp = max_wnorm/wnorm;
                if (temp < 1)
                    w.scale_by(temp);
            }
            else
            {
                w.scale_by(1 - learning_rate*lambda);
            }

            // return the current learning rate
            return 1/(std::min(lambda_c1,lambda_c2)*train_count);
        }

        scalar_type operator() (
            const sample_type& x
        ) const
        {
            return w.inner_product(x);
        }

        const decision_function<kernel_type> get_decision_function (
        ) const
        {
            distance_function<offset_kernel<kernel_type> > df = w.get_distance_function();
            return decision_function<kernel_type>(df.get_alpha(), -tau*sum(df.get_alpha()), kernel, df.get_basis_vectors());
        }

        void swap (
            svm_pegasos& item
        )
        {
            exchange(max_sv,         item.max_sv);
            exchange(kernel,         item.kernel);
            exchange(lambda_c1,      item.lambda_c1);
            exchange(lambda_c2,      item.lambda_c2);
            exchange(max_wnorm,      item.max_wnorm);
            exchange(tau,            item.tau);
            exchange(tolerance,      item.tolerance);
            exchange(train_count,    item.train_count);
            exchange(w,              item.w);
        }

        friend void serialize(const svm_pegasos& item, std::ostream& out)
        {
            serialize(item.max_sv, out);
            serialize(item.kernel, out);
            serialize(item.lambda_c1, out);
            serialize(item.lambda_c2, out);
            serialize(item.max_wnorm, out);
            serialize(item.tau, out);
            serialize(item.tolerance, out);
            serialize(item.train_count, out);
            serialize(item.w, out);
        }

        friend void deserialize(svm_pegasos& item, std::istream& in)
        {
            deserialize(item.max_sv, in);
            deserialize(item.kernel, in);
            deserialize(item.lambda_c1, in);
            deserialize(item.lambda_c2, in);
            deserialize(item.max_wnorm, in);
            deserialize(item.tau, in);
            deserialize(item.tolerance, in);
            deserialize(item.train_count, in);
            deserialize(item.w, in);
        }

    private:

        unsigned long max_sv;
        kernel_type kernel;
        scalar_type lambda_c1;
        scalar_type lambda_c2;
        scalar_type max_wnorm;
        scalar_type tau;
        scalar_type tolerance;
        scalar_type train_count;
        kc_type w;

    }; // end of class svm_pegasos

    template <
        typename K 
        >
    void swap (
        svm_pegasos<K>& a,
        svm_pegasos<K>& b
    ) { a.swap(b); }

// ----------------------------------------------------------------------------------------

    template <
        typename T,
        typename U
        >
    void replicate_settings (
        const svm_pegasos<T>& source,
        svm_pegasos<U>& dest
    )
    {
        dest.set_tolerance(source.get_tolerance());
        dest.set_lambda_class1(source.get_lambda_class1());
        dest.set_lambda_class2(source.get_lambda_class2());
        dest.set_max_num_sv(source.get_max_num_sv());
    }

// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------
// ----------------------------------------------------------------------------------------

    template <
        typename trainer_type
        >
    class batch_trainer 
    {

    // ------------------------------------------------------------------------------------

        template <
            typename K,
            typename sample_vector_type
            >
        class caching_kernel 
        {
        public:
            typedef typename K::scalar_type scalar_type;
            typedef long sample_type;
            //typedef typename K::sample_type sample_type;
            typedef typename K::mem_manager_type mem_manager_type;

            caching_kernel () : samples(0), counter(0), counter_threshold(0) {}

            caching_kernel (
                const K& kern,
                const sample_vector_type& samps,
                long cache_size_
            ) : real_kernel(kern), samples(&samps), counter(0)  
            {
                cache_size = std::min<long>(cache_size_, samps.size());

                cache.reset(new cache_type);
                cache->frequency_of_use.resize(samps.size());
                for (long i = 0; i < samps.size(); ++i)
                    cache->frequency_of_use[i] = std::make_pair(0, i);

                // Set the cache build/rebuild threshold so that we have to have
                // as many cache misses as there are entries in the cache before
                // we build/rebuild.
                counter_threshold = samps.size()*cache_size;
                cache->sample_location.assign(samples->size(), -1);
            }

            scalar_type operator() (
                const sample_type& a,
                const sample_type& b
            )  const
            { 
                // rebuild the cache every so often
                if (counter > counter_threshold )
                {
                    build_cache();
                }

                const long a_loc = cache->sample_location[a];
                const long b_loc = cache->sample_location[b];

                cache->frequency_of_use[a].first += 1;
                cache->frequency_of_use[b].first += 1;

                if (a_loc != -1)
                {
                    return cache->kernel(a_loc, b);
                }
                else if (b_loc != -1)
                {
                    return cache->kernel(b_loc, a);
                }
                else
                {
                    ++counter;
                    return real_kernel((*samples)(a), (*samples)(b));
                }
            }

            bool operator== (
                const caching_kernel& item
            ) const
            {
                return item.real_kernel == real_kernel &&
                    item.samples == samples;
            }

        private:
            K real_kernel;

            void build_cache (
            ) const
            {
                std::sort(cache->frequency_of_use.rbegin(), cache->frequency_of_use.rend());
                counter = 0;


                cache->kernel.set_size(cache_size, samples->size());
                cache->sample_location.assign(samples->size(), -1);

                // loop over all the samples in the cache
                for (long i = 0; i < cache_size; ++i)
                {
                    const long cur = cache->frequency_of_use[i].second;
                    cache->sample_location[cur] = i;

                    // now populate all possible kernel products with the current sample
                    for (long j = 0; j < samples->size(); ++j)
                    {
                        cache->kernel(i, j) = real_kernel((*samples)(cur), (*samples)(j));
                    }

                }

                // reset the frequency of use metrics
                for (long i = 0; i < samples->size(); ++i)
                    cache->frequency_of_use[i] = std::make_pair(0, i);
            }


            struct cache_type
            {
                matrix<scalar_type> kernel;  

                std::vector<long> sample_location; // where in the cache a sample is.  -1 means not in cache
                std::vector<std::pair<long,long> > frequency_of_use;  
            };

            const sample_vector_type* samples;

            shared_ptr<cache_type> cache;
            mutable unsigned long counter;
            unsigned long counter_threshold;
            long cache_size;
        };

    // ------------------------------------------------------------------------------------

    public:
        typedef typename trainer_type::kernel_type kernel_type;
        typedef typename trainer_type::scalar_type scalar_type;
        typedef typename trainer_type::sample_type sample_type;
        typedef typename trainer_type::mem_manager_type mem_manager_type;
        typedef typename trainer_type::trained_function_type trained_function_type;


        batch_trainer (
        ) :
            min_learning_rate(0.1),
            use_cache(false),
            cache_size(100)
        {
        }

        batch_trainer (
            const trainer_type& trainer_, 
            const scalar_type min_learning_rate_,
            bool verbose_,
            bool use_cache_,
            long cache_size_ = 100
        ) :
            trainer(trainer_),
            min_learning_rate(min_learning_rate_),
            verbose(verbose_),
            use_cache(use_cache_),
            cache_size(cache_size_)
        {
            // make sure requires clause is not broken
            DLIB_ASSERT(0 < min_learning_rate_ &&
                        cache_size_ > 0,
                        "\tbatch_trainer::batch_trainer()"
                        << "\n\t invalid inputs were given to this function"
                        << "\n\t min_learning_rate_: " << min_learning_rate_ 
                        << "\n\t cache_size_: " << cache_size_ 
            );
            
            trainer.clear();
        }

        const scalar_type get_min_learning_rate (
        ) const 
        {
            return min_learning_rate;
        }

        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
        {
            if (use_cache)
                return do_train_cached(vector_to_matrix(x), vector_to_matrix(y));
            else
                return do_train(vector_to_matrix(x), vector_to_matrix(y));
        }

    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 decision_function<kernel_type>::sample_vector_type sample_vector_type;
            typedef typename decision_function<kernel_type>::scalar_vector_type scalar_vector_type;

            dlib::rand rnd;

            trainer_type my_trainer(trainer);

            scalar_type cur_learning_rate = min_learning_rate + 10;
            unsigned long count = 0;

            while (cur_learning_rate > min_learning_rate)
            {
                const long i = rnd.get_random_32bit_number()%x.size();
                // keep feeding the trainer data until its learning rate goes below our threshold
                cur_learning_rate = my_trainer.train(x(i), y(i));

                if (verbose)
                {
                    if ( (count&0x7FF) == 0)
                    {
                        std::cout << "\rbatch_trainer(): Percent complete: " 
                                  << 100*min_learning_rate/cur_learning_rate << "             " << std::flush;
                    }
                    ++count;
                }
            }

            if (verbose)
            {
                decision_function<kernel_type> df = my_trainer.get_decision_function();
                std::cout << "\rbatch_trainer(): Percent complete: 100           " << std::endl;
                std::cout << "    Num sv: " << df.basis_vectors.size() << std::endl;
                std::cout << "    bias:   " << df.b << std::endl;
                return df;
            }
            else
            {
                return my_trainer.get_decision_function();
            }
        }

        template <
            typename in_sample_vector_type,
            typename in_scalar_vector_type
            >
        const decision_function<kernel_type> do_train_cached (
            const in_sample_vector_type& x,
            const in_scalar_vector_type& y
        ) const
        {
            typedef typename decision_function<kernel_type>::sample_vector_type sample_vector_type;
            typedef typename decision_function<kernel_type>::scalar_vector_type scalar_vector_type;

            dlib::rand rnd;

            // make a caching kernel
            typedef caching_kernel<kernel_type, in_sample_vector_type> ckernel_type;
            ckernel_type ck(trainer.get_kernel(), x, cache_size);

            // now rebind the trainer to use the caching kernel
            typedef typename trainer_type::template rebind<ckernel_type>::other rebound_trainer_type;
            rebound_trainer_type my_trainer;
            my_trainer.set_kernel(ck);
            replicate_settings(trainer, my_trainer);

            scalar_type cur_learning_rate = min_learning_rate + 10;
            unsigned long count = 0;

            while (cur_learning_rate > min_learning_rate)
            {
                const long i = rnd.get_random_32bit_number()%x.size();
                // keep feeding the trainer data until its learning rate goes below our threshold
                cur_learning_rate = my_trainer.train(i, y(i));

                if (verbose)
                {
                    if ( (count&0x7FF) == 0)
                    {
                        std::cout << "\rbatch_trainer(): Percent complete: " 
                                  << 100*min_learning_rate/cur_learning_rate << "             " << std::flush;
                    }
                    ++count;
                }
            }

            if (verbose)
            {
                decision_function<ckernel_type> cached_df;
                cached_df = my_trainer.get_decision_function();

                std::cout << "\rbatch_trainer(): Percent complete: 100           " << std::endl;
                std::cout << "    Num sv: " << cached_df.basis_vectors.size() << std::endl;
                std::cout << "    bias:   " << cached_df.b << std::endl;

                return decision_function<kernel_type> (
                        cached_df.alpha,
                        cached_df.b,
                        trainer.get_kernel(),
                        rowm(x, cached_df.basis_vectors)
                        );
            }
            else
            {
                decision_function<ckernel_type> cached_df;
                cached_df = my_trainer.get_decision_function();

                return decision_function<kernel_type> (
                        cached_df.alpha,
                        cached_df.b,
                        trainer.get_kernel(),
                        rowm(x, cached_df.basis_vectors)
                        );
            }
        }

        trainer_type trainer;
        scalar_type min_learning_rate;
        bool verbose;
        bool use_cache;
        long cache_size;

    }; // end of class batch_trainer

// ----------------------------------------------------------------------------------------

    template <
        typename trainer_type
        >
    const batch_trainer<trainer_type> batch (
        const trainer_type& trainer,
        const typename trainer_type::scalar_type min_learning_rate = 0.1
    ) { return batch_trainer<trainer_type>(trainer, min_learning_rate, false, false); }

// ----------------------------------------------------------------------------------------

    template <
        typename trainer_type
        >
    const batch_trainer<trainer_type> verbose_batch (
        const trainer_type& trainer,
        const typename trainer_type::scalar_type min_learning_rate = 0.1
    ) { return batch_trainer<trainer_type>(trainer, min_learning_rate, true, false); }

// ----------------------------------------------------------------------------------------

    template <
        typename trainer_type
        >
    const batch_trainer<trainer_type> batch_cached (
        const trainer_type& trainer,
        const typename trainer_type::scalar_type min_learning_rate = 0.1,
        long cache_size = 100
    ) { return batch_trainer<trainer_type>(trainer, min_learning_rate, false, true, cache_size); }

// ----------------------------------------------------------------------------------------

    template <
        typename trainer_type
        >
    const batch_trainer<trainer_type> verbose_batch_cached (
        const trainer_type& trainer,
        const typename trainer_type::scalar_type min_learning_rate = 0.1,
        long cache_size = 100
    ) { return batch_trainer<trainer_type>(trainer, min_learning_rate, true, true, cache_size); }

// ----------------------------------------------------------------------------------------

}

#endif // DLIB_PEGASoS_