core.h 95.8 KB
Newer Older
1
2
3
4
5
6
7
8
9
// Copyright (C) 2015  Davis E. King (davis@dlib.net)
// License: Boost Software License   See LICENSE.txt for the full license.
#ifndef DLIB_DNn_CORE_H_
#define DLIB_DNn_CORE_H_

#include "core_abstract.h"
#include "tensor.h"
#include <iterator>
#include <memory>
10
#include <sstream>
11
#include <type_traits>
Davis King's avatar
Davis King committed
12
13
#include "../statistics.h"
#include "../rand.h"
14
#include "../algs.h"
15
#include <utility>
16
#include <tuple>
Davis King's avatar
Davis King committed
17
#include <cmath>
18
#include <vector>
19
20
#include "tensor_tools.h"

21
22
23
24
25


namespace dlib
{

Davis King's avatar
Davis King committed
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
// ----------------------------------------------------------------------------------------

    inline double log1pexp(double x)
    {
        using std::exp;
        using namespace std; // Do this instead of using std::log1p because some compilers
                             // error out otherwise (E.g. gcc 4.9 in cygwin)
        if (x <= -37)
            return exp(x);
        else if (-37 < x && x <= 18)
            return log1p(exp(x));
        else if (18 < x && x <= 33.3)
            return x + exp(-x);
        else
            return x;
    }
    
43
44
// ----------------------------------------------------------------------------------------

Davis King's avatar
Davis King committed
45
46
47
48
    // Tell us if T is one of the special layer types (i.e. add_layer, add_tag_layer, or
    // add_skip_layer).
    template <typename T> struct is_nonloss_layer_type : std::false_type {};
    // Tell us if T is an instance of add_loss_layer.
49
50
    template <typename T> struct is_loss_layer_type : std::false_type {};

51
52
53
54
55
56
57
58
59
60
61
62
63
64
    namespace impl
    {
        template <size_t... n>
        struct ct_integers_list {
            template <size_t m>
            struct push_back
            {
                typedef ct_integers_list<n..., m> type;
            };
        };

        template <size_t max>
        struct ct_make_integer_range
        {
65
            // recursively call push_back on ct_integers_list to build a range from 1 to max
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
            // inclusive.
            typedef typename ct_make_integer_range<max-1>::type::template push_back<max>::type type;
        };

        template <>
        struct ct_make_integer_range<0>
        {
            typedef ct_integers_list<> type;
        };

        template <size_t... indices, typename Tuple>
        auto tuple_subset(
            const Tuple& item, 
            ct_integers_list<indices...>
        ) -> decltype(std::make_tuple(std::get<indices>(item)...))
        {
            return std::make_tuple(std::get<indices>(item)...);
        }

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
        template <typename Head, typename... Tail>
        std::tuple<Tail...> basic_tuple_tail(
            const std::tuple<Head, Tail...>& item
        )
        {
            return tuple_subset(item, typename ct_make_integer_range<sizeof...(Tail)>::type());
        }

        template <typename T>
        std::tuple<T> tuple_flatten(const T& t) 
        {
            return std::make_tuple(t);
        }

        template <typename... T>
        auto tuple_flatten(
            const std::tuple<T...>& item
        ) -> decltype(tuple_flatten(item, typename ct_make_integer_range<sizeof...(T)>::type()))
        {
            return tuple_flatten(item, typename ct_make_integer_range<sizeof...(T)>::type());
        }

        template <size_t... indices, typename... T>
        auto tuple_flatten(
            const std::tuple<T...>& item, 
            ct_integers_list<indices...>
        ) -> decltype(std::tuple_cat(tuple_flatten(std::get<indices-1>(item))...))
        {
            return std::tuple_cat(tuple_flatten(std::get<indices-1>(item))...);
        }

        template <typename T>
        struct tuple_head_helper
        {
            typedef T type;
            static const type& get(const T& item) 
            {
                return item;
            }
        };

        template <typename T, typename... U>
        struct tuple_head_helper<std::tuple<T, U...>>
        {
            typedef typename tuple_head_helper<T>::type type;
            static const type& get(const std::tuple<T,U...>& item) 
            {
                return tuple_head_helper<T>::get(std::get<0>(item));
            }
        };

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
        template <typename T> struct alwaysbool { typedef bool type; };

        resizable_tensor& rt();

        // The significance of a layer's backward method requiring forward's outputs is
        // that such as layer can't have an in-place layer stacked on top of it because
        // in-place layers overwrite the output of the layer they sit on top of.
        template <typename layer_type, typename SUBNET>
        constexpr auto backward_requires_forward_output(
            layer_type& layer,
            SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.backward(rt(),rt(),sub,rt()))>::type
        {
            return true;
        }

        template <typename layer_type, typename SUBNET>
        constexpr auto backward_requires_forward_output(
            layer_type& layer,
            SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.backward(rt(),sub,rt()))>::type
        {
            return false;
        }

        template <typename layer_type, typename SUBNET>
        constexpr auto backward_requires_forward_output(
            layer_type& layer,
            SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.backward_inplace(rt(),rt(),sub.get_gradient_input(),rt()))>::type
        {
            return true;
        }

170
171
172
173
174
175
176
177
178
        template <typename layer_type, typename SUBNET>
        constexpr auto backward_requires_forward_output(
            layer_type& layer,
            SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.backward_inplace(rt(),sub.get_gradient_input(),rt()))>::type
        {
            return false;
        }

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
        template <typename layer_type, typename SUBNET>
        constexpr auto has_inplace_backward(
            layer_type& layer,
            SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.backward(rt(),rt(),sub,rt()))>::type
        {
            return false;
        }

        template <typename layer_type, typename SUBNET>
        constexpr auto has_inplace_backward(
            layer_type& layer,
            SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.backward(rt(),sub,rt()))>::type
        {
            return false;
        }

        template <typename layer_type, typename SUBNET>
        constexpr auto has_inplace_backward(
            layer_type& layer,
            SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.backward_inplace(rt(),rt(),sub.get_gradient_input(),rt()))>::type
        {
            return true;
        }

206
207
208
209
210
211
212
213
214
        template <typename layer_type, typename SUBNET>
        constexpr auto has_inplace_backward(
            layer_type& layer,
            SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.backward_inplace(rt(),sub.get_gradient_input(),rt()))>::type
        {
            return true;
        }

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
        template <typename layer_type, typename SUBNET>
        constexpr auto is_inplace_layer(
            layer_type& layer,
            const SUBNET& sub 
        ) -> typename alwaysbool<decltype(layer.forward(sub,rt()))>::type
        {
            return false;
        }

        template <typename layer_type, typename SUBNET>
        constexpr auto is_inplace_layer(
            layer_type& layer,
            const SUBNET& sub
        ) -> typename alwaysbool<decltype(layer.forward_inplace(sub.get_output(),rt()))>::type
        {
            return true;
        }

        template <typename layer_type, typename SUBNET>
        auto call_layer_backward(
            layer_type& layer,
            const tensor& computed_output, 
            const tensor& gradient_input, 
            SUBNET& sub, 
            tensor& params_grad
        ) -> decltype(layer.backward(computed_output,gradient_input,sub,params_grad))
        {
            layer.backward(computed_output,gradient_input,sub,params_grad);
        }

        template <typename layer_type, typename SUBNET>
        auto call_layer_backward(
            layer_type& layer,
            const tensor& , 
            const tensor& gradient_input, 
            SUBNET& sub, 
            tensor& params_grad
        ) -> decltype(layer.backward(gradient_input,sub,params_grad))
        {
            layer.backward(gradient_input,sub,params_grad);
        }

        template <typename layer_type, typename SUBNET>
        auto call_layer_backward(
            layer_type& layer,
            const tensor& computed_output, 
            const tensor& gradient_input, 
            SUBNET& sub, 
            tensor& params_grad
        ) -> decltype(layer.backward_inplace(computed_output,gradient_input,sub.get_gradient_input(),params_grad))
        {
            layer.backward_inplace(computed_output,gradient_input,sub.get_gradient_input(),params_grad);
        }

269
270
271
272
273
274
275
276
277
278
279
280
        template <typename layer_type, typename SUBNET>
        auto call_layer_backward(
            layer_type& layer,
            const tensor& , 
            const tensor& gradient_input, 
            SUBNET& sub, 
            tensor& params_grad
        ) -> decltype(layer.backward_inplace(gradient_input,sub.get_gradient_input(),params_grad))
        {
            layer.backward_inplace(gradient_input,sub.get_gradient_input(),params_grad);
        }

281
282
283
284
285

        template <typename layer_type, typename SUBNET>
        auto call_layer_forward(
            layer_type& layer,
            const SUBNET& sub, 
286
            tensor& /*data_output*/
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
        ) -> decltype(layer.forward(sub,rt()))
        {
            // This overload of call_layer_forward() is here because this template
            // naturally gets instantiated but only on code paths that never get executed.
            // So rather than writing a bunch of hard to read template magic around call
            // sites we just have this overload that doesn't do anything (and an assert to
            // make sure that's the case).
            DLIB_CASSERT(false, "This should never happen");
        }

        template <typename layer_type, typename SUBNET>
        auto call_layer_forward(
            layer_type& layer,
            const SUBNET& sub, 
            resizable_tensor& data_output
        ) -> decltype(layer.forward(sub,data_output))
        {
            layer.forward(sub,data_output);
        }

        template <typename layer_type, typename SUBNET>
        auto call_layer_forward(
            layer_type& layer,
            const SUBNET& sub, 
            tensor& data_output
        ) -> decltype(layer.forward_inplace(sub.get_output(),data_output))
        {
            layer.forward_inplace(sub.get_output(),data_output);
        }

        template <typename layer_type, typename SUBNET>
        auto call_layer_forward(
            layer_type& layer,
            const SUBNET& sub, 
            resizable_tensor& data_output
        ) -> decltype(layer.forward_inplace(sub.get_output(),data_output))
        {
            if (!have_same_dimensions(data_output, sub.get_output()))
                data_output.copy_size(sub.get_output());
            layer.forward_inplace(sub.get_output(),data_output);
        }


    } // end namespace impl
331

332
333
334
335
336
337
338
339
340
341
342
343
    template <typename... T>
    typename impl::tuple_head_helper<std::tuple<T...>>::type tuple_head (
        const std::tuple<T...>& item
    ) 
    {
        return impl::tuple_head_helper<std::tuple<T...>>::get(item);
    }

    template <typename... T>
    auto tuple_tail(
        const std::tuple<T...>& item
    ) -> decltype(impl::basic_tuple_tail(impl::tuple_flatten(item)))
344
    {
345
        return impl::basic_tuple_tail(impl::tuple_flatten(item));
346
347
    }

348
349
350
351
352
353
    inline std::tuple<> tuple_tail(
        const std::tuple<>& item
    ) 
    {
        return item;
    }
354
355
356
357
358
359
360
361
// ----------------------------------------------------------------------------------------

    inline void randomize_parameters (
        tensor& params,
        unsigned long num_inputs_and_outputs,
        dlib::rand& rnd
    )
    {
Davis King's avatar
Davis King committed
362
        for (auto& val : params)
363
364
365
366
        {
            // Draw a random number to initialize the layer according to formula (16)
            // from Understanding the difficulty of training deep feedforward neural
            // networks by Xavier Glorot and Yoshua Bengio.
Davis King's avatar
Davis King committed
367
            val = 2*rnd.get_random_float()-1;
368
369
370
371
372
373
            val *= std::sqrt(6.0/(num_inputs_and_outputs));
        }
    }

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

374
    template <typename T>
375
376
    class sstack
    {
Davis King's avatar
Davis King committed
377
378
    public:
        typedef T value_type;
379

380
        sstack() = delete;
381

382
383
384
385
        sstack (
            T* data_,
            size_t s
        ) : data(data_), mysize(s) {}
386

387
388
389
390
391
392
393
394
395
        const T& top() const 
        { 
            DLIB_CASSERT(size() != 0, "You can't call top() on an empty stack");
            return *data;
        }
        T& top()  
        { 
            DLIB_CASSERT(size() != 0, "You can't call top() on an empty stack");
            return *data;
396
397
        }

398
399
400
401
402
403
        size_t size() const { return mysize; }

        sstack pop(size_t num=1) 
        { 
            DLIB_CASSERT(num < size(), "You can't pop more things from the stack than it has in it.");
            return sstack(data+num, mysize-num);
404
405
        }

Davis King's avatar
Davis King committed
406
    private:
407
408
409

        T* data;
        size_t mysize;
410
411
412
    };

    template <typename T>
413
    sstack<T> make_sstack(std::vector<T>& item)
414
    {
415
416
        return sstack<T>(item.data(), item.size());
    }
417
418
419
420
421
422
423

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

    namespace dimpl
    {
424
        template <typename T, bool is_first = true, typename enabled=void>
Davis King's avatar
Davis King committed
425
        class subnet_wrapper
426
427
428
        {
            /*!
                WHAT THIS OBJECT REPRESENTS
Davis King's avatar
Davis King committed
429
                    This is a tool that makes an add_layer or add_loss_layer object
Davis King's avatar
Davis King committed
430
                    expose only the part of its interface defined by the SUBNET
431
                    type in layers_abstract.h.  This way, when we pass subnetwork
432
                    objects to the layer callbacks those callbacks won't be able to 
433
                    interact with the subnetworks in a way other than specified 
Davis King's avatar
Davis King committed
434
                    by the SUBNET interface spec.
435
436
437
438
439
440
441

                    We also allow the top layer of a subnet_wrapper stack to call the
                    private_get_output() and private_get_gradient_input() functions.  This
                    way, layers that have had their output/gradient overwritten by in-place
                    layers can only be accessed from the in-place layers that sit directly
                    on top of them since those in-place layers are the only layers that
                    know how to interact with them properly.
442
443
444
            !*/

        public:
Davis King's avatar
Davis King committed
445
446
            subnet_wrapper(const subnet_wrapper&) = delete;
            subnet_wrapper& operator=(const subnet_wrapper&) = delete;
447

Davis King's avatar
Davis King committed
448
            subnet_wrapper(T& l_) {}
449
450
451
452
453
            // Nothing here because in this case T is one of the input layer types 
            // that doesn't have anything in it.
        };

        template <typename T>
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
        class subnet_wrapper<T,true, typename std::enable_if<is_nonloss_layer_type<T>::value>::type>
        {

        public:
            subnet_wrapper(const subnet_wrapper&) = delete;
            subnet_wrapper& operator=(const subnet_wrapper&) = delete;

            typedef T wrapped_type;
            const static size_t num_layers = T::num_layers;

            subnet_wrapper(T& l_) : l(l_),subnetwork(l.subnet()) {}

            const tensor& get_output() const { return l.private_get_output(); }
            tensor& get_gradient_input() { return l.private_get_gradient_input(); }

Davis King's avatar
Davis King committed
469
470
            const subnet_wrapper<typename T::subnet_type,false>& subnet() const { return subnetwork; }
            subnet_wrapper<typename T::subnet_type,false>& subnet() { return subnetwork; }
471
472
473
474
475
476
477
478

        private:
            T& l;
            subnet_wrapper<typename T::subnet_type,false> subnetwork;
        };

        template <typename T>
        class subnet_wrapper<T,false, typename std::enable_if<is_nonloss_layer_type<T>::value>::type>
479
480
481
        {

        public:
Davis King's avatar
Davis King committed
482
483
            subnet_wrapper(const subnet_wrapper&) = delete;
            subnet_wrapper& operator=(const subnet_wrapper&) = delete;
484

485
486
487
            typedef T wrapped_type;
            const static size_t num_layers = T::num_layers;

Davis King's avatar
Davis King committed
488
            subnet_wrapper(T& l_) : l(l_),subnetwork(l.subnet()) {}
489
490
491
492

            const tensor& get_output() const { return l.get_output(); }
            tensor& get_gradient_input() { return l.get_gradient_input(); }

Davis King's avatar
Davis King committed
493
494
            const subnet_wrapper<typename T::subnet_type,false>& subnet() const { return subnetwork; }
            subnet_wrapper<typename T::subnet_type,false>& subnet() { return subnetwork; }
495
496
497

        private:
            T& l;
Davis King's avatar
Davis King committed
498
            subnet_wrapper<typename T::subnet_type,false> subnetwork;
499
500
501
        };
    }

Davis King's avatar
Davis King committed
502
503
// ----------------------------------------------------------------------------------------

Davis King's avatar
Davis King committed
504
    template <typename LAYER_DETAILS, typename SUBNET, typename enabled = void>
505
506
507
    class add_layer;

    template <typename T, typename U>
Davis King's avatar
Davis King committed
508
    struct is_nonloss_layer_type<add_layer<T,U>> : std::true_type {};
509

Davis King's avatar
Davis King committed
510
511
512
    template <typename LAYER_DETAILS, typename SUBNET>
    class add_layer<LAYER_DETAILS,SUBNET, 
            typename std::enable_if<is_nonloss_layer_type<SUBNET>::value>::type>
513
514
515
    {
    public:
        typedef LAYER_DETAILS layer_details_type;
Davis King's avatar
Davis King committed
516
517
518
519
        typedef SUBNET subnet_type;
        typedef typename subnet_type::input_type input_type;
        const static size_t num_layers = subnet_type::num_layers + 1;
        const static unsigned int sample_expansion_factor = subnet_type::sample_expansion_factor;
520
521
522

        add_layer(
        ):
523
            subnetwork(new subnet_type()),
524
            this_layer_setup_called(false),
525
526
            gradient_input_is_stale(true),
            get_output_and_gradient_input_disabled(false)
527
        {
528
            if (this_layer_operates_inplace())
529
                subnetwork->disable_output_and_gradient_getters();
530
531
        }

532
533
534
535
536
537
538
539
540
541
542
543
544
        add_layer(const add_layer& item)
        {
            details = item.details;
            subnetwork.reset(new subnet_type(*item.subnetwork));
            this_layer_setup_called = item.this_layer_setup_called;
            gradient_input_is_stale = item.gradient_input_is_stale;
            get_output_and_gradient_input_disabled = item.get_output_and_gradient_input_disabled;
            x_grad = item.x_grad;
            cached_output = item.cached_output; 
            params_grad = item.params_grad; 
            temp_tensor = item.temp_tensor;
        }
        add_layer& operator=(const add_layer& item) { add_layer(item).swap(*this); return *this;}
545
546
        add_layer(add_layer&& item) : add_layer() { swap(item); }
        add_layer& operator=(add_layer&& item) { swap(item); return *this; }
547
548
549

        template <typename T, typename U, typename E>
        friend class add_layer;
550
551
        template <typename T, bool is_first, typename E>
        friend class dimpl::subnet_wrapper;
552
553
554
555
        template <unsigned long T, typename U, typename E>
        friend class add_tag_layer;
        template <template<typename> class T, typename U>
        friend class add_skip_layer;
556
557
        template <size_t N, template<typename> class L, typename S>
        friend class repeat;
558
559
560
561
562
563
564

        // Allow copying networks from one to another as long as their corresponding 
        // layers can be constructed from each other.
        template <typename T, typename U, typename E>
        add_layer(
            const add_layer<T,U,E>& item
        ) :
565
            subnetwork(new subnet_type(item.subnet())),
566
567
568
            details(item.layer_details()), 
            this_layer_setup_called(item.this_layer_setup_called),
            gradient_input_is_stale(item.gradient_input_is_stale),
569
            get_output_and_gradient_input_disabled(item.get_output_and_gradient_input_disabled),
570
571
572
            x_grad(item.x_grad),
            cached_output(item.cached_output)
        {
573
            if (this_layer_operates_inplace())
574
                subnetwork->disable_output_and_gradient_getters();
575
576
577
578
579
580
581
582
        }

        template <typename ...T>
        add_layer(
            const LAYER_DETAILS& layer_det, 
            T&& ...args
        ) : 
            details(layer_det), 
583
            subnetwork(new subnet_type(std::forward<T>(args)...)),
584
            this_layer_setup_called(false),
585
586
            gradient_input_is_stale(true),
            get_output_and_gradient_input_disabled(false)
587
        {
588
            if (this_layer_operates_inplace())
589
                subnetwork->disable_output_and_gradient_getters();
590
591
592
593
594
595
596
597
        }

        template <typename ...T>
        add_layer(
            LAYER_DETAILS&& layer_det, 
            T&& ...args
        ) : 
            details(std::move(layer_det)), 
598
            subnetwork(new subnet_type(std::forward<T>(args)...)),
599
            this_layer_setup_called(false),
600
601
            gradient_input_is_stale(true),
            get_output_and_gradient_input_disabled(false)
602
        {
603
            if (this_layer_operates_inplace())
604
                subnetwork->disable_output_and_gradient_getters();
605
606
        }

607
        template <typename ...T, typename LD, typename ...U>
608
        add_layer(
609
            const std::tuple<LD,U...>& layer_det, 
610
611
            T&& ...args
        ) : 
612
            details(tuple_head(layer_det)), 
613
            subnetwork(new subnet_type(tuple_tail(layer_det),std::forward<T>(args)...)),
614
            this_layer_setup_called(false),
615
616
            gradient_input_is_stale(true),
            get_output_and_gradient_input_disabled(false)
617
        {
618
            if (this_layer_operates_inplace())
619
                subnetwork->disable_output_and_gradient_getters();
620
621
        }

622
        template <typename ...T, typename LD, typename ...U>
623
624
        add_layer(
            std::tuple<>,
625
            const std::tuple<LD,U...>& layer_det, 
626
627
628
629
630
631
632
633
634
635
            T&& ...args
        ) : add_layer(layer_det,args...) { }

        template <typename ...T>
        add_layer(
            std::tuple<>, 
            LAYER_DETAILS&& layer_det, 
            T&& ...args
        ) : add_layer(layer_det, args...) { }

636
637
        template <typename input_iterator>
        void to_tensor (
638
639
            input_iterator ibegin,
            input_iterator iend,
640
641
642
            resizable_tensor& data
        ) const
        {
643
            subnetwork->to_tensor(ibegin,iend,data);
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
        }

        template <typename input_iterator>
        const tensor& operator() (
            input_iterator ibegin,
            input_iterator iend
        )
        {
            to_tensor(ibegin,iend,temp_tensor);
            return forward(temp_tensor);
        }


        const tensor& operator() (const input_type& x)
        {
            return (*this)(&x, &x+1);
        }

        const tensor& forward(const tensor& x)
        {
664
665
            subnetwork->forward(x);
            const dimpl::subnet_wrapper<subnet_type> wsub(*subnetwork);
666
667
668
669
670
            if (!this_layer_setup_called)
            {
                details.setup(wsub);
                this_layer_setup_called = true;
            }
671
672
673
674
675
            if (this_layer_operates_inplace())
                impl::call_layer_forward(details, wsub, private_get_output());
            else
                impl::call_layer_forward(details, wsub, cached_output);

676
            gradient_input_is_stale = true;
677
            return private_get_output();
678
679
        }

680
681
    private:
        tensor& private_get_output() const
682
        { 
683
            if (const_cast<add_layer&>(*this).this_layer_operates_inplace())
684
                return subnetwork->private_get_output();
685
686
687
688
689
690
            else
                return const_cast<resizable_tensor&>(cached_output); 
        }
        tensor& private_get_gradient_input() 
        { 
            if (this_layer_operates_inplace())
691
            {
692
                return subnetwork->private_get_gradient_input();
693
            }
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
            else
            {
                if (gradient_input_is_stale)
                {
                    gradient_input_is_stale = false;
                    x_grad.copy_size(private_get_output());
                    x_grad = 0;
                }
                return x_grad; 
            }
        }
        void disable_output_and_gradient_getters (
        ) { get_output_and_gradient_input_disabled = true; }
    public:
        const tensor& get_output() const 
        { 
            if (get_output_and_gradient_input_disabled)
                throw dlib::error("Accessing this layer's get_output() is disabled because an in-place layer has been stacked on top of it.");
            return private_get_output(); 
        }
        tensor& get_gradient_input() 
        { 
            if (get_output_and_gradient_input_disabled)
                throw dlib::error("Accessing this layer's get_gradient_input() is disabled because an in-place layer has been stacked on top of it.");
            return private_get_gradient_input();
719
720
        }

721
        const tensor& get_final_data_gradient(
722
        ) const { return subnetwork->get_final_data_gradient(); }
723

724
        template <typename solver_type>
725
        void update(const tensor& x, sstack<solver_type> solvers, double step_size)
726
        {
727
            update(x,private_get_gradient_input(),solvers,step_size);
728
729
730
        }

        template <typename solver_type>
731
        void update(const tensor& x, const tensor& gradient_input, sstack<solver_type> solvers, double step_size)
732
        {
733
            DLIB_CASSERT(solvers.size()>=num_layers,"");
734
            dimpl::subnet_wrapper<subnet_type> wsub(*subnetwork);
735
            params_grad.copy_size(details.get_layer_params());
736
            impl::call_layer_backward(details, private_get_output(),
737
                gradient_input, wsub, static_cast<tensor&>(params_grad));
738

739
740
            // Don't try to adjust the parameters if this layer doesn't have any.
            if (params_grad.size() != 0)
741
742
743
744
745
            {
                const tensor& step = solvers.top()(details.get_layer_params(), static_cast<const tensor&>(params_grad));
                tt::add(1,details.get_layer_params(), step_size, step);
            }
            subnetwork->update(x, solvers.pop(), step_size);
746
            gradient_input_is_stale = true;
747
748
        }

749
750
        const subnet_type& subnet() const { return *subnetwork; }
        subnet_type& subnet() { return *subnetwork; }
751
752
753
754
755
756
757
758
759
760
761

        const layer_details_type& layer_details() const { return details; } 
        layer_details_type& layer_details() { return details; } 

        void clean()
        {
            x_grad.clear();
            cached_output.clear();
            params_grad.clear();
            temp_tensor.clear();
            gradient_input_is_stale = true;
762
            subnetwork->clean();
763
764
        }

765
766
767
768
        friend void serialize(const add_layer& item, std::ostream& out)
        {
            int version = 1;
            serialize(version, out);
769
            serialize(*item.subnetwork, out);
770
771
772
            serialize(item.details, out);
            serialize(item.this_layer_setup_called, out);
            serialize(item.gradient_input_is_stale, out);
773
            serialize(item.get_output_and_gradient_input_disabled, out);
774
775
776
777
778
779
780
781
782
783
            serialize(item.x_grad, out);
            serialize(item.cached_output, out);
        }

        friend void deserialize(add_layer& item, std::istream& in)
        {
            int version = 0;
            deserialize(version, in);
            if (version != 1)
                throw serialization_error("Unexpected version found while deserializing dlib::add_layer.");
784
            deserialize(*item.subnetwork, in);
785
786
787
            deserialize(item.details, in);
            deserialize(item.this_layer_setup_called, in);
            deserialize(item.gradient_input_is_stale, in);
788
            deserialize(item.get_output_and_gradient_input_disabled, in);
789
790
791
792
            deserialize(item.x_grad, in);
            deserialize(item.cached_output, in);
        }

793
794
    private:

795
796
797
798
        bool this_layer_operates_inplace(
        ) 
        {
            // This layer can run in-place if it's an in-place capable layer and also if
799
            // the layer it's on top of doesn't need its own output tensor (since in-place
800
            // layers overwrite that tensor)
801
            return impl::is_inplace_layer(details, *subnetwork) && !subnetwork->this_layer_requires_forward_output();
802
803
804
805
        }
        bool this_layer_requires_forward_output(
        ) 
        {
806
            return impl::backward_requires_forward_output(details, *subnetwork);
807
808
        }

809
810
811
812
813
814
        void swap(add_layer& item)
        {
            std::swap(subnetwork,item.subnetwork);
            std::swap(details, item.details);
            std::swap(this_layer_setup_called, item.this_layer_setup_called);
            std::swap(gradient_input_is_stale, item.gradient_input_is_stale);
815
            std::swap(get_output_and_gradient_input_disabled, item.get_output_and_gradient_input_disabled);
816
817
818
819
            std::swap(x_grad, item.x_grad);
            std::swap(cached_output, item.cached_output);
        }

820
821

        LAYER_DETAILS details;
822
        std::unique_ptr<subnet_type> subnetwork;
823
824
        bool this_layer_setup_called;
        bool gradient_input_is_stale;
825
826
827
828
        bool get_output_and_gradient_input_disabled;
        // Note that if this_layer_operates_inplace()==true then x_grad and cached_output
        // are not used at all.  Instead, this layer uses these variables from the lower
        // layer.
829
830
831
832
833
834
835
836
837
838
839
840
841
        resizable_tensor x_grad;
        resizable_tensor cached_output; 

        // The following 2 objects don't logically contribute to the state of this class.
        // They are only here to prevent them from being reallocated over and over in
        // member functions.
        resizable_tensor params_grad; 
        resizable_tensor temp_tensor;

    };

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

842
// This version of add_layer handles the special case where the subnetwork being given is
Davis King's avatar
Davis King committed
843
// just an input layer object.
844
845
846
847
848
    template <typename LAYER_DETAILS, typename INPUT_LAYER, typename enabled>
    class add_layer
    {
    public:
        typedef LAYER_DETAILS layer_details_type;
Davis King's avatar
Davis King committed
849
        typedef INPUT_LAYER subnet_type;
850
851
852
853
854
855
856
857
858
        typedef typename INPUT_LAYER::input_type input_type;
        const static unsigned int sample_expansion_factor = INPUT_LAYER::sample_expansion_factor;
        const static size_t num_layers = 1;
        static_assert(sample_expansion_factor >= 1,
            "The input layer can't produce fewer output tensors than there are inputs.");

        add_layer(
        ): 
            this_layer_setup_called(false),
859
860
            gradient_input_is_stale(true),
            get_output_and_gradient_input_disabled(false)
861
862
863
        {}

        add_layer(const add_layer&) = default;
864
        add_layer(add_layer&& item) : add_layer() { swap(item); }
865
        add_layer& operator=(const add_layer&) = default;
866
        add_layer& operator=(add_layer&& item) { swap(item); return *this; }
867
868
869

        template <typename T, typename U, typename E>
        friend class add_layer;
870
871
        template <typename T, bool is_first, typename E>
        friend class dimpl::subnet_wrapper;
872
873
874
875
        template <unsigned long T, typename U, typename E>
        friend class add_tag_layer;
        template <template<typename> class T, typename U>
        friend class add_skip_layer;
876
877
        template <size_t N, template<typename> class L, typename S>
        friend class repeat;
878
879
880
881
882
883
884

        // Allow copying networks from one to another as long as their corresponding 
        // layers can be constructed from each other.
        template <typename T, typename U, typename E>
        add_layer(
            const add_layer<T,U,E>& item
        ):
Davis King's avatar
Davis King committed
885
            input_layer(item.subnet()),
886
887
888
            details(item.layer_details()),
            this_layer_setup_called(item.this_layer_setup_called),
            gradient_input_is_stale(item.gradient_input_is_stale),
889
            get_output_and_gradient_input_disabled(false),
890
            x_grad(item.x_grad),
891
892
            cached_output(item.cached_output),
            grad_final(item.grad_final)
893
894
895
896
897
898
899
900
        {
        }

        add_layer(
            const LAYER_DETAILS& layer_det
        ) : 
            details(layer_det), 
            this_layer_setup_called(false),
901
902
            gradient_input_is_stale(true),
            get_output_and_gradient_input_disabled(false)
903
904
905
906
907
908
909
        {}

        add_layer(
            LAYER_DETAILS&& layer_det
        ) : 
            details(std::move(layer_det)), 
            this_layer_setup_called(false),
910
911
            gradient_input_is_stale(true),
            get_output_and_gradient_input_disabled(false)
912
913
914
915
916
917
        {}

        add_layer(
            LAYER_DETAILS layer_det, 
            INPUT_LAYER il
        ) : 
918
919
            details(std::move(layer_det)),
            input_layer(std::move(il)),
920
            this_layer_setup_called(false),
921
922
            gradient_input_is_stale(true),
            get_output_and_gradient_input_disabled(false)
923
924
        {}

925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
        add_layer(
            std::tuple<>,
            const LAYER_DETAILS& layer_det
        ) : add_layer(layer_det) {}

        add_layer(
            std::tuple<>,
            LAYER_DETAILS&& layer_det
        ) : add_layer(layer_det) {}

        add_layer(
            std::tuple<>,
            LAYER_DETAILS layer_det, 
            INPUT_LAYER il
        ) : add_layer(layer_det,il) {}

        add_layer(
            const std::tuple<LAYER_DETAILS>& layer_det
943
        ) : add_layer(tuple_head(layer_det)) {}
944
945
946
947

        add_layer(
            const std::tuple<LAYER_DETAILS>& layer_det,
            INPUT_LAYER il
948
        ) : add_layer(tuple_head(layer_det),il) {}
949

950
951
        template <typename input_iterator>
        void to_tensor (
952
953
            input_iterator ibegin,
            input_iterator iend,
954
955
956
            resizable_tensor& data
        ) const
        {
957
            input_layer.to_tensor(ibegin, iend, data);
958
            // make sure the input layer's to_tensor() function is implemented properly.
959
            DLIB_CASSERT(std::distance(ibegin,iend)*sample_expansion_factor == data.num_samples(),"");
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
            data.async_copy_to_device();
        }


        template <typename input_iterator>
        const tensor& operator() (
            input_iterator ibegin,
            input_iterator iend
        )
        {
            to_tensor(ibegin,iend,temp_tensor);
            return forward(temp_tensor);
        }


        const tensor& operator() (const input_type& x)
        {
            return (*this)(&x, &x+1);
        }

        const tensor& forward (const tensor& x)
        {
            DLIB_CASSERT(x.num_samples()%sample_expansion_factor == 0,"");
983
            subnet_wrapper wsub(x, grad_final);
984
985
986
987
988
            if (!this_layer_setup_called)
            {
                details.setup(wsub);
                this_layer_setup_called = true;
            }
989
            impl::call_layer_forward(details, wsub, cached_output);
990
            gradient_input_is_stale = true;
991
            return private_get_output();
992
993
        }

994
995
996
    private:
        tensor& private_get_output() const { return const_cast<resizable_tensor&>(cached_output); }
        tensor& private_get_gradient_input() 
997
998
999
1000
        { 
            if (gradient_input_is_stale)
            {
                gradient_input_is_stale = false;
1001
                x_grad.copy_size(private_get_output());
1002
1003
1004
1005
                x_grad = 0;
            }
            return x_grad; 
        }
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
        void disable_output_and_gradient_getters (
        ) { get_output_and_gradient_input_disabled = true; }
    public:
        const tensor& get_output() const 
        { 
            if (get_output_and_gradient_input_disabled)
                throw dlib::error("Accessing this layer's get_output() is disabled because an in-place layer has been stacked on top of it.");
            return private_get_output(); 
        }
        tensor& get_gradient_input() 
        { 
            if (get_output_and_gradient_input_disabled)
                throw dlib::error("Accessing this layer's get_gradient_input() is disabled because an in-place layer has been stacked on top of it.");
            return private_get_gradient_input();
        }
1021

1022
        const tensor& get_final_data_gradient(
1023
        ) const { return grad_final; }
1024

1025
        template <typename solver_type>
1026
        void update(const tensor& x, sstack<solver_type> solvers, double step_size)
1027
        {
1028
            return update(x,private_get_gradient_input(),solvers, step_size);
1029
1030
1031
        }

        template <typename solver_type>
1032
        void update(const tensor& x, const tensor& gradient_input, sstack<solver_type> solvers, double step_size)
1033
        {
1034
1035
1036
1037
1038
1039
1040
            DLIB_CASSERT(solvers.size()>=num_layers,"");
            // make sure grad_final is initialized to 0
            if (!have_same_dimensions(x, grad_final))
                grad_final.copy_size(x);
            grad_final = 0;  

            subnet_wrapper wsub(x, grad_final);
1041
            params_grad.copy_size(details.get_layer_params());
1042
            impl::call_layer_backward(details, private_get_output(),
1043
                gradient_input, wsub, static_cast<tensor&>(params_grad));
1044

1045
1046
            // Don't try to adjust the parameters if this layer doesn't have any.
            if (params_grad.size() != 0)
1047
1048
1049
1050
            {
                const tensor& step = solvers.top()(details.get_layer_params(), static_cast<const tensor&>(params_grad));
                tt::add(1,details.get_layer_params(), step_size, step);
            }
1051
            gradient_input_is_stale = true;
1052
1053
        }

Davis King's avatar
Davis King committed
1054
1055
        const subnet_type& subnet() const { return input_layer; } 
        subnet_type& subnet() { return input_layer; } 
1056
1057
1058
1059
1060
1061
1062

        const layer_details_type& layer_details() const { return details; } 
        layer_details_type& layer_details() { return details; } 

        void clean()
        {
            x_grad.clear();
1063
            grad_final.clear();
1064
1065
1066
1067
1068
1069
            cached_output.clear();
            params_grad.clear();
            temp_tensor.clear();
            gradient_input_is_stale = true;
        }

1070
1071
        friend void serialize(const add_layer& item, std::ostream& out)
        {
1072
            int version = 2;
1073
1074
1075
1076
1077
            serialize(version, out);
            serialize(item.input_layer, out);
            serialize(item.details, out);
            serialize(item.this_layer_setup_called, out);
            serialize(item.gradient_input_is_stale, out);
1078
            serialize(item.get_output_and_gradient_input_disabled, out);
1079
1080
            serialize(item.x_grad, out);
            serialize(item.cached_output, out);
1081
            serialize(item.grad_final, out);
1082
1083
1084
1085
1086
1087
        }

        friend void deserialize(add_layer& item, std::istream& in)
        {
            int version = 0;
            deserialize(version, in);
1088
            if (version != 2)
1089
1090
1091
1092
1093
                throw serialization_error("Unexpected version found while deserializing dlib::add_layer.");
            deserialize(item.input_layer, in);
            deserialize(item.details, in);
            deserialize(item.this_layer_setup_called, in);
            deserialize(item.gradient_input_is_stale, in);
1094
            deserialize(item.get_output_and_gradient_input_disabled, in);
1095
1096
            deserialize(item.x_grad, in);
            deserialize(item.cached_output, in);
1097
            deserialize(item.grad_final, in);
1098
1099
        }

1100
1101
    private:

1102
1103
1104
        bool this_layer_requires_forward_output(
        ) 
        {
1105
            subnet_wrapper wsub(grad_final, grad_final);
1106
1107
1108
            return impl::backward_requires_forward_output(details, wsub);
        }

Davis King's avatar
Davis King committed
1109
        class subnet_wrapper
1110
1111
        {
        public:
1112
1113
            subnet_wrapper(const tensor& x_, resizable_tensor& grad_final_) :
                x(x_), grad_final(grad_final_) {}
1114

Davis King's avatar
Davis King committed
1115
1116
            subnet_wrapper(const subnet_wrapper&) = delete;
            subnet_wrapper& operator=(const subnet_wrapper&) = delete;
1117

1118
1119
1120
            const tensor& get_output() const { return x; }
            tensor& get_gradient_input() 
            { 
1121
                if (!have_same_dimensions(x, grad_final))
1122
                {
1123
1124
                    grad_final.copy_size(x);
                    grad_final = 0;  
1125
                }
1126
                return grad_final; 
1127
1128
1129
1130
            }

        private:
            const tensor& x;
1131
            resizable_tensor& grad_final;
1132
1133
        };

1134
1135
1136
1137
1138
1139
        void swap(add_layer& item)
        {
            std::swap(input_layer, item.input_layer);
            std::swap(details, item.details);
            std::swap(this_layer_setup_called, item.this_layer_setup_called);
            std::swap(gradient_input_is_stale, item.gradient_input_is_stale);
1140
            std::swap(get_output_and_gradient_input_disabled, item.get_output_and_gradient_input_disabled);
1141
1142
            std::swap(x_grad, item.x_grad); 
            std::swap(cached_output, item.cached_output); 
1143
            std::swap(grad_final, item.grad_final); 
1144
1145
        }

Davis King's avatar
Davis King committed
1146
        subnet_type input_layer;
1147
1148
1149
        LAYER_DETAILS details;
        bool this_layer_setup_called;
        bool gradient_input_is_stale;
1150
        bool get_output_and_gradient_input_disabled;
1151
1152
        resizable_tensor x_grad; 
        resizable_tensor cached_output; 
1153
        resizable_tensor grad_final;
1154

1155
        // The following 2 objects don't logically contribute to the state of this class.
1156
1157
1158
1159
1160
1161
1162
1163
        // They are only here to prevent them from being reallocated over and over in
        // member functions.
        resizable_tensor params_grad; 
        resizable_tensor temp_tensor; 
    };

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

Davis King's avatar
Davis King committed
1164
    template <unsigned long ID, typename SUBNET, typename enabled=void>
1165
1166
    class add_tag_layer;

Davis King's avatar
Davis King committed
1167
1168
1169
    template <unsigned long ID, typename SUBNET>
    class add_tag_layer<ID,SUBNET,
            typename std::enable_if<is_nonloss_layer_type<SUBNET>::value>::type>
1170
1171
    {
    public:
Davis King's avatar
Davis King committed
1172
1173
        typedef SUBNET subnet_type;
        typedef typename subnet_type::input_type input_type;
1174
        const static size_t num_layers = subnet_type::num_layers;
Davis King's avatar
Davis King committed
1175
        const static unsigned int sample_expansion_factor = subnet_type::sample_expansion_factor;
1176
1177
1178
        static_assert(sample_expansion_factor >= 1,
            "The input layer can't produce fewer output tensors than there are inputs.");

Davis King's avatar
Davis King committed
1179
1180
1181
1182
1183
        add_tag_layer() = default;
        add_tag_layer(const add_tag_layer&) = default;
        add_tag_layer(add_tag_layer&&) = default;
        add_tag_layer& operator=(add_tag_layer&&) = default;
        add_tag_layer& operator=(const add_tag_layer&) = default;
1184
1185

        template <typename T>
Davis King's avatar
Davis King committed
1186
1187
        add_tag_layer(
            const add_tag_layer<ID,T>& item
Davis King's avatar
Davis King committed
1188
        ) : subnetwork(item.subnet())
1189
1190
1191
        {}

        template <typename ...T>
Davis King's avatar
Davis King committed
1192
        add_tag_layer(
1193
1194
            T ...args
        ) : 
Davis King's avatar
Davis King committed
1195
            subnetwork(std::move(args)...) 
1196
1197
1198
1199
1200
        {
        }

        template <typename input_iterator>
        void to_tensor (
1201
1202
            input_iterator ibegin,
            input_iterator iend,
1203
1204
1205
            resizable_tensor& data
        ) const
        {
Davis King's avatar
Davis King committed
1206
            subnetwork.to_tensor(ibegin,iend,data);
1207
1208
1209
1210
1211
1212
1213
1214
        }

        template <typename input_iterator>
        const tensor& operator() (
            input_iterator ibegin,
            input_iterator iend
        )
        {
Davis King's avatar
Davis King committed
1215
            return subnetwork(ibegin,iend);
1216
1217
1218
1219
        }

        const tensor& operator() (const input_type& x)
        {
Davis King's avatar
Davis King committed
1220
            return subnetwork(x);
1221
1222
1223
1224
        }

        const tensor& forward(const tensor& x)
        {
Davis King's avatar
Davis King committed
1225
            return subnetwork.forward(x);
1226
1227
        }

Davis King's avatar
Davis King committed
1228
        const tensor& get_output() const { return subnetwork.get_output(); }
1229
1230
1231

        tensor& get_gradient_input() 
        { 
Davis King's avatar
Davis King committed
1232
            return subnetwork.get_gradient_input();
1233
1234
        }

1235
1236
1237
        const tensor& get_final_data_gradient(
        ) const { return subnetwork.get_final_data_gradient(); }

1238
        template <typename solver_type>
1239
        void update(const tensor& x, sstack<solver_type> solvers, double step_size)
1240
        {
1241
            subnetwork.update(x,solvers, step_size);
1242
1243
        }

1244
        template <typename solver_type>
1245
        void update(const tensor& x, const tensor& gradient_input, sstack<solver_type> solvers, double step_size)
1246
        {
1247
            subnetwork.update(x,gradient_input,solvers, step_size);
1248
1249
        }

Davis King's avatar
Davis King committed
1250
1251
        const subnet_type& subnet() const { return subnetwork; }
        subnet_type& subnet() { return subnetwork; }
1252
1253
1254

        void clean()
        {
Davis King's avatar
Davis King committed
1255
            subnetwork.clean();
1256
1257
        }

1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
        friend void serialize(const add_tag_layer& item, std::ostream& out)
        {
            int version = 1;
            serialize(version, out);
            serialize(item.subnetwork, out);
        }

        friend void deserialize(add_tag_layer& item, std::istream& in)
        {
            int version = 0;
            deserialize(version, in);
            if (version != 1)
                throw serialization_error("Unexpected version found while deserializing dlib::add_tag_layer.");
            deserialize(item.subnetwork, in);
        }

1274
1275
    private:

1276
1277
1278
1279
1280
1281
1282
1283
        template <typename T, typename U, typename E>
        friend class add_layer;
        template <typename T, bool is_first, typename E>
        friend class dimpl::subnet_wrapper;
        template <unsigned long T, typename U, typename E>
        friend class add_tag_layer;
        template <template<typename> class T, typename U>
        friend class add_skip_layer;
1284
1285
        template <size_t N, template<typename> class L, typename S>
        friend class repeat;
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307

        // You woudln't put a tag on a layer if you didn't want to access its forward
        // outputs.  So this is always true.
        bool this_layer_requires_forward_output(
        ) { return true; } 

        void disable_output_and_gradient_getters (
        ) 
        { 
            // This should never happen because only inplace layers call
            // disable_output_and_gradient_getters(), however, putting a tag layer right
            // before an inplace layer basically means you don't want the following layer
            // to operate in place.  So the inplace layer should turn itself into an
            // out-of-place layer and not call disable_output_and_gradient_getters(). 
            DLIB_CASSERT(false,"This should never happen");
        }

        tensor& private_get_output() const
        { return subnetwork.private_get_output(); }
        tensor& private_get_gradient_input() 
        { return subnetwork.private_get_gradient_input(); }

Davis King's avatar
Davis King committed
1308
        subnet_type subnetwork;
1309
1310
    };

1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
// ----------------------------------------------------------------------------------------

    namespace impl
    {
        class repeat_input_layer 
        {
            /*!
                None of the declarations in this object are really used. The only reason it
                exists is to allow the repeat object to use a special input layer in its
                internal networks which will cause add_tag_layer objects that happen to be
                right at the input to not create copies of their input tensors.  So
                introducing the repeat_input_layer object allows us to optimize the
                implementation of add_tag_layer for a special case that arises when it's
                used in the context of the repeat layer.
            !*/
        public:
            typedef int input_type;
            const static unsigned int sample_expansion_factor = 1;

            template <typename input_iterator>
            void to_tensor (
                input_iterator ,
                input_iterator ,
                resizable_tensor& 
            ) const
            {
                DLIB_CASSERT(false,"This function should never be called");
            }

            friend void serialize(const repeat_input_layer&, std::ostream&){}
            friend void deserialize(repeat_input_layer&, std::istream&){}
        };
    }

    template <
        size_t num,
        template<typename> class LAYER, 
        typename SUBNET
        >
    class repeat
    {
        static_assert(num > 0, "You can't have a layer repeated 0 times.");
    public:
        typedef SUBNET subnet_type;
        typedef typename SUBNET::input_type input_type;
        const static size_t num_layers = (LAYER<SUBNET>::num_layers-SUBNET::num_layers)*num + SUBNET::num_layers;
        const static unsigned int sample_expansion_factor = SUBNET::sample_expansion_factor;

        typedef LAYER<impl::repeat_input_layer> repeated_layer_type;

        repeat(
        ) : 
            details(num)
        {
        }

        size_t num_repetitions (
        ) const { return num; }

        const repeated_layer_type& get_repeated_layer (
            size_t i 
        ) const
        { 
            DLIB_CASSERT(i < num_repetitions(), "");
            return details[i]; 
        }

        repeated_layer_type& get_repeated_layer (
            size_t i 
        ) 
        { 
            DLIB_CASSERT(i < num_repetitions(), "");
            return details[i]; 
        }

        repeat(const repeat&) = default;
        repeat(repeat&&) = default;
        repeat& operator=(repeat&&) = default;
        repeat& operator=(const repeat&) = default;

        template <template<typename> class T, typename U>
        repeat(
            const repeat<num,T,U>& item
        ) : 
            subnetwork(item.subnetwork)
        {
            for (auto&& d : item.details)
                details.emplace_back(d);
        }

        template <typename T, typename ...U>
        repeat(
            T arg1,
            U ...args2
        ): 
            details(num, std::move(arg1)),
            subnetwork(std::move(args2)...)
        {
        }

        template <typename T, typename ...U>
        repeat(
            std::tuple<>,
            T arg1,
            U ...args2
        ): 
            details(num, std::move(arg1)),
            subnetwork(std::move(args2)...)
        {
        }

        template <typename input_iterator>
        void to_tensor (
            input_iterator ibegin,
            input_iterator iend,
            resizable_tensor& data
        ) const
        {
            subnetwork.to_tensor(ibegin,iend,data);
        }

        template <typename input_iterator>
        const tensor& operator() (
            input_iterator ibegin,
            input_iterator iend
        )
        {
            to_tensor(ibegin,iend,temp_tensor);
            return forward(temp_tensor);
        }

        const tensor& operator() (const input_type& x)
        {
            return (*this)(&x, &x+1);
        }

        const tensor& forward(const tensor& x)
        {
            subnetwork.forward(x);
            details[details.size()-1].forward(subnetwork.get_output());
            for (long i = details.size()-2; i >= 0; --i)
                details[i].forward(details[i+1].get_output());
            return private_get_output();
        }

    private:
        tensor& private_get_output() const
        { 
            return details[0].private_get_output();
        }
        tensor& private_get_gradient_input() 
        { 
            return details[0].private_get_gradient_input();
        }
    public:
        const tensor& get_output() const 
        { 
            return details[0].get_output(); 
        }
        tensor& get_gradient_input() 
        { 
            return details[0].get_gradient_input();
        }

        template <typename solver_type>
1476
        void update(const tensor& x, sstack<solver_type> solvers, double step_size)
1477
        {
1478
            update(x,private_get_gradient_input(),solvers,step_size);
1479
1480
1481
        }

        template <typename solver_type>
1482
        void update(const tensor& x, const tensor& gradient_input, sstack<solver_type> solvers, double step_size)
1483
1484
1485
1486
        {
            const auto cnt = (LAYER<SUBNET>::num_layers-SUBNET::num_layers);
            if (details.size() > 1)
            {
1487
                details[0].update(details[1].get_output(), gradient_input, solvers,step_size);
1488
1489
1490
                for (size_t i = 1; i < details.size(); ++i)
                {
                    if (i+1 < details.size())
1491
                        details[i].update(details[i+1].get_output(), details[i-1].get_final_data_gradient(), solvers.pop(cnt*i),step_size);
1492
                    else
1493
                        details[i].update(subnetwork.get_output(), details[i-1].get_final_data_gradient(), solvers.pop(cnt*i),step_size);
1494
1495
1496
1497
                }
            }
            else
            {
1498
                details[0].update(subnetwork.get_output(), gradient_input, solvers,step_size);
1499
            }
1500
            subnetwork.update(x, details.back().get_final_data_gradient(), solvers.pop(cnt*details.size()),step_size);
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
        }

        const subnet_type& subnet() const { return subnetwork; }
        subnet_type& subnet() { return subnetwork; }

        void clean()
        {
            temp_tensor.clear();
            subnetwork.clean();
            for (auto&& d : details)
                d.clean();
        }

        friend void serialize(const repeat& item, std::ostream& out)
        {
            int version = 1;
            serialize(version, out);
            serialize(item.details, out);
            serialize(item.subnetwork, out);
        }

        friend void deserialize(repeat& item, std::istream& in)
        {
            int version = 0;
            deserialize(version, in);
            if (version != 1)
                throw serialization_error("Unexpected version found while deserializing dlib::repeat.");
            deserialize(item.details, in);
            deserialize(item.subnetwork, in);
        }

    private:

        template <typename T, typename U, typename E>
        friend class add_layer;
        template <typename T, bool is_first, typename E>
        friend class dimpl::subnet_wrapper;
        template <unsigned long T, typename U, typename E>
        friend class add_tag_layer;
        template <template<typename> class T, typename U>
        friend class add_skip_layer;
        template <size_t N, template<typename> class L, typename S>
        friend class repeat;

        bool this_layer_requires_forward_output(
        ) 
        { 
            return details[0].this_layer_requires_forward_output(); 
        } 

        void disable_output_and_gradient_getters (
        ) 
        { 
            details[0].disable_output_and_gradient_getters();
        }


        std::vector<repeated_layer_type> details; 
        subnet_type subnetwork;

        // temp_tensor doesn't logically contribute to the state of this class.
        // It is here only to void needing to reallocate it over and over.
        resizable_tensor temp_tensor;
    };

    template <
        size_t num,
        template<typename> class LAYER, 
        typename SUBNET
        >
    struct is_nonloss_layer_type<repeat<num,LAYER,SUBNET>> : std::true_type {};

1573
1574
// ----------------------------------------------------------------------------------------

1575
// This version of add_tag_layer handles the special case where the subnetwork being given
1576
1577
1578
1579
1580
// is just an input layer object.
    template <unsigned long ID, typename INPUT_LAYER, typename enabled>
    class add_tag_layer
    {
    public:
Davis King's avatar
Davis King committed
1581
1582
        typedef INPUT_LAYER subnet_type;
        typedef typename subnet_type::input_type input_type;
1583
        const static size_t num_layers = 0;
Davis King's avatar
Davis King committed
1584
        const static unsigned int sample_expansion_factor = subnet_type::sample_expansion_factor;
1585
1586
1587
        static_assert(sample_expansion_factor >= 1,
            "The input layer can't produce fewer output tensors than there are inputs.");

1588
1589
        add_tag_layer():gradient_input_is_stale(true),cached_output_ptr(nullptr) {}

1590
1591
        add_tag_layer(const add_tag_layer&) = default;
        add_tag_layer& operator=(const add_tag_layer&) = default;
1592
1593
        add_tag_layer(add_tag_layer&& item) : add_tag_layer() { swap(item); }
        add_tag_layer& operator=(add_tag_layer&& item) { swap(item); return *this; }
1594
1595
1596
1597

        template <typename T, typename E>
        add_tag_layer(
            const add_tag_layer<ID,T,E>& item
1598
1599
1600
1601
1602
        ) : input_layer(item.subnet()), 
            cached_output(item.cached_output),
            cached_output_ptr(nullptr),
            grad_final(item.grad_final),
            gradient_input_is_stale(item.gradient_input_is_stale)
1603
1604
1605
1606
1607
1608
        {}

        template <typename ...T>
        add_tag_layer(
            T ...args
        ) : 
1609
1610
1611
            input_layer(std::move(args)...),
            cached_output_ptr(nullptr),
            gradient_input_is_stale(true)
1612
1613
1614
        {
        }

1615
1616
1617
1618
1619
1620
1621
        add_tag_layer (
            std::tuple<>
        ) : 
            cached_output_ptr(nullptr),
            gradient_input_is_stale(true)
        {}

1622
1623
        template <typename input_iterator>
        void to_tensor (
1624
1625
            input_iterator ibegin,
            input_iterator iend,
1626
1627
1628
            resizable_tensor& data
        ) const
        {
1629
            input_layer.to_tensor(ibegin,iend,data);
1630
1631
1632
1633
        }

        template <typename input_iterator>
        const tensor& operator() (
1634
            input_iterator ibegin, 
1635
1636
1637
1638
            input_iterator iend
        )
        {
            input_layer.to_tensor(ibegin,iend,cached_output);
1639
            cached_output_ptr = nullptr;
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
            return get_output();
        }

        const tensor& operator() (const input_type& x)
        {
            return (*this)(&x, &x+1);
        }

        const tensor& forward(const tensor& x)
        {
1650
1651
1652
1653
1654
1655
1656
1657
1658
            // If this tag is the first layer in one of the sub networks inside a repeat
            // layer then we don't want it to be creating copies of x.  This is because, we
            // can just hold a pointer to x since the way repeat is constructed guarantees
            // that x will have a lifetime larger than this pointer. 
            if (is_same_type<INPUT_LAYER, impl::repeat_input_layer>::value)
                cached_output_ptr = const_cast<tensor*>(&x);
            else
                cached_output = x;
            gradient_input_is_stale = true;
1659
1660
1661
1662
1663
            return get_output();
        }

        const tensor& get_output() const 
        { 
1664
1665
1666
1667
            if (cached_output_ptr)
                return *cached_output_ptr;
            else
                return cached_output; 
1668
1669
        }

1670
        const tensor& get_final_data_gradient(
1671
        ) const { return grad_final; }
1672

1673
1674
        tensor& get_gradient_input() 
        { 
1675
1676
            if (!have_same_dimensions(get_output(), grad_final) ||
                gradient_input_is_stale)
1677
            {
1678
1679
1680
                grad_final.copy_size(get_output());
                grad_final = 0;
                gradient_input_is_stale = false;
1681
            }
1682
            return grad_final; 
1683
1684
1685
        }

        template <typename solver_type>
1686
1687
1688
1689
1690
        void update(
            const tensor& /*x*/, 
            sstack<solver_type> /*solvers*/,
            double /*step_size*/
        )
1691
1692
1693
1694
        {
            // nothing to update
        }

1695
        template <typename solver_type>
1696
1697
1698
1699
1700
1701
        void update(
            const tensor& /*x*/,
            const tensor& /*gradient_input*/,
            sstack<solver_type> /*solvers*/,
            double /*step_size*/
        )
1702
1703
1704
1705
        {
            // nothing to update
        }

Davis King's avatar
Davis King committed
1706
1707
        const subnet_type& subnet() const { return input_layer; }
        subnet_type& subnet() { return input_layer; }
1708
1709
1710

        void clean()
        {
1711
            grad_final.clear();
1712
            cached_output.clear();
1713
            cached_output_ptr = 0;
1714
1715
        }

1716
1717
1718
1719
1720
1721
        friend void serialize(const add_tag_layer& item, std::ostream& out)
        {
            int version = 1;
            serialize(version, out);
            serialize(item.input_layer, out);
            serialize(item.cached_output, out);
1722
1723
            serialize(item.grad_final, out);
            serialize(item.gradient_input_is_stale, out);
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
        }

        friend void deserialize(add_tag_layer& item, std::istream& in)
        {
            int version = 0;
            deserialize(version, in);
            if (version != 1)
                throw serialization_error("Unexpected version found while deserializing dlib::add_tag_layer.");
            deserialize(item.input_layer, in);
            deserialize(item.cached_output, in);
1734
1735
1736
            deserialize(item.grad_final, in);
            deserialize(item.gradient_input_is_stale, in);
            item.cached_output_ptr = nullptr;
1737
1738
        }

1739
1740
    private:

1741
1742
1743
1744
1745
1746
1747
1748
        template <typename T, typename U, typename E>
        friend class add_layer;
        template <typename T, bool is_first, typename E>
        friend class dimpl::subnet_wrapper;
        template <unsigned long T, typename U, typename E>
        friend class add_tag_layer;
        template <template<typename> class T, typename U>
        friend class add_skip_layer;
1749
1750
        template <size_t N, template<typename> class L, typename S>
        friend class repeat;
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768

        // You woudln't put a tag on a layer if you didn't want to access its forward
        // outputs.  So this is always true.
        bool this_layer_requires_forward_output(
        ) { return true; } 

        void disable_output_and_gradient_getters (
        ) 
        { 
            // This should never happen because only inplace layers call
            // disable_output_and_gradient_getters(), however, putting a tag layer right
            // before an inplace layer basically means you don't want the following layer
            // to operate in place.  So the inplace layer should turn itself into an
            // out-of-place layer and not call disable_output_and_gradient_getters(). 
            DLIB_CASSERT(false,"This should never happen");
        }

        tensor& private_get_output() const
1769
        { return const_cast<tensor&>(get_output()); }
1770
1771
1772
        tensor& private_get_gradient_input() 
        { return get_gradient_input(); }

1773
1774
1775
1776
        void swap(add_tag_layer& item)
        {
            std::swap(input_layer, item.input_layer);
            std::swap(cached_output, item.cached_output);
1777
1778
1779
            std::swap(cached_output_ptr, item.cached_output_ptr);
            std::swap(grad_final, item.grad_final);
            std::swap(gradient_input_is_stale, item.gradient_input_is_stale);
1780
1781
        }

Davis King's avatar
Davis King committed
1782
        subnet_type input_layer;
1783
        resizable_tensor cached_output;
1784
1785
1786
        tensor* cached_output_ptr;
        resizable_tensor grad_final;
        bool gradient_input_is_stale;
1787
1788
1789
1790
    };

    template <unsigned long ID, typename U, typename E>
    struct is_nonloss_layer_type<add_tag_layer<ID,U,E>> : std::true_type {};
1791
1792
1793
1794
1795
1796


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

Davis King's avatar
Davis King committed
1797
    template <typename LOSS_DETAILS, typename SUBNET>
Davis King's avatar
Davis King committed
1798
    class add_loss_layer;
1799
1800
1801
1802
1803

    class no_label_type
    {
    private:
        // We don't want anyone making these no_label_type objects.  They are here only to
Davis King's avatar
Davis King committed
1804
        // allow add_loss_layer::label_type and dnn_trainer::label_type to exist which avoids
Davis King's avatar
Davis King committed
1805
        // needing to overload add_loss_layer and dnn_trainer for supervised an unsupervised
1806
1807
        // losses.  It also can be a type to use in template metaprogramming to indicate
        // "no label".  So here we make the constructor private with the exception that
Davis King's avatar
Davis King committed
1808
        // add_loss_layer objects can make it (again, just to simplify add_loss_layer's
1809
        // implementation).
1810
        no_label_type(){};
Davis King's avatar
Davis King committed
1811
        template <typename LOSS_DETAILS, typename SUBNET> friend class add_loss_layer;
1812
        template < typename net_type, typename solver_type > friend class dnn_trainer; 
1813
1814
1815
1816
    };

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

Davis King's avatar
Davis King committed
1817
    template <typename LOSS_DETAILS, typename SUBNET>
Davis King's avatar
Davis King committed
1818
    class add_loss_layer
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
    {
        template <typename T, typename enabled=void>
        struct get_loss_layer_label_type
        {
            typedef no_label_type type;
        };
        template <typename T>
        struct get_loss_layer_label_type<T,typename std::enable_if<sizeof(typename T::label_type)!=0>::type>
        {
            typedef typename T::label_type type;
        };

    public:
        typedef LOSS_DETAILS loss_details_type;
Davis King's avatar
Davis King committed
1833
1834
        typedef SUBNET subnet_type;
        typedef typename subnet_type::input_type input_type;
1835
        // Note that the loss layer doesn't count as an additional layer.
Davis King's avatar
Davis King committed
1836
1837
        const static size_t num_layers = subnet_type::num_layers;
        const static unsigned int sample_expansion_factor = subnet_type::sample_expansion_factor;
1838
1839
        typedef typename get_loss_layer_label_type<LOSS_DETAILS>::type label_type;

1840
1841
        static_assert(is_nonloss_layer_type<SUBNET>::value, 
            "SUBNET must be of type add_layer, add_skip_layer, or add_tag_layer."); 
1842
1843
1844
1845
        static_assert(sample_expansion_factor == LOSS_DETAILS::sample_expansion_factor,
            "The loss layer and input layer must agree on the sample_expansion_factor.");


1846
        add_loss_layer() {};
Davis King's avatar
Davis King committed
1847
1848
        add_loss_layer(const add_loss_layer&) = default;
        add_loss_layer& operator=(const add_loss_layer&) = default;
1849
1850
        add_loss_layer(add_loss_layer&& item) : add_loss_layer() { swap(item); }
        add_loss_layer& operator=(add_loss_layer&& item) { swap(item); return *this; }
1851
1852

        template <typename T, typename U>
Davis King's avatar
Davis King committed
1853
1854
        add_loss_layer(
            const add_loss_layer<T,U>& item
1855
1856
        ) : 
            loss(item.loss_details()),
Davis King's avatar
Davis King committed
1857
            subnetwork(item.subnet())
1858
1859
1860
        {}

        template <typename ...T>
Davis King's avatar
Davis King committed
1861
        add_loss_layer(
1862
1863
1864
1865
            const LOSS_DETAILS& layer_det, 
            T&& ...args
        ) : 
            loss(layer_det), 
Davis King's avatar
Davis King committed
1866
            subnetwork(std::forward<T>(args)...)
1867
1868
1869
1870
        {
        }

        template <typename ...T>
Davis King's avatar
Davis King committed
1871
        add_loss_layer(
1872
1873
1874
1875
            LOSS_DETAILS&& layer_det, 
            T&& ...args
        ) : 
            loss(std::move(layer_det)), 
Davis King's avatar
Davis King committed
1876
            subnetwork(std::forward<T>(args)...)
1877
1878
1879
1880
        {
        }

        template <typename ...T>
Davis King's avatar
Davis King committed
1881
        add_loss_layer(
1882
1883
            T ...args
        ) : 
Davis King's avatar
Davis King committed
1884
            subnetwork(std::move(args)...)
1885
        {
1886
1887
1888
1889
1890
1891
1892
1893
1894
        }

        template <typename input_iterator>
        void to_tensor (
            input_iterator ibegin,
            input_iterator iend,
            resizable_tensor& data
        ) const
        {
Davis King's avatar
Davis King committed
1895
            subnetwork.to_tensor(ibegin,iend,data);
1896
1897
1898
1899
1900
1901
1902
1903
        }

        template <typename output_iterator>
        void operator() (
            const tensor& x, 
            output_iterator obegin
        )
        {
Davis King's avatar
Davis King committed
1904
1905
            subnetwork.forward(x);
            const dimpl::subnet_wrapper<subnet_type> wsub(subnetwork);
1906
            loss.to_label(x, wsub, obegin);
1907
1908
1909
1910
1911
1912
1913
1914
1915
        }

        template <typename input_iterator, typename output_iterator>
        void operator() (
            input_iterator ibegin,
            input_iterator iend,
            output_iterator obegin
        )
        {
1916
1917
            to_tensor(ibegin,iend,temp_tensor);
            (*this)(temp_tensor, obegin);
1918
1919
1920
1921
1922
1923
1924
1925
        }

        const label_type& operator() (const input_type& x)
        {
            (*this)(&x, &x+1, &temp_label);
            return temp_label;
        }

1926
        template <typename iterable_type>
1927
        std::vector<label_type> operator() (
1928
            const iterable_type& data,
1929
1930
1931
            size_t batch_size = 128
        )
        {
1932
            std::vector<label_type> results(std::distance(data.begin(), data.end()));
1933
1934
1935
1936
1937
1938
1939
1940
1941
            auto o = results.begin();
            for (auto i = data.begin(); i < data.end(); i+=batch_size, o+=batch_size)
            {
                auto end = std::min(i+batch_size, data.end());
                (*this)(i, end, o);
            }
            return results;
        }

1942
1943
1944
1945
1946
1947
        template <typename label_iterator>
        double compute_loss (
            const tensor& x,
            label_iterator lbegin 
        )
        {
Davis King's avatar
Davis King committed
1948
1949
            subnetwork.forward(x);
            dimpl::subnet_wrapper<subnet_type> wsub(subnetwork);
1950
1951
            return loss.compute_loss(x, lbegin, wsub);
        }
1952
1953
1954
1955
1956
1957
1958
1959

        template <typename input_iterator, typename label_iterator>
        double compute_loss (
            input_iterator ibegin,
            input_iterator iend,
            label_iterator lbegin 
        )
        {
1960
1961
1962
1963
1964
1965
1966
1967
            to_tensor(ibegin,iend,temp_tensor);
            return compute_loss(temp_tensor, lbegin);
        }

        double compute_loss (
            const tensor& x
        )
        {
Davis King's avatar
Davis King committed
1968
1969
            subnetwork.forward(x);
            dimpl::subnet_wrapper<subnet_type> wsub(subnetwork);
1970
            return loss.compute_loss(x, wsub);
1971
1972
1973
1974
1975
1976
1977
1978
        }

        template <typename input_iterator>
        double compute_loss (
            input_iterator ibegin,
            input_iterator iend
        )
        {
1979
1980
1981
1982
1983
1984
1985
1986
            to_tensor(ibegin,iend,temp_tensor);
            return compute_loss(temp_tensor);
        }

        template <typename label_iterator, typename solver_type>
        double update (
            const tensor& x,
            label_iterator lbegin,
1987
1988
            sstack<solver_type> solvers,
            double step_size
1989
1990
        )
        {
Davis King's avatar
Davis King committed
1991
1992
            subnetwork.forward(x);
            dimpl::subnet_wrapper<subnet_type> wsub(subnetwork);
1993
            double l = loss.compute_loss(x, lbegin, wsub);
1994
            subnetwork.update(x, solvers, step_size);
1995
            return l;
1996
1997
1998
1999
2000
2001
2002
        }

        template <typename input_iterator, typename label_iterator, typename solver_type>
        double update (
            input_iterator ibegin,
            input_iterator iend,
            label_iterator lbegin,
2003
2004
            sstack<solver_type> solvers,
            double step_size
2005
2006
        )
        {
2007
            to_tensor(ibegin,iend,temp_tensor);
2008
            return update(temp_tensor, lbegin, solvers, step_size);
2009
2010
2011
2012
2013
        }

        template <typename solver_type>
        double update (
            const tensor& x,
2014
2015
            sstack<solver_type> solvers,
            double step_size
2016
2017
        )
        {
Davis King's avatar
Davis King committed
2018
2019
            subnetwork.forward(x);
            dimpl::subnet_wrapper<subnet_type> wsub(subnetwork);
2020
            double l = loss.compute_loss(x, wsub);
2021
            subnetwork.update(x, solvers, step_size);
2022
2023
2024
2025
2026
2027
2028
            return l;
        }

        template <typename input_iterator, typename solver_type>
        double update (
            input_iterator ibegin,
            input_iterator iend,
2029
2030
            sstack<solver_type> solvers,
            double step_size
2031
2032
        )
        {
2033
            to_tensor(ibegin,iend,temp_tensor);
2034
            return update(temp_tensor, solvers, step_size);
2035
2036
        }

Davis King's avatar
Davis King committed
2037
2038
        const subnet_type& subnet() const { return subnetwork; }
        subnet_type& subnet() { return subnetwork; }
2039
2040
2041
2042
2043
2044
2045
        const loss_details_type& loss_details() const { return loss; }
        loss_details_type& loss_details() { return loss; }

        void clean (
        )
        {
            temp_tensor.clear();
2046
            subnetwork.clean();
2047
2048
        }

2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
        friend void serialize(const add_loss_layer& item, std::ostream& out)
        {
            int version = 1;
            serialize(version, out);
            serialize(item.loss, out);
            serialize(item.subnetwork, out);
        }

        friend void deserialize(add_loss_layer& item, std::istream& in)
        {
            int version = 0;
            deserialize(version, in);
            if (version != 1)
                throw serialization_error("Unexpected version found while deserializing dlib::add_loss_layer.");
            deserialize(item.loss, in);
            deserialize(item.subnetwork, in);
        }

2067
2068
    private:

2069
2070
2071
2072
2073
2074
        void swap(add_loss_layer& item)
        {
            std::swap(loss, item.loss);
            std::swap(subnetwork, item.subnetwork);
        }

2075
        loss_details_type loss;
Davis King's avatar
Davis King committed
2076
        subnet_type subnetwork;
2077
2078
2079
2080
2081
2082
2083
2084
2085

        // These two objects don't logically contribute to the state of this object.  They
        // are here to prevent them from being reallocated over and over.
        label_type temp_label;
        resizable_tensor temp_tensor;
    };


    template <typename T, typename U>
Davis King's avatar
Davis King committed
2086
    struct is_loss_layer_type<add_loss_layer<T,U>> : std::true_type {};
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097

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

    namespace impl
    {
        template <unsigned int i, typename T>
        struct layer_helper
        {
            static T& makeT();
Davis King's avatar
Davis King committed
2098
            using next_type = typename std::remove_reference<decltype(makeT().subnet())>::type;
2099
2100
2101
            using type = typename layer_helper<i-1,next_type>::type;
            static type& layer(T& n)
            {
Davis King's avatar
Davis King committed
2102
                return layer_helper<i-1,next_type>::layer(n.subnet());
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
            }
        };
        template <typename T>
        struct layer_helper<0,T>
        {
            using type = T;
            static type& layer(T& n)
            {
                return n;
            }
        };

        template <template<typename> class Match, typename T, unsigned int i, typename enabled = void>
        struct layer_helper_match
        {
            static T& makeT();
Davis King's avatar
Davis King committed
2119
            using next_type = typename std::remove_reference<decltype(makeT().subnet())>::type;
2120
2121
2122
            using type = typename layer_helper_match<Match,next_type,i>::type;
            static type& layer(T& n)
            {
Davis King's avatar
Davis King committed
2123
                return layer_helper_match<Match,next_type,i>::layer(n.subnet());
2124
2125
            }
        };
Davis King's avatar
Davis King committed
2126
        // This overload catches add_layer and add_loss_layer templates.
2127
2128
        template <template<typename> class Match, typename T, unsigned int i>
        struct layer_helper_match<Match,T,i,
Davis King's avatar
Davis King committed
2129
            typename std::enable_if<std::is_same<const T,const  Match<typename T::subnet_type>>::value>::type>
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
        {
            using type = typename layer_helper<i,T>::type;
            static type& layer(T& n)
            {
                return layer_helper<i,T>::layer(n);
            }
        };
        // This overload catches input templates.
        template <template<typename> class Match, typename T, unsigned int i>
        struct layer_helper_match<Match,T,i,
            typename std::enable_if<std::is_same<const T,const  Match<typename T::input_type>>::value>::type>
        {
            using type = typename layer_helper<i,T>::type;
            static type& layer(T& n)
            {
                return layer_helper<i,T>::layer(n);
            }
        };
Davis King's avatar
Davis King committed
2148
        // This overload catches subnet_wrapper templates.
2149
2150
2151
        template <template<typename> class Match, typename T, unsigned int i>
        struct layer_helper_match<Match,T,i,
            typename std::enable_if<std::is_same<const typename T::wrapped_type, 
Davis King's avatar
Davis King committed
2152
                                                 const Match<typename T::wrapped_type::subnet_type>>::value>::type>
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
        {
            using type = typename layer_helper<i,T>::type;
            static type& layer(T& n)
            {
                return layer_helper<i,T>::layer(n);
            }
        };
    }

    template <unsigned int i, typename T>
    typename impl::layer_helper<i,T>::type& layer (T& n) 
    {
        return impl::layer_helper<i,T>::layer(n);
    }

    template <template<typename> class Match, typename T>
    typename impl::layer_helper_match<Match,T,0>::type& layer (T& n) 
    {
        return impl::layer_helper_match<Match,T,0>::layer(n);
    }

    template <template<typename> class Match, unsigned int i, typename T>
    typename impl::layer_helper_match<Match,T,i>::type& layer (T& n) 
    {
        return impl::layer_helper_match<Match,T,i>::layer(n);
    }

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

Davis King's avatar
Davis King committed
2182
    template <template<typename> class TAG_TYPE, typename SUBNET>
Davis King's avatar
Davis King committed
2183
    class add_skip_layer
2184
2185
    {
    public:
Davis King's avatar
Davis King committed
2186
2187
        typedef SUBNET subnet_type;
        typedef typename subnet_type::input_type input_type;
2188
        const static size_t num_layers = subnet_type::num_layers;
Davis King's avatar
Davis King committed
2189
        const static unsigned int sample_expansion_factor = subnet_type::sample_expansion_factor;
2190
2191
2192
        static_assert(sample_expansion_factor >= 1,
            "The input layer can't produce fewer output tensors than there are inputs.");

Davis King's avatar
Davis King committed
2193
2194
2195
2196
2197
        add_skip_layer() = default;
        add_skip_layer(const add_skip_layer&) = default;
        add_skip_layer(add_skip_layer&&) = default;
        add_skip_layer& operator=(add_skip_layer&&) = default;
        add_skip_layer& operator=(const add_skip_layer&) = default;
2198
2199

        template <typename T>
Davis King's avatar
Davis King committed
2200
2201
        add_skip_layer(
            const add_skip_layer<TAG_TYPE,T>& item
Davis King's avatar
Davis King committed
2202
        ) : subnetwork(item.subnet())
2203
2204
2205
        {}

        template <typename ...T>
Davis King's avatar
Davis King committed
2206
        add_skip_layer(
2207
2208
            T ...args
        ) : 
Davis King's avatar
Davis King committed
2209
            subnetwork(std::move(args)...) 
2210
2211
2212
2213
2214
        {
        }

        template <typename input_iterator>
        void to_tensor (
2215
2216
            input_iterator ibegin,
            input_iterator iend,
2217
2218
2219
            resizable_tensor& data
        ) const
        {
Davis King's avatar
Davis King committed
2220
            subnetwork.to_tensor(ibegin,iend,data);
2221
2222
2223
2224
2225
2226
2227
2228
        }

        template <typename input_iterator>
        const tensor& operator() (
            input_iterator ibegin,
            input_iterator iend
        )
        {
Davis King's avatar
Davis King committed
2229
2230
            subnetwork(ibegin,iend);
            return layer<TAG_TYPE>(subnetwork).get_output();
2231
2232
2233
2234
        }

        const tensor& operator() (const input_type& x)
        {
Davis King's avatar
Davis King committed
2235
2236
            subnetwork(x);
            return layer<TAG_TYPE>(subnetwork).get_output();
2237
2238
2239
2240
        }

        const tensor& forward(const tensor& x)
        {
Davis King's avatar
Davis King committed
2241
2242
            subnetwork.forward(x);
            return layer<TAG_TYPE>(subnetwork).get_output();
2243
2244
2245
2246
        }

        const tensor& get_output() const 
        { 
Davis King's avatar
Davis King committed
2247
            return layer<TAG_TYPE>(subnetwork).get_output();
2248
2249
2250
2251
        }

        tensor& get_gradient_input() 
        { 
Davis King's avatar
Davis King committed
2252
            return layer<TAG_TYPE>(subnetwork).get_gradient_input();
2253
2254
        }

2255
2256
2257
2258
2259
2260
        const tensor& get_final_data_gradient(
        ) const 
        { 
            return subnetwork.get_final_data_gradient(); 
        }

2261
        template <typename solver_type>
2262
        void update(const tensor& x, sstack<solver_type> solvers)
2263
        {
2264
            subnetwork.update(x,solvers);
2265
2266
        }

2267
        template <typename solver_type>
2268
        void update(const tensor& x, const tensor& gradient_input, sstack<solver_type> solvers)
2269
        {
2270
            subnetwork.update(x,gradient_input,solvers);
2271
2272
        }

Davis King's avatar
Davis King committed
2273
        const subnet_type& subnet() const 
2274
        { 
Davis King's avatar
Davis King committed
2275
            return subnetwork; 
2276
2277
        }

Davis King's avatar
Davis King committed
2278
        subnet_type& subnet() 
2279
        { 
Davis King's avatar
Davis King committed
2280
            return subnetwork; 
2281
2282
2283
2284
        }

        void clean()
        {
Davis King's avatar
Davis King committed
2285
            subnetwork.clean();
2286
2287
        }

2288
2289
2290
2291
2292
2293
2294
2295
2296
2297
2298
2299
2300
2301
2302
2303
        friend void serialize(const add_skip_layer& item, std::ostream& out)
        {
            int version = 1;
            serialize(version, out);
            serialize(item.subnetwork, out);
        }

        friend void deserialize(add_skip_layer& item, std::istream& in)
        {
            int version = 0;
            deserialize(version, in);
            if (version != 1)
                throw serialization_error("Unexpected version found while deserializing dlib::add_skip_layer.");
            deserialize(item.subnetwork, in);
        }

2304
2305
    private:

2306
2307
2308
2309
2310
2311
2312
2313
        template <typename T, typename U, typename E>
        friend class add_layer;
        template <typename T, bool is_first, typename E>
        friend class dimpl::subnet_wrapper;
        template <unsigned long T, typename U, typename E>
        friend class add_tag_layer;
        template <template<typename> class T, typename U>
        friend class add_skip_layer;
2314
2315
        template <size_t N, template<typename> class L, typename S>
        friend class repeat;
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327

        bool this_layer_requires_forward_output(
        ) { return layer<TAG_TYPE>(subnetwork).this_layer_requires_forward_output(); } 

        void disable_output_and_gradient_getters (
        ) { layer<TAG_TYPE>(subnetwork).disable_output_and_gradient_getters(); }

        tensor& private_get_output() const
        { return layer<TAG_TYPE>(subnetwork).private_get_output(); }
        tensor& private_get_gradient_input() 
        { return layer<TAG_TYPE>(subnetwork).private_get_gradient_input(); }

Davis King's avatar
Davis King committed
2328
        subnet_type subnetwork;
2329
2330
    };
    template <template<typename> class T, typename U>
Davis King's avatar
Davis King committed
2331
2332
    struct is_nonloss_layer_type<add_skip_layer<T,U>> : std::true_type {};

Davis King's avatar
Davis King committed
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
    template <typename SUBNET> using tag1  = add_tag_layer< 1, SUBNET>;
    template <typename SUBNET> using tag2  = add_tag_layer< 2, SUBNET>;
    template <typename SUBNET> using tag3  = add_tag_layer< 3, SUBNET>;
    template <typename SUBNET> using tag4  = add_tag_layer< 4, SUBNET>;
    template <typename SUBNET> using tag5  = add_tag_layer< 5, SUBNET>;
    template <typename SUBNET> using tag6  = add_tag_layer< 6, SUBNET>;
    template <typename SUBNET> using tag7  = add_tag_layer< 7, SUBNET>;
    template <typename SUBNET> using tag8  = add_tag_layer< 8, SUBNET>;
    template <typename SUBNET> using tag9  = add_tag_layer< 9, SUBNET>;
    template <typename SUBNET> using tag10 = add_tag_layer<10, SUBNET>;

    template <typename SUBNET> using skip1  = add_skip_layer< tag1, SUBNET>;
    template <typename SUBNET> using skip2  = add_skip_layer< tag2, SUBNET>;
    template <typename SUBNET> using skip3  = add_skip_layer< tag3, SUBNET>;
    template <typename SUBNET> using skip4  = add_skip_layer< tag4, SUBNET>;
    template <typename SUBNET> using skip5  = add_skip_layer< tag5, SUBNET>;
    template <typename SUBNET> using skip6  = add_skip_layer< tag6, SUBNET>;
    template <typename SUBNET> using skip7  = add_skip_layer< tag7, SUBNET>;
    template <typename SUBNET> using skip8  = add_skip_layer< tag8, SUBNET>;
    template <typename SUBNET> using skip9  = add_skip_layer< tag9, SUBNET>;
    template <typename SUBNET> using skip10 = add_skip_layer<tag10, SUBNET>;
2354
2355
2356
2357
2358

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

    namespace timpl
    {
2359
        inline void fill_with_gassuan_random_numbers (
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
            tensor& t,
            dlib::rand& rnd,
            double sigma = 1
        )
        {
            float* data = t.host();
            for (size_t i = 0; i < t.size(); ++i)
                data[i] = rnd.get_random_gaussian()*sigma;
        }

Davis King's avatar
Davis King committed
2370
        class test_layer_subnet 
2371
2372
        {
        public:
Davis King's avatar
Davis King committed
2373
            test_layer_subnet (
2374
2375
2376
2377
2378
2379
                dlib::rand& rnd_
            ) : rnd(rnd_) 
            {
                // Output and gradient_input have to have the same dimensions in each
                // layer.
                const long num_samples = rnd.get_random_32bit_number()%4+3;
2380
                const long k  = rnd.get_random_32bit_number()%4+2;
2381
2382
2383
                const long nr = rnd.get_random_32bit_number()%4+2;
                const long nc = rnd.get_random_32bit_number()%4+2;

2384
2385
                output.set_size(num_samples, k, nr, nc);
                gradient_input.set_size(num_samples, k, nr, nc);
2386
2387
2388
2389
2390
2391
2392
2393
2394

                // Use a non-zero initial gradient to make sure the layers add to it
                // rather than assign and blow away the initial value.
                fill_with_gassuan_random_numbers(gradient_input, rnd, 0.01);

                fill_with_gassuan_random_numbers(output, rnd);
            }


2395
            tensor& get_mutable_output() { return output; }
2396
            const tensor& get_output() const { return output; }
2397
            const tensor& private_get_output() const { return get_output(); }
Davis King's avatar
Davis King committed
2398
            const test_layer_subnet& subnet() const { init_sub(); return *subnetwork; }
2399
2400

            tensor& get_gradient_input() { return gradient_input; }
2401
            tensor& private_get_gradient_input() { return get_gradient_input(); }
Davis King's avatar
Davis King committed
2402
            test_layer_subnet& subnet() { init_sub(); return *subnetwork; }
2403
2404
2405
2406
2407



            unsigned long count_outputs() const
            {
Davis King's avatar
Davis King committed
2408
2409
                if (subnetwork)
                    return subnetwork->count_outputs() + output.size();
2410
2411
2412
2413
2414
2415
2416
2417
2418
                else
                    return output.size();
            }

            float& get_output_element(unsigned long i)
            {
                if (i < output.size())
                    return output.host()[i];
                else
Davis King's avatar
Davis King committed
2419
                    return subnet().get_output_element(i-output.size());
2420
2421
2422
2423
2424
2425
2426
            }

            float get_gradient_input_element(unsigned long i) const
            {
                if (i < gradient_input.size())
                    return gradient_input.host()[i];
                else
Davis King's avatar
Davis King committed
2427
                    return subnet().get_gradient_input_element(i-gradient_input.size());
2428
2429
2430
2431
2432
            }


        private:
            // We lazily initialize sub-layers as needed when someone tries to call
Davis King's avatar
Davis King committed
2433
            // subnet()
2434
2435
            void init_sub() const
            {
Davis King's avatar
Davis King committed
2436
2437
                if (!subnetwork)
                    subnetwork.reset(new test_layer_subnet(rnd));
2438
2439
2440
            }

            dlib::rand& rnd;
Davis King's avatar
Davis King committed
2441
            mutable std::unique_ptr<test_layer_subnet> subnetwork;
2442
2443
2444
2445
            resizable_tensor output;
            resizable_tensor gradient_input;
        };

2446
    }
2447

2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
    struct layer_test_results
    {
        layer_test_results() : was_good(true) {}
        explicit layer_test_results(const std::string& l) : log(l),was_good(false) {}

        std::string log;
        bool was_good;

        operator bool() const { return was_good; }
    };

    inline std::ostream& operator<< (std::ostream& out, const layer_test_results& item)
    {
        out << item.log;
        return out;
2463
2464
2465
2466
2467
    }

    template <
        typename layer_details_type
        >
Davis King's avatar
Davis King committed
2468
2469
2470
    layer_test_results impl_test_layer (
        layer_details_type l,
        const float base_eps 
2471
2472
2473
2474
    )
    {
        using namespace timpl;
        // Do some setup
2475
        running_stats<double> rs_data, rs_params;
2476
        dlib::rand rnd;
2477
2478
        std::ostringstream sout;
        for (int iter = 0; iter < 10; ++iter)
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
2505
2506
        {
            test_layer_subnet subnetwork(rnd);
            resizable_tensor output, out2, out3;
            // Run setup() and forward() as well to make sure any calls to subnet() have
            // happened before we start assuming we know how many data elements there are
            // (since we do a lazy layer creation thing based on calls to subnet() inside
            // test_layer_subnet).
            l.setup(subnetwork);
            impl::call_layer_forward(l, subnetwork, output);

            resizable_tensor input_grad;
            input_grad.copy_size(output);
            fill_with_gassuan_random_numbers(input_grad, rnd);


            // The f() we are computing gradients of is this thing.  It's value at the current
            // parameter and data values is:
            //sout << "f(data,params): " << dot(output, input_grad) << std::endl;

            // We are going to save a copy of the subnetwork.get_gradient_input() data before we do
            // backpropagation since the backward() function is supposed to *add* to the
            // gradients rather than overwrite them.  We will use this saved data to check if
            // that is the case.
            const unsigned long num_data_inputs = subnetwork.count_outputs();
            std::vector<float> initial_gradient_input(num_data_inputs);
            for (unsigned long i = 0; i < num_data_inputs; ++i)
                initial_gradient_input[i] = subnetwork.get_gradient_input_element(i);

2507

2508
2509
2510
            // Now tell the layer to compute all the gradients.  In the rest of this function
            // we will just be checking that these gradients were computed correctly by
            // comparing them to a central differences approximation.
2511
            resizable_tensor params_grad;
2512
2513
2514
            params_grad.copy_size(l.get_layer_params());
            // But first, set the params grad to something crazy so that it's very obvious if
            // it doesn't get fully assigned.
2515
            params_grad = std::numeric_limits<float>::infinity();
2516
            impl::call_layer_backward(l, output, input_grad, subnetwork, params_grad);
2517

2518
2519
2520
2521
2522
2523
            static_assert(impl::is_inplace_layer(l, subnetwork) == impl::has_inplace_backward(l, subnetwork),
                "Layer not defined correctly.  forward and backward methods must either both be in-place or both out-of-place. ");

            // Make sure the outputs of forward() and backward() are the same when they are run
            // in in-place mode.
            if (impl::is_inplace_layer(l, subnetwork))
2524
            {
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
                test_layer_subnet subnetwork2(rnd);
                layer_details_type ll(l);
                ll.setup(subnetwork2);
                resizable_tensor ip_out;
                impl::call_layer_forward(ll, subnetwork2, ip_out);
                impl::call_layer_forward(ll, subnetwork2, subnetwork2.get_mutable_output());
                const auto forward_error = max(abs(mat(ip_out) - mat(subnetwork2.get_output())));
                if (forward_error > 0.00001)
                {
                    using namespace std;
                    sout << "This layer is supposed to support in-place computations but the output of forward_inplace()\n";
                    sout << "changes when invoked in-place vs. out-of-place. The error was: " << forward_error << endl;
                    return layer_test_results(sout.str()); 
                }

                resizable_tensor params_grad;
                params_grad.copy_size(ll.get_layer_params());
                params_grad = std::numeric_limits<float>::infinity();

                resizable_tensor input_grad;
                input_grad.copy_size(ip_out);
                fill_with_gassuan_random_numbers(input_grad, rnd);
                resizable_tensor params_grad1, params_grad2, data_grad1, data_grad2;
                params_grad1 = params_grad;
                params_grad2 = params_grad;
                // Now call backward() and make sure it works as well.
                subnetwork2.get_gradient_input() = 9999;
                impl::call_layer_backward(ll, ip_out, input_grad, subnetwork2, params_grad1);
                data_grad1 = subnetwork2.get_gradient_input();

                subnetwork2.get_gradient_input() = mat(input_grad);
                impl::call_layer_backward(ll, ip_out, subnetwork2.get_gradient_input(), subnetwork2, params_grad2);
                data_grad2 = subnetwork2.get_gradient_input();
                if (params_grad.size() != 0)
                {
                    const auto backward_param_error = max(abs(mat(params_grad1) - mat(params_grad2)));
                    if (backward_param_error > 0.00001)
                    {
                        using namespace std;
                        sout << "This layer is supposed to support in-place computations but the output of backward_inplace()\n";
                        sout << "changes when invoked in-place vs. out-of-place. The error was: " << backward_param_error << endl;
                        return layer_test_results(sout.str()); 
                    }
                }
                const auto backward_data_error = max(abs(mat(data_grad1) - mat(data_grad2)));
                if (backward_data_error > 0.00001)
2571
2572
2573
                {
                    using namespace std;
                    sout << "This layer is supposed to support in-place computations but the output of backward_inplace()\n";
2574
                    sout << "changes when invoked in-place vs. out-of-place. The error was: " << backward_data_error << endl;
2575
2576
2577
                    return layer_test_results(sout.str()); 
                }
            }
2578

2579
2580
2581
            // ==================================================================
            // first validate the way the parameter gradients are computed
            for (unsigned long i = 0; i < params_grad.size(); ++i)
2582
            {
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
                layer_details_type l1(l);

                float eps = l1.get_layer_params().host()[i]*base_eps;
                if (eps == 0)
                    eps = base_eps;
                const float oldval = l1.get_layer_params().host()[i];
                l1.get_layer_params().host()[i] = oldval+eps;
                impl::call_layer_forward(l1, subnetwork, out2);
                l1.get_layer_params().host()[i] = oldval-eps;
                impl::call_layer_forward(l1, subnetwork, out3);
                l1.get_layer_params().host()[i] = oldval;

                // Compute a reference derivative via a central differences approximation and
                // compare it to the one output by the layer and make sure they match.
                double reference_derivative = (dot(out2,input_grad)-dot(out3, input_grad))/(2*eps);
                double output_derivative = params_grad.host()[i];
2599
2600
2601
2602
2603
                double relative_error;
                if (reference_derivative != 0)
                    relative_error = (reference_derivative - output_derivative)/(reference_derivative);
                else
                    relative_error = (reference_derivative - output_derivative);
2604
                double absolute_error = (reference_derivative - output_derivative);
2605
                rs_params.add(std::abs(relative_error));
Davis King's avatar
Davis King committed
2606
                if (std::abs(relative_error) > 0.05 && std::abs(absolute_error) > 0.006)
2607
2608
2609
2610
2611
                {
                    using namespace std;
                    sout << "Gradient error in parameter #" << i <<".  Relative error: "<< relative_error << endl;
                    sout << "expected derivative: " << reference_derivative << endl;
                    sout << "output derivative:   " << output_derivative << endl;
2612
                    sout << "iteration:           " << iter << endl;
2613
2614
2615
                    return layer_test_results(sout.str()); 
                }
            }
2616

2617
2618
2619
            // ==================================================================
            // now validate the data gradients
            for (unsigned long i = 0; i < num_data_inputs; ++i)
2620
            {
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
                const float oldval = subnetwork.get_output_element(i);
                float eps = oldval*base_eps;
                if (eps == 0)
                    eps = base_eps;
                subnetwork.get_output_element(i) = oldval+eps;
                impl::call_layer_forward(l, subnetwork, out2);
                subnetwork.get_output_element(i) = oldval-eps;
                impl::call_layer_forward(l, subnetwork, out3);
                subnetwork.get_output_element(i) = oldval;

                // Compute a reference derivative via a central differences approximation and
                // compare it to the one output by the layer and make sure they match.
                double reference_derivative = (dot(out2,input_grad)-dot(out3, input_grad))/(2*eps);
                double output_derivative = subnetwork.get_gradient_input_element(i);
                if (!impl::is_inplace_layer(l,subnetwork))
                    output_derivative -= initial_gradient_input[i];
2637
2638
2639
2640
2641
                double relative_error;
                if (reference_derivative != 0)
                    relative_error = (reference_derivative - output_derivative)/(reference_derivative);
                else
                    relative_error = (reference_derivative - output_derivative);
2642
                double absolute_error = (reference_derivative - output_derivative);
2643
                rs_data.add(std::abs(relative_error));
Davis King's avatar
Davis King committed
2644
                if (std::abs(relative_error) > 0.05 && std::abs(absolute_error) > 0.006)
2645
2646
2647
2648
2649
                {
                    using namespace std;
                    sout << "Gradient error in data variable #" << i <<".  Relative error: "<< relative_error << endl;
                    sout << "expected derivative: " << reference_derivative << endl;
                    sout << "output derivative:   " << output_derivative << endl;
2650
                    sout << "iteration:           " << iter << endl;
2651
2652
                    return layer_test_results(sout.str()); 
                }
2653
            }
2654
2655

        } // end for (int iter = 0; iter < 5; ++iter)
2656

2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
        if (rs_params.mean() > 0.003)
        {
            using namespace std;
            sout << "Average parameter gradient error is somewhat large at: "<< rs_params.mean() << endl;
            return layer_test_results(sout.str()); 
        }
        if (rs_data.mean() > 0.003)
        {
            using namespace std;
            sout << "Average data gradient error is somewhat large at: "<< rs_data.mean() << endl;
            return layer_test_results(sout.str()); 
        }

2670
        return layer_test_results();
2671
2672
    }

Davis King's avatar
Davis King committed
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
2686
2687
2688
2689
2690
2691
    template <
        typename layer_details_type
        >
    layer_test_results test_layer (
        layer_details_type l
    )
    {
        // Try a few different derivative step sizes to see if any work. 
        for (float base_eps = 0.0001; base_eps < 0.1; base_eps *= 2)
        {
            auto result = impl_test_layer(l, base_eps);
            if (result)
                return result;
        }
        // However, if none of the step sizes worked then try this one and probably result
        // in returning an error.
        return impl_test_layer(l, 0.01);
    }

2692
2693
2694
2695
// ----------------------------------------------------------------------------------------

}

2696
#endif // DLIB_DNn_CORE_H_
2697
2698