File size: 24,978 Bytes
9375c9a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
// Copyright (C) 2015  Davis E. King ([email protected])
// License: Boost Software License   See LICENSE.txt for the full license.
#ifndef DLIB_DNN_CPU_H_
#define DLIB_DNN_CPU_H_

// This file contains CPU implementations of the GPU based functions in cuda_dlib.h
// and cudnn_dlibapi.h

#include "tensor.h"
#include "../geometry/rectangle.h"
#include "../dnn/misc.h"

namespace dlib
{
    namespace cpu 
    {

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

        void multiply (
            bool add_to,
            tensor& dest,
            const tensor& src1,
            const tensor& src2
        );

        void multiply_conv (
            bool add_to,
            tensor& dest,
            const tensor& src1,
            const tensor& src2
        );

        void multiply_zero_padded (
            bool add_to,
            tensor& dest,
            const tensor& src1,
            const tensor& src2
        );

        void scale_channels (
            bool add_to,
            tensor& dest,
            const tensor& src,
            const tensor& scales
        );

        void add(
            float beta,
            tensor& dest,
            float alpha,
            const tensor& src
        );

        void assign_bias_gradient (
            tensor& grad,
            const tensor& gradient_input
        );

        void add (
            tensor& dest,
            const tensor& src1,
            const tensor& src2
        );

        void assign_conv_bias_gradient (
            tensor& grad,
            const tensor& gradient_input
        );

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

        void affine_transform(
            tensor& dest,
            const tensor& src,
            const float A,
            const float B
        );

        void affine_transform(
            tensor& dest,
            const tensor& src1,
            const tensor& src2,
            const float A,
            const float B,
            const float C
        );

        void affine_transform(
            tensor& dest,
            const tensor& src1,
            const tensor& src2,
            const tensor& src3,
            const float A,
            const float B,
            const float C,
            const float D
        );

        void affine_transform_range(
            size_t begin,
            size_t end,
            tensor& dest,
            const tensor& src1,
            const tensor& src2,
            const tensor& src3,
            const float A,
            const float B,
            const float C
        );

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

        void affine_transform(
            tensor& dest,
            const tensor& src,
            const tensor& A,
            const tensor& B
        );

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

        void affine_transform_conv(
            tensor& dest,
            const tensor& src,
            const tensor& A,
            const tensor& B
        );

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

        void affine_transform(
            const rectangle& rect,
            tensor& dest, 
            const tensor& src1, 
            const tensor& src2, 
            const tensor& src3, 
            float A, 
            float B,
            float C
        );

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

        void compute_adam_update (
            size_t begin,
            size_t end,
            tensor& s,
            tensor& m,
            tensor& v,
            const float t,
            const float learning_rate,
            const float weight_decay,
            const float momentum1,
            const float momentum2,
            const tensor& params,
            const tensor& params_grad
        );

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

        void batch_normalize_inference (
            const double eps,
            resizable_tensor& dest,
            const tensor& src,
            const tensor& gamma, 
            const tensor& beta,
            const tensor& running_means,
            const tensor& running_variances
        );

        void batch_normalize (
            const double eps,
            resizable_tensor& dest,
            resizable_tensor& means,
            resizable_tensor& invstds,
            const double averaging_factor,
            resizable_tensor& running_means,
            resizable_tensor& running_variances,
            const tensor& src,
            const tensor& gamma, 
            const tensor& beta 
        );

        void batch_normalize_gradient (
            const double eps,
            const tensor& gradient_input,
            const tensor& means,
            const tensor& invstds,
            const tensor& src,
            const tensor& gamma,
            tensor& src_grad,
            tensor& gamma_grad, 
            tensor& beta_grad 
        );

        void batch_normalize_conv_inference (
            const double eps,
            resizable_tensor& dest,
            const tensor& src,
            const tensor& gamma, 
            const tensor& beta,
            const tensor& running_means,
            const tensor& running_variances
        );

        void batch_normalize_conv (
            const double eps,
            resizable_tensor& dest,
            resizable_tensor& means,
            resizable_tensor& invstds,
            const double averaging_factor,
            resizable_tensor& running_means,
            resizable_tensor& running_variances,
            const tensor& src,
            const tensor& gamma, 
            const tensor& beta 
        );

        void batch_normalize_conv_gradient (
            const double eps,
            const tensor& gradient_input,
            const tensor& means,
            const tensor& invstds,
            const tensor& src,
            const tensor& gamma,
            tensor& src_grad,
            tensor& gamma_grad, 
            tensor& beta_grad 
        );

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

        void layer_normalize (
            const double eps,
            resizable_tensor& dest,
            resizable_tensor& means,
            resizable_tensor& invstds,
            const tensor& src,
            const tensor& gamma,
            const tensor& beta
        );

        void layer_normalize_gradient (
            const double eps,
            const tensor& gradient_input,
            const tensor& means,
            const tensor& invstds,
            const tensor& src,
            const tensor& gamma,
            tensor& src_grad,
            tensor& gamma_grad,
            tensor& beta_grad
        );

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

        void threshold (
            tensor& data,
            float thresh
        );

        void dot (
            const tensor& a,
            const tensor& b,
            tensor& result,
            size_t idx
        );

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

        void softmax (
            tensor& dest,
            const tensor& src
        );

        void softmax_gradient (
            tensor& grad,
            const tensor& dest,
            const tensor& gradient_input
        );

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

        void softmax_all (
            tensor& dest,
            const tensor& src
        );

        void softmax_all_gradient (
            tensor& grad,
            const tensor& dest,
            const tensor& gradient_input
        );

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

        void sigmoid (
            tensor& dest,
            const tensor& src
        );

        void sigmoid_gradient (
            tensor& grad,
            const tensor& dest,
            const tensor& gradient_input
        );

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

        void mish (
            tensor& dest,
            const tensor& src
        );

        void mish_gradient (
            tensor& grad,
            const tensor& src,
            const tensor& gradient_input
        );

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

        void relu (
            tensor& dest,
            const tensor& src
        );

        void relu_gradient (
            tensor& grad,
            const tensor& dest,
            const tensor& gradient_input
        );

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

        void prelu (
            tensor& dest,
            const tensor& src,
            const tensor& param
        );

        void prelu_gradient (
            tensor& grad,
            const tensor& src,
            const tensor& gradient_input,
            const tensor& param,
            tensor& params_grad 
        );

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

        void leaky_relu (
            tensor& dest,
            const tensor& src,
            const float alpha
        );

        void leaky_relu_gradient (
            tensor& grad,
            const tensor& dest,
            const tensor& gradient_input,
            const float alpha
        );

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

        void tanh (
            tensor& dest,
            const tensor& src
        );

        void tanh_gradient (
            tensor& grad,
            const tensor& dest,
            const tensor& gradient_input
        );

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

        void gelu (
            tensor& dest,
            const tensor& src
        );

        void gelu_gradient (
            tensor& grad,
            const tensor& dest,
            const tensor& gradient_input
        );

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

        void resize_bilinear (
            tensor& dest,
            long dest_row_stride,
            long dest_channel_stride,
            const tensor& src,
            long src_row_stride,
            long src_channel_stride
        );

        void resize_bilinear_gradient (
            tensor& grad,
            long grad_row_stride,
            long grad_channel_stride,
            const tensor& gradient_input,
            long gradient_input_row_stride,
            long gradient_input_channel_stride
        );

        inline void resize_bilinear (
            tensor& dest,
            const tensor& src
        ) { resize_bilinear(dest, dest.nc(), dest.nr()*dest.nc(), src, src.nc(), src.nr()*src.nc()); }

        inline void resize_bilinear_gradient (
            tensor& grad,
            const tensor& gradient_input
        ) { resize_bilinear_gradient(grad, grad.nc(), grad.nr()*grad.nc(), gradient_input, gradient_input.nc(), gradient_input.nr()*gradient_input.nc()); }

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

        class pooling
        {
        public:

            pooling(const pooling&) = delete;
            pooling& operator=(const pooling&) = delete;

            pooling (
            );

            void clear(
            );

            void setup_max_pooling(
                int window_height,
                int window_width,
                int stride_y,
                int stride_x,
                int padding_y,
                int padding_x
            );

            void setup_avg_pooling(
                int window_height,
                int window_width,
                int stride_y,
                int stride_x,
                int padding_y,
                int padding_x
            );

            bool does_max_pooling(
            ) const { return do_max_pooling; }

            void operator() (
                resizable_tensor& dest,
                const tensor& src
            );

            void get_gradient(
                const tensor& gradient_input, 
                const tensor& dest,
                const tensor& src,
                tensor& grad 
            );

        private:
            int window_height;
            int window_width;
            int stride_y;
            int stride_x;
            int padding_y;
            int padding_x;
            bool do_max_pooling;

        };

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

        class tensor_conv
        {
        public:
            tensor_conv(const tensor_conv&) = delete;
            tensor_conv& operator=(const tensor_conv&) = delete;

            tensor_conv() {}

            void clear(
            ) {}

            void setup(
                const tensor& data,    /* not used but required for interface */
                const tensor& filters, /* not used but required for interface */
                int stride_y,
                int stride_x,
                int padding_y,
                int padding_x
            ) 
            {
                (void)data;    /* silence compiler */
                DLIB_CASSERT(stride_y > 0 && stride_x > 0);
                DLIB_CASSERT(0 <= padding_y && padding_y < filters.nr());
                DLIB_CASSERT(0 <= padding_x && padding_x < filters.nc());
                last_stride_y = stride_y;
                last_stride_x = stride_x;
                last_padding_y = padding_y;
                last_padding_x = padding_x;            
            }

             void operator() (
                const bool add_to_output,
                resizable_tensor& output,
                const tensor& data,
                const tensor& filters
            );

             void operator() (
                const bool add_to_output,
                tensor& output,
                const tensor& data,
                const tensor& filters
            );

            void get_gradient_for_data (
                const bool add_to_output,
                const tensor& gradient_input, 
                const tensor& filters,
                tensor& data_gradient
            );

            void get_gradient_for_filters (
                const bool add_to_output,
                const tensor& gradient_input, 
                const tensor& data,
                tensor& filters_gradient
            );

        private:

            long last_stride_y = 0;
            long last_stride_x = 0;
            long last_padding_y = 0;
            long last_padding_x = 0;
        };

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

        void copy_tensor(
            bool add_to,
            tensor& dest,
            size_t dest_k_offset,
            const tensor& src,
            size_t src_k_offset,
            size_t count_k
        );

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

    class compute_loss_binary_log_per_pixel
    {

        /*! The point of this class is to compute the loss for loss_binary_log_per_pixel_
            on the cpu to provide an analogous implementation of the cuda version
        !*/
    public:
        compute_loss_binary_log_per_pixel(
        )
        {
        }

        template <
            typename const_label_iterator
            >
        void operator()(
            const_label_iterator truth,
            const tensor& output_tensor,
            tensor& grad,
            double& loss
        ) const
        {
            sigmoid(grad, output_tensor);
            // The loss we output is the average loss over the mini-batch, and also over each element of the matrix output.
            const double scale = 1.0/(output_tensor.num_samples()*output_tensor.nr()*output_tensor.nc());
            loss = 0;
            float* const g = grad.host();
            const float* const out_data = output_tensor.host();
            for (long i = 0; i < output_tensor.num_samples(); ++i, ++truth)
            {
                for (long r = 0; r < output_tensor.nr(); ++r)
                {
                    for (long c = 0; c < output_tensor.nc(); ++c)
                    {
                        const float y = truth->operator()(r, c);
                        const size_t idx = tensor_index(output_tensor, i, 0, r, c);

                        if (y > 0.f)
                        {
                            const float temp = log1pexp(-out_data[idx]);
                            loss += y*scale*temp;
                            g[idx] = y*scale*(g[idx]-1);
                        }
                        else if (y < 0.f)
                        {
                            const float temp = -(-out_data[idx]-log1pexp(-out_data[idx]));
                            loss += -y*scale*temp;
                            g[idx] = -y*scale*g[idx];
                        }
                        else
                        {
                            g[idx] = 0.f;
                        }
                    }
                }
            }
        }
    };

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

    class compute_loss_multiclass_log_per_pixel
    {

        /*! The point of this class is to compute the loss for loss_multiclass_log_per_pixel_
            on the cpu to provide an analogous implementation of the cuda version
        !*/
    public:
        compute_loss_multiclass_log_per_pixel(
        )
        {
        }

        template <
            typename const_label_iterator
            >
        void operator()(
            const_label_iterator truth,
            const tensor& output_tensor,
            tensor& grad,
            double& loss
        ) const
        {
            softmax(grad, output_tensor);
            // The loss we output is the average loss over the mini-batch, and also over each element of the matrix output.
            const double scale = 1.0 / (output_tensor.num_samples() * output_tensor.nr() * output_tensor.nc());
            loss = 0;
            float* const g = grad.host();
            for (long i = 0; i < output_tensor.num_samples(); ++i, ++truth)
            {
                for (long r = 0; r < output_tensor.nr(); ++r)
                {
                    for (long c = 0; c < output_tensor.nc(); ++c)
                    {
                        const uint16_t y = truth->operator()(r, c);
                        // The network must produce a number of outputs that is equal to the number
                        // of labels when using this type of loss.
                        DLIB_CASSERT(static_cast<long>(y) < output_tensor.k() || y == label_to_ignore,
                                        "y: " << y << ", output_tensor.k(): " << output_tensor.k());
                        for (long k = 0; k < output_tensor.k(); ++k)
                        {
                            const size_t idx = tensor_index(output_tensor, i, k, r, c);
                            if (k == y)
                            {
                                loss += scale*-safe_log(g[idx]);
                                g[idx] = scale*(g[idx] - 1);
                            }
                            else if (y == label_to_ignore)
                            {
                                g[idx] = 0.f;
                            }
                            else
                            {
                                g[idx] = scale*g[idx];
                            }
                        }
                    }
                }
            }
        }
    private:
        static const uint16_t label_to_ignore = std::numeric_limits<uint16_t>::max();
    };

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

    class compute_loss_multiclass_log_per_pixel_weighted
    {

        /*! The point of this class is to compute the loss for loss_multiclass_log_per_pixel_weighted_
            on the cpu to provide an analogous implementation of the cuda version
        !*/
    public:
        compute_loss_multiclass_log_per_pixel_weighted(
        )
        {
        }

        template <
            typename const_label_iterator
            >
        void operator()(
            const_label_iterator truth,
            const tensor& output_tensor,
            tensor& grad,
            double& loss
        ) const
        {
            softmax(grad, output_tensor);
            // The loss we output is the weighted average loss over the mini-batch, and also over each element of the matrix output.
            const double scale = 1.0 / (output_tensor.num_samples() * output_tensor.nr() * output_tensor.nc());
            loss = 0;
            float* const g = grad.host();
            for (long i = 0; i < output_tensor.num_samples(); ++i, ++truth)
            {
                for (long r = 0; r < output_tensor.nr(); ++r)
                {
                    for (long c = 0; c < output_tensor.nc(); ++c)
                    {
                        const weighted_label<uint16_t>& weighted_label = truth->operator()(r, c);
                        const uint16_t y = weighted_label.label;
                        const float weight = weighted_label.weight;
                        // The network must produce a number of outputs that is equal to the number
                        // of labels when using this type of loss.
                        DLIB_CASSERT(static_cast<long>(y) < output_tensor.k() || weight == 0.f,
                                        "y: " << y << ", output_tensor.k(): " << output_tensor.k());
                        for (long k = 0; k < output_tensor.k(); ++k)
                        {
                            const size_t idx = tensor_index(output_tensor, i, k, r, c);
                            if (k == y)
                            {
                                loss += weight*scale*-safe_log(g[idx]);
                                g[idx] = weight*scale*(g[idx] - 1);
                            }
                            else
                            {
                                g[idx] = weight*scale*g[idx];
                            }
                        }
                    }
                }
            }
        }

    };

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

    class compute_loss_mean_squared_per_channel_and_pixel
    {
        /*! The point of this class is to compute the loss for loss_mean_squared_per_channel_and_pixel_
            on the cpu to provide an analogous implementation of the cuda version
        !*/
        public:

            compute_loss_mean_squared_per_channel_and_pixel(
            )
            {
            }

        template <
            typename const_label_iterator
            >
        void operator()(
            const_label_iterator truth,
            const tensor& output_tensor,
            tensor& grad,
            double& loss
        ) const
        {
            // The loss we output is the average loss over the mini-batch, and also over each element of the matrix output.
            const double scale = 1.0 / (output_tensor.num_samples() * output_tensor.k() * output_tensor.nr() * output_tensor.nc());
            loss = 0;
            float* const g = grad.host();
            const float* out_data = output_tensor.host();
            for (long i = 0; i < output_tensor.num_samples(); ++i, ++truth)
            {
                for (long k = 0; k < output_tensor.k(); ++k)
                {
                    for (long r = 0; r < output_tensor.nr(); ++r)
                    {
                        for (long c = 0; c < output_tensor.nc(); ++c)
                        {
                            const float y = (*truth)[k].operator()(r, c);
                            const size_t idx = tensor_index(output_tensor, i, k, r, c);
                            const float temp1 = y - out_data[idx];
                            const float temp2 = scale*temp1;
                            loss += temp2*temp1;
                            g[idx] = -temp2;
                        }
                    }
                }
            }
        }

    };

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

    } 
}

#ifdef NO_MAKEFILE
#include "cpu_dlib.cpp"
#endif

#endif // DLIB_DNN_CPU_H_