• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*M///////////////////////////////////////////////////////////////////////////////////////
2 //
3 //  IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
4 //
5 //  By downloading, copying, installing or using the software you agree to this license.
6 //  If you do not agree to this license, do not download, install,
7 //  copy or use the software.
8 //
9 //
10 //                        Intel License Agreement
11 //
12 // Copyright (C) 2000, Intel Corporation, all rights reserved.
13 // Third party copyrights are property of their respective owners.
14 //
15 // Redistribution and use in source and binary forms, with or without modification,
16 // are permitted provided that the following conditions are met:
17 //
18 //   * Redistribution's of source code must retain the above copyright notice,
19 //     this list of conditions and the following disclaimer.
20 //
21 //   * Redistribution's in binary form must reproduce the above copyright notice,
22 //     this list of conditions and the following disclaimer in the documentation
23 //     and/or other materials provided with the distribution.
24 //
25 //   * The name of Intel Corporation may not be used to endorse or promote products
26 //     derived from this software without specific prior written permission.
27 //
28 // This software is provided by the copyright holders and contributors "as is" and
29 // any express or implied warranties, including, but not limited to, the implied
30 // warranties of merchantability and fitness for a particular purpose are disclaimed.
31 // In no event shall the Intel Corporation or contributors be liable for any direct,
32 // indirect, incidental, special, exemplary, or consequential damages
33 // (including, but not limited to, procurement of substitute goods or services;
34 // loss of use, data, or profits; or business interruption) however caused
35 // and on any theory of liability, whether in contract, strict liability,
36 // or tort (including negligence or otherwise) arising in any way out of
37 // the use of this software, even if advised of the possibility of such damage.
38 //
39 //M*/
40 
41 #include "_ml.h"
42 
CvANN_MLP_TrainParams()43 CvANN_MLP_TrainParams::CvANN_MLP_TrainParams()
44 {
45     term_crit = cvTermCriteria( CV_TERMCRIT_ITER + CV_TERMCRIT_EPS, 1000, 0.01 );
46     train_method = RPROP;
47     bp_dw_scale = bp_moment_scale = 0.1;
48     rp_dw0 = 0.1; rp_dw_plus = 1.2; rp_dw_minus = 0.5;
49     rp_dw_min = FLT_EPSILON; rp_dw_max = 50.;
50 }
51 
52 
CvANN_MLP_TrainParams(CvTermCriteria _term_crit,int _train_method,double _param1,double _param2)53 CvANN_MLP_TrainParams::CvANN_MLP_TrainParams( CvTermCriteria _term_crit,
54                                               int _train_method,
55                                               double _param1, double _param2 )
56 {
57     term_crit = _term_crit;
58     train_method = _train_method;
59     bp_dw_scale = bp_moment_scale = 0.1;
60     rp_dw0 = 1.; rp_dw_plus = 1.2; rp_dw_minus = 0.5;
61     rp_dw_min = FLT_EPSILON; rp_dw_max = 50.;
62 
63     if( train_method == RPROP )
64     {
65         rp_dw0 = _param1;
66         if( rp_dw0 < FLT_EPSILON )
67             rp_dw0 = 1.;
68         rp_dw_min = _param2;
69         rp_dw_min = MAX( rp_dw_min, 0 );
70     }
71     else if( train_method == BACKPROP )
72     {
73         bp_dw_scale = _param1;
74         if( bp_dw_scale <= 0 )
75             bp_dw_scale = 0.1;
76         bp_dw_scale = MAX( bp_dw_scale, 1e-3 );
77         bp_dw_scale = MIN( bp_dw_scale, 1 );
78         bp_moment_scale = _param2;
79         if( bp_moment_scale < 0 )
80             bp_moment_scale = 0.1;
81         bp_moment_scale = MIN( bp_moment_scale, 1 );
82     }
83     else
84         train_method = RPROP;
85 }
86 
87 
~CvANN_MLP_TrainParams()88 CvANN_MLP_TrainParams::~CvANN_MLP_TrainParams()
89 {
90 }
91 
92 
CvANN_MLP()93 CvANN_MLP::CvANN_MLP()
94 {
95     layer_sizes = wbuf = 0;
96     min_val = max_val = min_val1 = max_val1 = 0.;
97     weights = 0;
98     rng = cvRNG(-1);
99     default_model_name = "my_nn";
100     clear();
101 }
102 
103 
CvANN_MLP(const CvMat * _layer_sizes,int _activ_func,double _f_param1,double _f_param2)104 CvANN_MLP::CvANN_MLP( const CvMat* _layer_sizes,
105                       int _activ_func,
106                       double _f_param1, double _f_param2 )
107 {
108     layer_sizes = wbuf = 0;
109     min_val = max_val = min_val1 = max_val1 = 0.;
110     weights = 0;
111     rng = cvRNG(-1);
112     default_model_name = "my_nn";
113     create( _layer_sizes, _activ_func, _f_param1, _f_param2 );
114 }
115 
116 
~CvANN_MLP()117 CvANN_MLP::~CvANN_MLP()
118 {
119     clear();
120 }
121 
122 
clear()123 void CvANN_MLP::clear()
124 {
125     cvReleaseMat( &layer_sizes );
126     cvReleaseMat( &wbuf );
127     cvFree( &weights );
128     activ_func = SIGMOID_SYM;
129     f_param1 = f_param2 = 1;
130     max_buf_sz = 1 << 12;
131 }
132 
133 
set_activ_func(int _activ_func,double _f_param1,double _f_param2)134 void CvANN_MLP::set_activ_func( int _activ_func, double _f_param1, double _f_param2 )
135 {
136     CV_FUNCNAME( "CvANN_MLP::set_activ_func" );
137 
138     __BEGIN__;
139 
140     if( _activ_func < 0 || _activ_func > GAUSSIAN )
141         CV_ERROR( CV_StsOutOfRange, "Unknown activation function" );
142 
143     activ_func = _activ_func;
144 
145     switch( activ_func )
146     {
147     case SIGMOID_SYM:
148         max_val = 0.95; min_val = -max_val;
149         max_val1 = 0.98; min_val1 = -max_val1;
150         if( fabs(_f_param1) < FLT_EPSILON )
151             _f_param1 = 2./3;
152         if( fabs(_f_param2) < FLT_EPSILON )
153             _f_param2 = 1.7159;
154         break;
155     case GAUSSIAN:
156         max_val = 1.; min_val = 0.05;
157         max_val1 = 1.; min_val1 = 0.02;
158         if( fabs(_f_param1) < FLT_EPSILON )
159             _f_param1 = 1.;
160         if( fabs(_f_param2) < FLT_EPSILON )
161             _f_param2 = 1.;
162         break;
163     default:
164         min_val = max_val = min_val1 = max_val1 = 0.;
165         _f_param1 = 1.;
166         _f_param2 = 0.;
167     }
168 
169     f_param1 = _f_param1;
170     f_param2 = _f_param2;
171 
172     __END__;
173 }
174 
175 
init_weights()176 void CvANN_MLP::init_weights()
177 {
178     int i, j, k;
179 
180     for( i = 1; i < layer_sizes->cols; i++ )
181     {
182         int n1 = layer_sizes->data.i[i-1];
183         int n2 = layer_sizes->data.i[i];
184         double val = 0, G = n2 > 2 ? 0.7*pow((double)n1,1./(n2-1)) : 1.;
185         double* w = weights[i];
186 
187         // initialize weights using Nguyen-Widrow algorithm
188         for( j = 0; j < n2; j++ )
189         {
190             double s = 0;
191             for( k = 0; k <= n1; k++ )
192             {
193                 val = cvRandReal(&rng)*2-1.;
194                 w[k*n2 + j] = val;
195                 s += val;
196             }
197 
198             if( i < layer_sizes->cols - 1 )
199             {
200                 s = 1./(s - val);
201                 for( k = 0; k <= n1; k++ )
202                     w[k*n2 + j] *= s;
203                 w[n1*n2 + j] *= G*(-1+j*2./n2);
204             }
205         }
206     }
207 }
208 
209 
create(const CvMat * _layer_sizes,int _activ_func,double _f_param1,double _f_param2)210 void CvANN_MLP::create( const CvMat* _layer_sizes, int _activ_func,
211                         double _f_param1, double _f_param2 )
212 {
213     CV_FUNCNAME( "CvANN_MLP::create" );
214 
215     __BEGIN__;
216 
217     int i, l_step, l_count, buf_sz = 0;
218     int *l_src, *l_dst;
219 
220     clear();
221 
222     if( !CV_IS_MAT(_layer_sizes) ||
223         _layer_sizes->cols != 1 && _layer_sizes->rows != 1 ||
224         CV_MAT_TYPE(_layer_sizes->type) != CV_32SC1 )
225         CV_ERROR( CV_StsBadArg,
226         "The array of layer neuron counters must be an integer vector" );
227 
228     CV_CALL( set_activ_func( _activ_func, _f_param1, _f_param2 ));
229 
230     l_count = _layer_sizes->rows + _layer_sizes->cols - 1;
231     l_src = _layer_sizes->data.i;
232     l_step = CV_IS_MAT_CONT(_layer_sizes->type) ? 1 :
233                 _layer_sizes->step / sizeof(l_src[0]);
234     CV_CALL( layer_sizes = cvCreateMat( 1, l_count, CV_32SC1 ));
235     l_dst = layer_sizes->data.i;
236 
237     max_count = 0;
238     for( i = 0; i < l_count; i++ )
239     {
240         int n = l_src[i*l_step];
241         if( n < 1 + (0 < i && i < l_count-1))
242             CV_ERROR( CV_StsOutOfRange,
243             "there should be at least one input and one output "
244             "and every hidden layer must have more than 1 neuron" );
245         l_dst[i] = n;
246         max_count = MAX( max_count, n );
247         if( i > 0 )
248             buf_sz += (l_dst[i-1]+1)*n;
249     }
250 
251     buf_sz += (l_dst[0] + l_dst[l_count-1]*2)*2;
252 
253     CV_CALL( wbuf = cvCreateMat( 1, buf_sz, CV_64F ));
254     CV_CALL( weights = (double**)cvAlloc( (l_count+1)*sizeof(weights[0]) ));
255 
256     weights[0] = wbuf->data.db;
257     weights[1] = weights[0] + l_dst[0]*2;
258     for( i = 1; i < l_count; i++ )
259         weights[i+1] = weights[i] + (l_dst[i-1] + 1)*l_dst[i];
260     weights[l_count+1] = weights[l_count] + l_dst[l_count-1]*2;
261 
262     __END__;
263 }
264 
265 
predict(const CvMat * _inputs,CvMat * _outputs) const266 float CvANN_MLP::predict( const CvMat* _inputs, CvMat* _outputs ) const
267 {
268     CV_FUNCNAME( "CvANN_MLP::predict" );
269 
270     __BEGIN__;
271 
272     double* buf;
273     int i, j, n, dn = 0, l_count, dn0, buf_sz, min_buf_sz;
274 
275     if( !layer_sizes )
276         CV_ERROR( CV_StsError, "The network has not been initialized" );
277 
278     if( !CV_IS_MAT(_inputs) || !CV_IS_MAT(_outputs) ||
279         !CV_ARE_TYPES_EQ(_inputs,_outputs) ||
280         CV_MAT_TYPE(_inputs->type) != CV_32FC1 &&
281         CV_MAT_TYPE(_inputs->type) != CV_64FC1 ||
282         _inputs->rows != _outputs->rows )
283         CV_ERROR( CV_StsBadArg, "Both input and output must be floating-point matrices "
284                                 "of the same type and have the same number of rows" );
285 
286     if( _inputs->cols != layer_sizes->data.i[0] )
287         CV_ERROR( CV_StsBadSize, "input matrix must have the same number of columns as "
288                                  "the number of neurons in the input layer" );
289 
290     if( _outputs->cols != layer_sizes->data.i[layer_sizes->cols - 1] )
291         CV_ERROR( CV_StsBadSize, "output matrix must have the same number of columns as "
292                                  "the number of neurons in the output layer" );
293     n = dn0 = _inputs->rows;
294     min_buf_sz = 2*max_count;
295     buf_sz = n*min_buf_sz;
296 
297     if( buf_sz > max_buf_sz )
298     {
299         dn0 = max_buf_sz/min_buf_sz;
300         dn0 = MAX( dn0, 1 );
301         buf_sz = dn0*min_buf_sz;
302     }
303 
304     buf = (double*)cvStackAlloc( buf_sz*sizeof(buf[0]) );
305     l_count = layer_sizes->cols;
306 
307     for( i = 0; i < n; i += dn )
308     {
309         CvMat hdr[2], _w, *layer_in = &hdr[0], *layer_out = &hdr[1], *temp;
310         dn = MIN( dn0, n - i );
311 
312         cvGetRows( _inputs, layer_in, i, i + dn );
313         cvInitMatHeader( layer_out, dn, layer_in->cols, CV_64F, buf );
314 
315         scale_input( layer_in, layer_out );
316         CV_SWAP( layer_in, layer_out, temp );
317 
318         for( j = 1; j < l_count; j++ )
319         {
320             double* data = buf + (j&1 ? max_count*dn0 : 0);
321             int cols = layer_sizes->data.i[j];
322 
323             cvInitMatHeader( layer_out, dn, cols, CV_64F, data );
324             cvInitMatHeader( &_w, layer_in->cols, layer_out->cols, CV_64F, weights[j] );
325             cvGEMM( layer_in, &_w, 1, 0, 0, layer_out );
326             calc_activ_func( layer_out, _w.data.db + _w.rows*_w.cols );
327 
328             CV_SWAP( layer_in, layer_out, temp );
329         }
330 
331         cvGetRows( _outputs, layer_out, i, i + dn );
332         scale_output( layer_in, layer_out );
333     }
334 
335     __END__;
336 
337     return 0.f;
338 }
339 
340 
scale_input(const CvMat * _src,CvMat * _dst) const341 void CvANN_MLP::scale_input( const CvMat* _src, CvMat* _dst ) const
342 {
343     int i, j, cols = _src->cols;
344     double* dst = _dst->data.db;
345     const double* w = weights[0];
346     int step = _src->step;
347 
348     if( CV_MAT_TYPE( _src->type ) == CV_32F )
349     {
350         const float* src = _src->data.fl;
351         step /= sizeof(src[0]);
352 
353         for( i = 0; i < _src->rows; i++, src += step, dst += cols )
354             for( j = 0; j < cols; j++ )
355                 dst[j] = src[j]*w[j*2] + w[j*2+1];
356     }
357     else
358     {
359         const double* src = _src->data.db;
360         step /= sizeof(src[0]);
361 
362         for( i = 0; i < _src->rows; i++, src += step, dst += cols )
363             for( j = 0; j < cols; j++ )
364                 dst[j] = src[j]*w[j*2] + w[j*2+1];
365     }
366 }
367 
368 
scale_output(const CvMat * _src,CvMat * _dst) const369 void CvANN_MLP::scale_output( const CvMat* _src, CvMat* _dst ) const
370 {
371     int i, j, cols = _src->cols;
372     const double* src = _src->data.db;
373     const double* w = weights[layer_sizes->cols];
374     int step = _dst->step;
375 
376     if( CV_MAT_TYPE( _dst->type ) == CV_32F )
377     {
378         float* dst = _dst->data.fl;
379         step /= sizeof(dst[0]);
380 
381         for( i = 0; i < _src->rows; i++, src += cols, dst += step )
382             for( j = 0; j < cols; j++ )
383                 dst[j] = (float)(src[j]*w[j*2] + w[j*2+1]);
384     }
385     else
386     {
387         double* dst = _dst->data.db;
388         step /= sizeof(dst[0]);
389 
390         for( i = 0; i < _src->rows; i++, src += cols, dst += step )
391             for( j = 0; j < cols; j++ )
392                 dst[j] = src[j]*w[j*2] + w[j*2+1];
393     }
394 }
395 
396 
calc_activ_func(CvMat * sums,const double * bias) const397 void CvANN_MLP::calc_activ_func( CvMat* sums, const double* bias ) const
398 {
399     int i, j, n = sums->rows, cols = sums->cols;
400     double* data = sums->data.db;
401     double scale = 0, scale2 = f_param2;
402 
403     switch( activ_func )
404     {
405     case IDENTITY:
406         scale = 1.;
407         break;
408     case SIGMOID_SYM:
409         scale = -f_param1;
410         break;
411     case GAUSSIAN:
412         scale = -f_param1*f_param1;
413         break;
414     default:
415         ;
416     }
417 
418     assert( CV_IS_MAT_CONT(sums->type) );
419 
420     if( activ_func != GAUSSIAN )
421     {
422         for( i = 0; i < n; i++, data += cols )
423             for( j = 0; j < cols; j++ )
424                 data[j] = (data[j] + bias[j])*scale;
425 
426         if( activ_func == IDENTITY )
427             return;
428     }
429     else
430     {
431         for( i = 0; i < n; i++, data += cols )
432             for( j = 0; j < cols; j++ )
433             {
434                 double t = data[j] + bias[j];
435                 data[j] = t*t*scale;
436             }
437     }
438 
439     cvExp( sums, sums );
440 
441     n *= cols;
442     data -= n;
443 
444     switch( activ_func )
445     {
446     case SIGMOID_SYM:
447         for( i = 0; i <= n - 4; i += 4 )
448         {
449             double x0 = 1.+data[i], x1 = 1.+data[i+1], x2 = 1.+data[i+2], x3 = 1.+data[i+3];
450             double a = x0*x1, b = x2*x3, d = scale2/(a*b), t0, t1;
451             a *= d; b *= d;
452             t0 = (2 - x0)*b*x1; t1 = (2 - x1)*b*x0;
453             data[i] = t0; data[i+1] = t1;
454             t0 = (2 - x2)*a*x3; t1 = (2 - x3)*a*x2;
455             data[i+2] = t0; data[i+3] = t1;
456         }
457 
458         for( ; i < n; i++ )
459         {
460             double t = scale2*(1. - data[i])/(1. + data[i]);
461             data[i] = t;
462         }
463         break;
464 
465     case GAUSSIAN:
466         for( i = 0; i < n; i++ )
467             data[i] = scale2*data[i];
468         break;
469 
470     default:
471         ;
472     }
473 }
474 
475 
calc_activ_func_deriv(CvMat * _xf,CvMat * _df,const double * bias) const476 void CvANN_MLP::calc_activ_func_deriv( CvMat* _xf, CvMat* _df,
477                                        const double* bias ) const
478 {
479     int i, j, n = _xf->rows, cols = _xf->cols;
480     double* xf = _xf->data.db;
481     double* df = _df->data.db;
482     double scale, scale2 = f_param2;
483     assert( CV_IS_MAT_CONT( _xf->type & _df->type ) );
484 
485     if( activ_func == IDENTITY )
486     {
487         for( i = 0; i < n; i++, xf += cols, df += cols )
488             for( j = 0; j < cols; j++ )
489             {
490                 xf[j] += bias[j];
491                 df[j] = 1;
492             }
493         return;
494     }
495     else if( activ_func == GAUSSIAN )
496     {
497         scale = -f_param1*f_param1;
498         scale2 *= scale;
499         for( i = 0; i < n; i++, xf += cols, df += cols )
500             for( j = 0; j < cols; j++ )
501             {
502                 double t = xf[j] + bias[j];
503                 df[j] = t*2*scale2;
504                 xf[j] = t*t*scale;
505             }
506     }
507     else
508     {
509         scale = -f_param1;
510         for( i = 0; i < n; i++, xf += cols, df += cols )
511             for( j = 0; j < cols; j++ )
512                 xf[j] = (xf[j] + bias[j])*scale;
513     }
514 
515     cvExp( _xf, _xf );
516 
517     n *= cols;
518     xf -= n; df -= n;
519 
520     // ((1+exp(-ax))^-1)'=a*((1+exp(-ax))^-2)*exp(-ax);
521     // ((1-exp(-ax))/(1+exp(-ax)))'=(a*exp(-ax)*(1+exp(-ax)) + a*exp(-ax)*(1-exp(-ax)))/(1+exp(-ax))^2=
522     // 2*a*exp(-ax)/(1+exp(-ax))^2
523     switch( activ_func )
524     {
525     case SIGMOID_SYM:
526         scale *= -2*f_param2;
527         for( i = 0; i <= n - 4; i += 4 )
528         {
529             double x0 = 1.+xf[i], x1 = 1.+xf[i+1], x2 = 1.+xf[i+2], x3 = 1.+xf[i+3];
530             double a = x0*x1, b = x2*x3, d = 1./(a*b), t0, t1;
531             a *= d; b *= d;
532 
533             t0 = b*x1; t1 = b*x0;
534             df[i] = scale*xf[i]*t0*t0;
535             df[i+1] = scale*xf[i+1]*t1*t1;
536             t0 *= scale2*(2 - x0); t1 *= scale2*(2 - x1);
537             xf[i] = t0; xf[i+1] = t1;
538 
539             t0 = a*x3; t1 = a*x2;
540             df[i+2] = scale*xf[i+2]*t0*t0;
541             df[i+3] = scale*xf[i+3]*t1*t1;
542             t0 *= scale2*(2 - x2); t1 *= scale2*(2 - x3);
543             xf[i+2] = t0; xf[i+3] = t1;
544         }
545 
546         for( ; i < n; i++ )
547         {
548             double t0 = 1./(1. + xf[i]);
549             double t1 = scale*xf[i]*t0*t0;
550             t0 *= scale2*(1. - xf[i]);
551             df[i] = t1;
552             xf[i] = t0;
553         }
554         break;
555 
556     case GAUSSIAN:
557         for( i = 0; i < n; i++ )
558             df[i] *= xf[i];
559         break;
560     default:
561         ;
562     }
563 }
564 
565 
calc_input_scale(const CvVectors * vecs,int flags)566 void CvANN_MLP::calc_input_scale( const CvVectors* vecs, int flags )
567 {
568     bool reset_weights = (flags & UPDATE_WEIGHTS) == 0;
569     bool no_scale = (flags & NO_INPUT_SCALE) != 0;
570     double* scale = weights[0];
571     int count = vecs->count;
572 
573     if( reset_weights )
574     {
575         int i, j, vcount = layer_sizes->data.i[0];
576         int type = vecs->type;
577         double a = no_scale ? 1. : 0.;
578 
579         for( j = 0; j < vcount; j++ )
580             scale[2*j] = a, scale[j*2+1] = 0.;
581 
582         if( no_scale )
583             return;
584 
585         for( i = 0; i < count; i++ )
586         {
587             const float* f = vecs->data.fl[i];
588             const double* d = vecs->data.db[i];
589             for( j = 0; j < vcount; j++ )
590             {
591                 double t = type == CV_32F ? (double)f[j] : d[j];
592                 scale[j*2] += t;
593                 scale[j*2+1] += t*t;
594             }
595         }
596 
597         for( j = 0; j < vcount; j++ )
598         {
599             double s = scale[j*2], s2 = scale[j*2+1];
600             double m = s/count, sigma2 = s2/count - m*m;
601             scale[j*2] = sigma2 < DBL_EPSILON ? 1 : 1./sqrt(sigma2);
602             scale[j*2+1] = -m*scale[j*2];
603         }
604     }
605 }
606 
607 
calc_output_scale(const CvVectors * vecs,int flags)608 void CvANN_MLP::calc_output_scale( const CvVectors* vecs, int flags )
609 {
610     int i, j, vcount = layer_sizes->data.i[layer_sizes->cols-1];
611     int type = vecs->type;
612     double m = min_val, M = max_val, m1 = min_val1, M1 = max_val1;
613     bool reset_weights = (flags & UPDATE_WEIGHTS) == 0;
614     bool no_scale = (flags & NO_OUTPUT_SCALE) != 0;
615     int l_count = layer_sizes->cols;
616     double* scale = weights[l_count];
617     double* inv_scale = weights[l_count+1];
618     int count = vecs->count;
619 
620     CV_FUNCNAME( "CvANN_MLP::calc_output_scale" );
621 
622     __BEGIN__;
623 
624     if( reset_weights )
625     {
626         double a0 = no_scale ? 1 : DBL_MAX, b0 = no_scale ? 0 : -DBL_MAX;
627 
628         for( j = 0; j < vcount; j++ )
629         {
630             scale[2*j] = inv_scale[2*j] = a0;
631             scale[j*2+1] = inv_scale[2*j+1] = b0;
632         }
633 
634         if( no_scale )
635             EXIT;
636     }
637 
638     for( i = 0; i < count; i++ )
639     {
640         const float* f = vecs->data.fl[i];
641         const double* d = vecs->data.db[i];
642 
643         for( j = 0; j < vcount; j++ )
644         {
645             double t = type == CV_32F ? (double)f[j] : d[j];
646 
647             if( reset_weights )
648             {
649                 double mj = scale[j*2], Mj = scale[j*2+1];
650                 if( mj > t ) mj = t;
651                 if( Mj < t ) Mj = t;
652 
653                 scale[j*2] = mj;
654                 scale[j*2+1] = Mj;
655             }
656             else
657             {
658                 t = t*scale[j*2] + scale[2*j+1];
659                 if( t < m1 || t > M1 )
660                     CV_ERROR( CV_StsOutOfRange,
661                     "Some of new output training vector components run exceed the original range too much" );
662             }
663         }
664     }
665 
666     if( reset_weights )
667         for( j = 0; j < vcount; j++ )
668         {
669             // map mj..Mj to m..M
670             double mj = scale[j*2], Mj = scale[j*2+1];
671             double a, b;
672             double delta = Mj - mj;
673             if( delta < DBL_EPSILON )
674                 a = 1, b = (M + m - Mj - mj)*0.5;
675             else
676                 a = (M - m)/delta, b = m - mj*a;
677             inv_scale[j*2] = a; inv_scale[j*2+1] = b;
678             a = 1./a; b = -b*a;
679             scale[j*2] = a; scale[j*2+1] = b;
680         }
681 
682     __END__;
683 }
684 
685 
prepare_to_train(const CvMat * _inputs,const CvMat * _outputs,const CvMat * _sample_weights,const CvMat * _sample_idx,CvVectors * _ivecs,CvVectors * _ovecs,double ** _sw,int _flags)686 bool CvANN_MLP::prepare_to_train( const CvMat* _inputs, const CvMat* _outputs,
687             const CvMat* _sample_weights, const CvMat* _sample_idx,
688             CvVectors* _ivecs, CvVectors* _ovecs, double** _sw, int _flags )
689 {
690     bool ok = false;
691     CvMat* sample_idx = 0;
692     CvVectors ivecs, ovecs;
693     double* sw = 0;
694     int count = 0;
695 
696     CV_FUNCNAME( "CvANN_MLP::prepare_to_train" );
697 
698     ivecs.data.ptr = ovecs.data.ptr = 0;
699     assert( _ivecs && _ovecs );
700 
701     __BEGIN__;
702 
703     const int* sidx = 0;
704     int i, sw_type = 0, sw_count = 0;
705     int sw_step = 0;
706     double sw_sum = 0;
707 
708     if( !layer_sizes )
709         CV_ERROR( CV_StsError,
710         "The network has not been created. Use method create or the appropriate constructor" );
711 
712     if( !CV_IS_MAT(_inputs) || CV_MAT_TYPE(_inputs->type) != CV_32FC1 &&
713         CV_MAT_TYPE(_inputs->type) != CV_64FC1 || _inputs->cols != layer_sizes->data.i[0] )
714         CV_ERROR( CV_StsBadArg,
715         "input training data should be a floating-point matrix with"
716         "the number of rows equal to the number of training samples and "
717         "the number of columns equal to the size of 0-th (input) layer" );
718 
719     if( !CV_IS_MAT(_outputs) || CV_MAT_TYPE(_outputs->type) != CV_32FC1 &&
720         CV_MAT_TYPE(_outputs->type) != CV_64FC1 ||
721         _outputs->cols != layer_sizes->data.i[layer_sizes->cols - 1] )
722         CV_ERROR( CV_StsBadArg,
723         "output training data should be a floating-point matrix with"
724         "the number of rows equal to the number of training samples and "
725         "the number of columns equal to the size of last (output) layer" );
726 
727     if( _inputs->rows != _outputs->rows )
728         CV_ERROR( CV_StsUnmatchedSizes, "The numbers of input and output samples do not match" );
729 
730     if( _sample_idx )
731     {
732         CV_CALL( sample_idx = cvPreprocessIndexArray( _sample_idx, _inputs->rows ));
733         sidx = sample_idx->data.i;
734         count = sample_idx->cols + sample_idx->rows - 1;
735     }
736     else
737         count = _inputs->rows;
738 
739     if( _sample_weights )
740     {
741         if( !CV_IS_MAT(_sample_weights) )
742             CV_ERROR( CV_StsBadArg, "sample_weights (if passed) must be a valid matrix" );
743 
744         sw_type = CV_MAT_TYPE(_sample_weights->type);
745         sw_count = _sample_weights->cols + _sample_weights->rows - 1;
746 
747         if( sw_type != CV_32FC1 && sw_type != CV_64FC1 ||
748             _sample_weights->cols != 1 && _sample_weights->rows != 1 ||
749             sw_count != count && sw_count != _inputs->rows )
750             CV_ERROR( CV_StsBadArg,
751             "sample_weights must be 1d floating-point vector containing weights "
752             "of all or selected training samples" );
753 
754         sw_step = CV_IS_MAT_CONT(_sample_weights->type) ? 1 :
755             _sample_weights->step/CV_ELEM_SIZE(sw_type);
756 
757         CV_CALL( sw = (double*)cvAlloc( count*sizeof(sw[0]) ));
758     }
759 
760     CV_CALL( ivecs.data.ptr = (uchar**)cvAlloc( count*sizeof(ivecs.data.ptr[0]) ));
761     CV_CALL( ovecs.data.ptr = (uchar**)cvAlloc( count*sizeof(ovecs.data.ptr[0]) ));
762 
763     ivecs.type = CV_MAT_TYPE(_inputs->type);
764     ovecs.type = CV_MAT_TYPE(_outputs->type);
765     ivecs.count = ovecs.count = count;
766 
767     for( i = 0; i < count; i++ )
768     {
769         int idx = sidx ? sidx[i] : i;
770         ivecs.data.ptr[i] = _inputs->data.ptr + idx*_inputs->step;
771         ovecs.data.ptr[i] = _outputs->data.ptr + idx*_outputs->step;
772         if( sw )
773         {
774             int si = sw_count == count ? i : idx;
775             double w = sw_type == CV_32FC1 ?
776                 (double)_sample_weights->data.fl[si*sw_step] :
777                 _sample_weights->data.db[si*sw_step];
778             sw[i] = w;
779             if( w < 0 )
780                 CV_ERROR( CV_StsOutOfRange, "some of sample weights are negative" );
781             sw_sum += w;
782         }
783     }
784 
785     // normalize weights
786     if( sw )
787     {
788         sw_sum = sw_sum > DBL_EPSILON ? 1./sw_sum : 0;
789         for( i = 0; i < count; i++ )
790             sw[i] *= sw_sum;
791     }
792 
793     calc_input_scale( &ivecs, _flags );
794     CV_CALL( calc_output_scale( &ovecs, _flags ));
795 
796     ok = true;
797 
798     __END__;
799 
800     if( !ok )
801     {
802         cvFree( &ivecs.data.ptr );
803         cvFree( &ovecs.data.ptr );
804         cvFree( &sw );
805     }
806 
807     cvReleaseMat( &sample_idx );
808     *_ivecs = ivecs;
809     *_ovecs = ovecs;
810     *_sw = sw;
811 
812     return ok;
813 }
814 
815 
train(const CvMat * _inputs,const CvMat * _outputs,const CvMat * _sample_weights,const CvMat * _sample_idx,CvANN_MLP_TrainParams _params,int flags)816 int CvANN_MLP::train( const CvMat* _inputs, const CvMat* _outputs,
817                       const CvMat* _sample_weights, const CvMat* _sample_idx,
818                       CvANN_MLP_TrainParams _params, int flags )
819 {
820     const int MAX_ITER = 1000;
821     const double DEFAULT_EPSILON = FLT_EPSILON;
822 
823     double* sw = 0;
824     CvVectors x0, u;
825     int iter = -1;
826 
827     x0.data.ptr = u.data.ptr = 0;
828 
829     CV_FUNCNAME( "CvANN_MLP::train" );
830 
831     __BEGIN__;
832 
833     int max_iter;
834     double epsilon;
835 
836     params = _params;
837 
838     // initialize training data
839     CV_CALL( prepare_to_train( _inputs, _outputs, _sample_weights,
840                                _sample_idx, &x0, &u, &sw, flags ));
841 
842     // ... and link weights
843     if( !(flags & UPDATE_WEIGHTS) )
844         init_weights();
845 
846     max_iter = params.term_crit.type & CV_TERMCRIT_ITER ? params.term_crit.max_iter : MAX_ITER;
847     max_iter = MIN( max_iter, MAX_ITER );
848     max_iter = MAX( max_iter, 1 );
849 
850     epsilon = params.term_crit.type & CV_TERMCRIT_EPS ? params.term_crit.epsilon : DEFAULT_EPSILON;
851     epsilon = MAX(epsilon, DBL_EPSILON);
852 
853     params.term_crit.type = CV_TERMCRIT_ITER + CV_TERMCRIT_EPS;
854     params.term_crit.max_iter = max_iter;
855     params.term_crit.epsilon = epsilon;
856 
857     if( params.train_method == CvANN_MLP_TrainParams::BACKPROP )
858     {
859         CV_CALL( iter = train_backprop( x0, u, sw ));
860     }
861     else
862     {
863         CV_CALL( iter = train_rprop( x0, u, sw ));
864     }
865 
866     __END__;
867 
868     cvFree( &x0.data.ptr );
869     cvFree( &u.data.ptr );
870     cvFree( &sw );
871 
872     return iter;
873 }
874 
875 
train_backprop(CvVectors x0,CvVectors u,const double * sw)876 int CvANN_MLP::train_backprop( CvVectors x0, CvVectors u, const double* sw )
877 {
878     CvMat* dw = 0;
879     CvMat* buf = 0;
880     double **x = 0, **df = 0;
881     CvMat* _idx = 0;
882     int iter = -1, count = x0.count;
883 
884     CV_FUNCNAME( "CvANN_MLP::train_backprop" );
885 
886     __BEGIN__;
887 
888     int i, j, k, ivcount, ovcount, l_count, total = 0, max_iter;
889     double *buf_ptr;
890     double prev_E = DBL_MAX*0.5, E = 0, epsilon;
891 
892     max_iter = params.term_crit.max_iter*count;
893     epsilon = params.term_crit.epsilon*count;
894 
895     l_count = layer_sizes->cols;
896     ivcount = layer_sizes->data.i[0];
897     ovcount = layer_sizes->data.i[l_count-1];
898 
899     // allocate buffers
900     for( i = 0; i < l_count; i++ )
901         total += layer_sizes->data.i[i] + 1;
902 
903     CV_CALL( dw = cvCreateMat( wbuf->rows, wbuf->cols, wbuf->type ));
904     cvZero( dw );
905     CV_CALL( buf = cvCreateMat( 1, (total + max_count)*2, CV_64F ));
906     CV_CALL( _idx = cvCreateMat( 1, count, CV_32SC1 ));
907     for( i = 0; i < count; i++ )
908         _idx->data.i[i] = i;
909 
910     CV_CALL( x = (double**)cvAlloc( total*2*sizeof(x[0]) ));
911     df = x + total;
912     buf_ptr = buf->data.db;
913 
914     for( j = 0; j < l_count; j++ )
915     {
916         x[j] = buf_ptr;
917         df[j] = x[j] + layer_sizes->data.i[j];
918         buf_ptr += (df[j] - x[j])*2;
919     }
920 
921     // run back-propagation loop
922     /*
923         y_i = w_i*x_{i-1}
924         x_i = f(y_i)
925         E = 1/2*||u - x_N||^2
926         grad_N = (x_N - u)*f'(y_i)
927         dw_i(t) = momentum*dw_i(t-1) + dw_scale*x_{i-1}*grad_i
928         w_i(t+1) = w_i(t) + dw_i(t)
929         grad_{i-1} = w_i^t*grad_i
930     */
931     for( iter = 0; iter < max_iter; iter++ )
932     {
933         int idx = iter % count;
934         double* w = weights[0];
935         double sweight = sw ? count*sw[idx] : 1.;
936         CvMat _w, _dw, hdr1, hdr2, ghdr1, ghdr2, _df;
937         CvMat *x1 = &hdr1, *x2 = &hdr2, *grad1 = &ghdr1, *grad2 = &ghdr2, *temp;
938 
939         if( idx == 0 )
940         {
941             if( fabs(prev_E - E) < epsilon )
942                 break;
943             prev_E = E;
944             E = 0;
945 
946             // shuffle indices
947             for( i = 0; i < count; i++ )
948             {
949                 int tt;
950                 j = (unsigned)cvRandInt(&rng) % count;
951                 k = (unsigned)cvRandInt(&rng) % count;
952                 CV_SWAP( _idx->data.i[j], _idx->data.i[k], tt );
953             }
954         }
955 
956         idx = _idx->data.i[idx];
957 
958         if( x0.type == CV_32F )
959         {
960             const float* x0data = x0.data.fl[idx];
961             for( j = 0; j < ivcount; j++ )
962                 x[0][j] = x0data[j]*w[j*2] + w[j*2 + 1];
963         }
964         else
965         {
966             const double* x0data = x0.data.db[idx];
967             for( j = 0; j < ivcount; j++ )
968                 x[0][j] = x0data[j]*w[j*2] + w[j*2 + 1];
969         }
970 
971         cvInitMatHeader( x1, 1, ivcount, CV_64F, x[0] );
972 
973         // forward pass, compute y[i]=w*x[i-1], x[i]=f(y[i]), df[i]=f'(y[i])
974         for( i = 1; i < l_count; i++ )
975         {
976             cvInitMatHeader( x2, 1, layer_sizes->data.i[i], CV_64F, x[i] );
977             cvInitMatHeader( &_w, x1->cols, x2->cols, CV_64F, weights[i] );
978             cvGEMM( x1, &_w, 1, 0, 0, x2 );
979             _df = *x2;
980             _df.data.db = df[i];
981             calc_activ_func_deriv( x2, &_df, _w.data.db + _w.rows*_w.cols );
982             CV_SWAP( x1, x2, temp );
983         }
984 
985         cvInitMatHeader( grad1, 1, ovcount, CV_64F, buf_ptr );
986         *grad2 = *grad1;
987         grad2->data.db = buf_ptr + max_count;
988 
989         w = weights[l_count+1];
990 
991         // calculate error
992         if( u.type == CV_32F )
993         {
994             const float* udata = u.data.fl[idx];
995             for( k = 0; k < ovcount; k++ )
996             {
997                 double t = udata[k]*w[k*2] + w[k*2+1] - x[l_count-1][k];
998                 grad1->data.db[k] = t*sweight;
999                 E += t*t;
1000             }
1001         }
1002         else
1003         {
1004             const double* udata = u.data.db[idx];
1005             for( k = 0; k < ovcount; k++ )
1006             {
1007                 double t = udata[k]*w[k*2] + w[k*2+1] - x[l_count-1][k];
1008                 grad1->data.db[k] = t*sweight;
1009                 E += t*t;
1010             }
1011         }
1012         E *= sweight;
1013 
1014         // backward pass, update weights
1015         for( i = l_count-1; i > 0; i-- )
1016         {
1017             int n1 = layer_sizes->data.i[i-1], n2 = layer_sizes->data.i[i];
1018             cvInitMatHeader( &_df, 1, n2, CV_64F, df[i] );
1019             cvMul( grad1, &_df, grad1 );
1020             cvInitMatHeader( &_w, n1+1, n2, CV_64F, weights[i] );
1021             cvInitMatHeader( &_dw, n1+1, n2, CV_64F, dw->data.db + (weights[i] - weights[0]) );
1022             cvInitMatHeader( x1, n1+1, 1, CV_64F, x[i-1] );
1023             x[i-1][n1] = 1.;
1024             cvGEMM( x1, grad1, params.bp_dw_scale, &_dw, params.bp_moment_scale, &_dw );
1025             cvAdd( &_w, &_dw, &_w );
1026             if( i > 1 )
1027             {
1028                 grad2->cols = n1;
1029                 _w.rows = n1;
1030                 cvGEMM( grad1, &_w, 1, 0, 0, grad2, CV_GEMM_B_T );
1031             }
1032             CV_SWAP( grad1, grad2, temp );
1033         }
1034     }
1035 
1036     iter /= count;
1037 
1038     __END__;
1039 
1040     cvReleaseMat( &dw );
1041     cvReleaseMat( &buf );
1042     cvReleaseMat( &_idx );
1043     cvFree( &x );
1044 
1045     return iter;
1046 }
1047 
1048 
train_rprop(CvVectors x0,CvVectors u,const double * sw)1049 int CvANN_MLP::train_rprop( CvVectors x0, CvVectors u, const double* sw )
1050 {
1051     const int max_buf_sz = 1 << 16;
1052     CvMat* dw = 0;
1053     CvMat* dEdw = 0;
1054     CvMat* prev_dEdw_sign = 0;
1055     CvMat* buf = 0;
1056     double **x = 0, **df = 0;
1057     int iter = -1, count = x0.count;
1058 
1059     CV_FUNCNAME( "CvANN_MLP::train" );
1060 
1061     __BEGIN__;
1062 
1063     int i, ivcount, ovcount, l_count, total = 0, max_iter, buf_sz, dcount0, dcount=0;
1064     double *buf_ptr;
1065     double prev_E = DBL_MAX*0.5, epsilon;
1066     double dw_plus, dw_minus, dw_min, dw_max;
1067     double inv_count;
1068 
1069     max_iter = params.term_crit.max_iter;
1070     epsilon = params.term_crit.epsilon;
1071     dw_plus = params.rp_dw_plus;
1072     dw_minus = params.rp_dw_minus;
1073     dw_min = params.rp_dw_min;
1074     dw_max = params.rp_dw_max;
1075 
1076     l_count = layer_sizes->cols;
1077     ivcount = layer_sizes->data.i[0];
1078     ovcount = layer_sizes->data.i[l_count-1];
1079 
1080     // allocate buffers
1081     for( i = 0; i < l_count; i++ )
1082         total += layer_sizes->data.i[i];
1083 
1084     CV_CALL( dw = cvCreateMat( wbuf->rows, wbuf->cols, wbuf->type ));
1085     cvSet( dw, cvScalarAll(params.rp_dw0) );
1086     CV_CALL( dEdw = cvCreateMat( wbuf->rows, wbuf->cols, wbuf->type ));
1087     cvZero( dEdw );
1088     CV_CALL( prev_dEdw_sign = cvCreateMat( wbuf->rows, wbuf->cols, CV_8SC1 ));
1089     cvZero( prev_dEdw_sign );
1090 
1091     inv_count = 1./count;
1092     dcount0 = max_buf_sz/(2*total);
1093     dcount0 = MAX( dcount0, 1 );
1094     dcount0 = MIN( dcount0, count );
1095     buf_sz = dcount0*(total + max_count)*2;
1096 
1097     CV_CALL( buf = cvCreateMat( 1, buf_sz, CV_64F ));
1098 
1099     CV_CALL( x = (double**)cvAlloc( total*2*sizeof(x[0]) ));
1100     df = x + total;
1101     buf_ptr = buf->data.db;
1102 
1103     for( i = 0; i < l_count; i++ )
1104     {
1105         x[i] = buf_ptr;
1106         df[i] = x[i] + layer_sizes->data.i[i]*dcount0;
1107         buf_ptr += (df[i] - x[i])*2;
1108     }
1109 
1110     // run rprop loop
1111     /*
1112         y_i(t) = w_i(t)*x_{i-1}(t)
1113         x_i(t) = f(y_i(t))
1114         E = sum_over_all_samples(1/2*||u - x_N||^2)
1115         grad_N = (x_N - u)*f'(y_i)
1116 
1117                       MIN(dw_i{jk}(t)*dw_plus, dw_max), if dE/dw_i{jk}(t)*dE/dw_i{jk}(t-1) > 0
1118         dw_i{jk}(t) = MAX(dw_i{jk}(t)*dw_minus, dw_min), if dE/dw_i{jk}(t)*dE/dw_i{jk}(t-1) < 0
1119                       dw_i{jk}(t-1) else
1120 
1121         if (dE/dw_i{jk}(t)*dE/dw_i{jk}(t-1) < 0)
1122            dE/dw_i{jk}(t)<-0
1123         else
1124            w_i{jk}(t+1) = w_i{jk}(t) + dw_i{jk}(t)
1125         grad_{i-1}(t) = w_i^t(t)*grad_i(t)
1126     */
1127     for( iter = 0; iter < max_iter; iter++ )
1128     {
1129         int n1, n2, si, j, k;
1130         double* w;
1131         CvMat _w, _dEdw, hdr1, hdr2, ghdr1, ghdr2, _df;
1132         CvMat *x1, *x2, *grad1, *grad2, *temp;
1133         double E = 0;
1134 
1135         // first, iterate through all the samples and compute dEdw
1136         for( si = 0; si < count; si += dcount )
1137         {
1138             dcount = MIN( count - si, dcount0 );
1139             w = weights[0];
1140             grad1 = &ghdr1; grad2 = &ghdr2;
1141             x1 = &hdr1; x2 = &hdr2;
1142 
1143             // grab and preprocess input data
1144             if( x0.type == CV_32F )
1145                 for( i = 0; i < dcount; i++ )
1146                 {
1147                     const float* x0data = x0.data.fl[si+i];
1148                     double* xdata = x[0]+i*ivcount;
1149                     for( j = 0; j < ivcount; j++ )
1150                         xdata[j] = x0data[j]*w[j*2] + w[j*2+1];
1151                 }
1152             else
1153                 for( i = 0; i < dcount; i++ )
1154                 {
1155                     const double* x0data = x0.data.db[si+i];
1156                     double* xdata = x[0]+i*ivcount;
1157                     for( j = 0; j < ivcount; j++ )
1158                         xdata[j] = x0data[j]*w[j*2] + w[j*2+1];
1159                 }
1160 
1161             cvInitMatHeader( x1, dcount, ivcount, CV_64F, x[0] );
1162 
1163             // forward pass, compute y[i]=w*x[i-1], x[i]=f(y[i]), df[i]=f'(y[i])
1164             for( i = 1; i < l_count; i++ )
1165             {
1166                 cvInitMatHeader( x2, dcount, layer_sizes->data.i[i], CV_64F, x[i] );
1167                 cvInitMatHeader( &_w, x1->cols, x2->cols, CV_64F, weights[i] );
1168                 cvGEMM( x1, &_w, 1, 0, 0, x2 );
1169                 _df = *x2;
1170                 _df.data.db = df[i];
1171                 calc_activ_func_deriv( x2, &_df, _w.data.db + _w.rows*_w.cols );
1172                 CV_SWAP( x1, x2, temp );
1173             }
1174 
1175             cvInitMatHeader( grad1, dcount, ovcount, CV_64F, buf_ptr );
1176             w = weights[l_count+1];
1177             grad2->data.db = buf_ptr + max_count*dcount;
1178 
1179             // calculate error
1180             if( u.type == CV_32F )
1181                 for( i = 0; i < dcount; i++ )
1182                 {
1183                     const float* udata = u.data.fl[si+i];
1184                     const double* xdata = x[l_count-1] + i*ovcount;
1185                     double* gdata = grad1->data.db + i*ovcount;
1186                     double sweight = sw ? sw[si+i] : inv_count, E1 = 0;
1187 
1188                     for( j = 0; j < ovcount; j++ )
1189                     {
1190                         double t = udata[j]*w[j*2] + w[j*2+1] - xdata[j];
1191                         gdata[j] = t*sweight;
1192                         E1 += t*t;
1193                     }
1194                     E += sweight*E1;
1195                 }
1196             else
1197                 for( i = 0; i < dcount; i++ )
1198                 {
1199                     const double* udata = u.data.db[si+i];
1200                     const double* xdata = x[l_count-1] + i*ovcount;
1201                     double* gdata = grad1->data.db + i*ovcount;
1202                     double sweight = sw ? sw[si+i] : inv_count, E1 = 0;
1203 
1204                     for( j = 0; j < ovcount; j++ )
1205                     {
1206                         double t = udata[j]*w[j*2] + w[j*2+1] - xdata[j];
1207                         gdata[j] = t*sweight;
1208                         E1 += t*t;
1209                     }
1210                     E += sweight*E1;
1211                 }
1212 
1213             // backward pass, update dEdw
1214             for( i = l_count-1; i > 0; i-- )
1215             {
1216                 n1 = layer_sizes->data.i[i-1]; n2 = layer_sizes->data.i[i];
1217                 cvInitMatHeader( &_df, dcount, n2, CV_64F, df[i] );
1218                 cvMul( grad1, &_df, grad1 );
1219                 cvInitMatHeader( &_dEdw, n1, n2, CV_64F, dEdw->data.db+(weights[i]-weights[0]) );
1220                 cvInitMatHeader( x1, dcount, n1, CV_64F, x[i-1] );
1221                 cvGEMM( x1, grad1, 1, &_dEdw, 1, &_dEdw, CV_GEMM_A_T );
1222                 // update bias part of dEdw
1223                 for( k = 0; k < dcount; k++ )
1224                 {
1225                     double* dst = _dEdw.data.db + n1*n2;
1226                     const double* src = grad1->data.db + k*n2;
1227                     for( j = 0; j < n2; j++ )
1228                         dst[j] += src[j];
1229                 }
1230                 cvInitMatHeader( &_w, n1, n2, CV_64F, weights[i] );
1231                 cvInitMatHeader( grad2, dcount, n1, CV_64F, grad2->data.db );
1232 
1233                 if( i > 1 )
1234                     cvGEMM( grad1, &_w, 1, 0, 0, grad2, CV_GEMM_B_T );
1235                 CV_SWAP( grad1, grad2, temp );
1236             }
1237         }
1238 
1239         // now update weights
1240         for( i = 1; i < l_count; i++ )
1241         {
1242             n1 = layer_sizes->data.i[i-1]; n2 = layer_sizes->data.i[i];
1243             for( k = 0; k <= n1; k++ )
1244             {
1245                 double* wk = weights[i]+k*n2;
1246                 size_t delta = wk - weights[0];
1247                 double* dwk = dw->data.db + delta;
1248                 double* dEdwk = dEdw->data.db + delta;
1249                 char* prevEk = (char*)(prev_dEdw_sign->data.ptr + delta);
1250 
1251                 for( j = 0; j < n2; j++ )
1252                 {
1253                     double Eval = dEdwk[j];
1254                     double dval = dwk[j];
1255                     double wval = wk[j];
1256                     int s = CV_SIGN(Eval);
1257                     int ss = prevEk[j]*s;
1258                     if( ss > 0 )
1259                     {
1260                         dval *= dw_plus;
1261                         dval = MIN( dval, dw_max );
1262                         dwk[j] = dval;
1263                         wk[j] = wval + dval*s;
1264                     }
1265                     else if( ss < 0 )
1266                     {
1267                         dval *= dw_minus;
1268                         dval = MAX( dval, dw_min );
1269                         prevEk[j] = 0;
1270                         dwk[j] = dval;
1271                         wk[j] = wval + dval*s;
1272                     }
1273                     else
1274                     {
1275                         prevEk[j] = (char)s;
1276                         wk[j] = wval + dval*s;
1277                     }
1278                     dEdwk[j] = 0.;
1279                 }
1280             }
1281         }
1282 
1283         if( fabs(prev_E - E) < epsilon )
1284             break;
1285         prev_E = E;
1286         E = 0;
1287     }
1288 
1289     __END__;
1290 
1291     cvReleaseMat( &dw );
1292     cvReleaseMat( &dEdw );
1293     cvReleaseMat( &prev_dEdw_sign );
1294     cvReleaseMat( &buf );
1295     cvFree( &x );
1296 
1297     return iter;
1298 }
1299 
1300 
write_params(CvFileStorage * fs)1301 void CvANN_MLP::write_params( CvFileStorage* fs )
1302 {
1303     //CV_FUNCNAME( "CvANN_MLP::write_params" );
1304 
1305     __BEGIN__;
1306 
1307     const char* activ_func_name = activ_func == IDENTITY ? "IDENTITY" :
1308                             activ_func == SIGMOID_SYM ? "SIGMOID_SYM" :
1309                             activ_func == GAUSSIAN ? "GAUSSIAN" : 0;
1310 
1311     if( activ_func_name )
1312         cvWriteString( fs, "activation_function", activ_func_name );
1313     else
1314         cvWriteInt( fs, "activation_function", activ_func );
1315 
1316     if( activ_func != IDENTITY )
1317     {
1318         cvWriteReal( fs, "f_param1", f_param1 );
1319         cvWriteReal( fs, "f_param2", f_param2 );
1320     }
1321 
1322     cvWriteReal( fs, "min_val", min_val );
1323     cvWriteReal( fs, "max_val", max_val );
1324     cvWriteReal( fs, "min_val1", min_val1 );
1325     cvWriteReal( fs, "max_val1", max_val1 );
1326 
1327     cvStartWriteStruct( fs, "training_params", CV_NODE_MAP );
1328     if( params.train_method == CvANN_MLP_TrainParams::BACKPROP )
1329     {
1330         cvWriteString( fs, "train_method", "BACKPROP" );
1331         cvWriteReal( fs, "dw_scale", params.bp_dw_scale );
1332         cvWriteReal( fs, "moment_scale", params.bp_moment_scale );
1333     }
1334     else if( params.train_method == CvANN_MLP_TrainParams::RPROP )
1335     {
1336         cvWriteString( fs, "train_method", "RPROP" );
1337         cvWriteReal( fs, "dw0", params.rp_dw0 );
1338         cvWriteReal( fs, "dw_plus", params.rp_dw_plus );
1339         cvWriteReal( fs, "dw_minus", params.rp_dw_minus );
1340         cvWriteReal( fs, "dw_min", params.rp_dw_min );
1341         cvWriteReal( fs, "dw_max", params.rp_dw_max );
1342     }
1343 
1344     cvStartWriteStruct( fs, "term_criteria", CV_NODE_MAP + CV_NODE_FLOW );
1345     if( params.term_crit.type & CV_TERMCRIT_EPS )
1346         cvWriteReal( fs, "epsilon", params.term_crit.epsilon );
1347     if( params.term_crit.type & CV_TERMCRIT_ITER )
1348         cvWriteInt( fs, "iterations", params.term_crit.max_iter );
1349     cvEndWriteStruct( fs );
1350 
1351     cvEndWriteStruct( fs );
1352 
1353     __END__;
1354 }
1355 
1356 
write(CvFileStorage * fs,const char * name)1357 void CvANN_MLP::write( CvFileStorage* fs, const char* name )
1358 {
1359     CV_FUNCNAME( "CvANN_MLP::write" );
1360 
1361     __BEGIN__;
1362 
1363     int i, l_count = layer_sizes->cols;
1364 
1365     if( !layer_sizes )
1366         CV_ERROR( CV_StsError, "The network has not been initialized" );
1367 
1368     cvStartWriteStruct( fs, name, CV_NODE_MAP, CV_TYPE_NAME_ML_ANN_MLP );
1369 
1370     cvWrite( fs, "layer_sizes", layer_sizes );
1371 
1372     write_params( fs );
1373 
1374     cvStartWriteStruct( fs, "input_scale", CV_NODE_SEQ + CV_NODE_FLOW );
1375     cvWriteRawData( fs, weights[0], layer_sizes->data.i[0]*2, "d" );
1376     cvEndWriteStruct( fs );
1377 
1378     cvStartWriteStruct( fs, "output_scale", CV_NODE_SEQ + CV_NODE_FLOW );
1379     cvWriteRawData( fs, weights[l_count], layer_sizes->data.i[l_count-1]*2, "d" );
1380     cvEndWriteStruct( fs );
1381 
1382     cvStartWriteStruct( fs, "inv_output_scale", CV_NODE_SEQ + CV_NODE_FLOW );
1383     cvWriteRawData( fs, weights[l_count+1], layer_sizes->data.i[l_count-1]*2, "d" );
1384     cvEndWriteStruct( fs );
1385 
1386     cvStartWriteStruct( fs, "weights", CV_NODE_SEQ );
1387     for( i = 1; i < l_count; i++ )
1388     {
1389         cvStartWriteStruct( fs, 0, CV_NODE_SEQ + CV_NODE_FLOW );
1390         cvWriteRawData( fs, weights[i], (layer_sizes->data.i[i-1]+1)*layer_sizes->data.i[i], "d" );
1391         cvEndWriteStruct( fs );
1392     }
1393 
1394     cvEndWriteStruct( fs );
1395 
1396     __END__;
1397 }
1398 
1399 
read_params(CvFileStorage * fs,CvFileNode * node)1400 void CvANN_MLP::read_params( CvFileStorage* fs, CvFileNode* node )
1401 {
1402     //CV_FUNCNAME( "CvANN_MLP::read_params" );
1403 
1404     __BEGIN__;
1405 
1406     const char* activ_func_name = cvReadStringByName( fs, node, "activation_function", 0 );
1407     CvFileNode* tparams_node;
1408 
1409     if( activ_func_name )
1410         activ_func = strcmp( activ_func_name, "SIGMOID_SYM" ) == 0 ? SIGMOID_SYM :
1411                      strcmp( activ_func_name, "IDENTITY" ) == 0 ? IDENTITY :
1412                      strcmp( activ_func_name, "GAUSSIAN" ) == 0 ? GAUSSIAN : 0;
1413     else
1414         activ_func = cvReadIntByName( fs, node, "activation_function" );
1415 
1416     f_param1 = cvReadRealByName( fs, node, "f_param1", 0 );
1417     f_param2 = cvReadRealByName( fs, node, "f_param2", 0 );
1418 
1419     set_activ_func( activ_func, f_param1, f_param2 );
1420 
1421     min_val = cvReadRealByName( fs, node, "min_val", 0. );
1422     max_val = cvReadRealByName( fs, node, "max_val", 1. );
1423     min_val1 = cvReadRealByName( fs, node, "min_val1", 0. );
1424     max_val1 = cvReadRealByName( fs, node, "max_val1", 1. );
1425 
1426     tparams_node = cvGetFileNodeByName( fs, node, "training_params" );
1427     params = CvANN_MLP_TrainParams();
1428 
1429     if( tparams_node )
1430     {
1431         const char* tmethod_name = cvReadStringByName( fs, tparams_node, "train_method", "" );
1432         CvFileNode* tcrit_node;
1433 
1434         if( strcmp( tmethod_name, "BACKPROP" ) == 0 )
1435         {
1436             params.train_method = CvANN_MLP_TrainParams::BACKPROP;
1437             params.bp_dw_scale = cvReadRealByName( fs, tparams_node, "dw_scale", 0 );
1438             params.bp_moment_scale = cvReadRealByName( fs, tparams_node, "moment_scale", 0 );
1439         }
1440         else if( strcmp( tmethod_name, "RPROP" ) == 0 )
1441         {
1442             params.train_method = CvANN_MLP_TrainParams::RPROP;
1443             params.rp_dw0 = cvReadRealByName( fs, tparams_node, "dw0", 0 );
1444             params.rp_dw_plus = cvReadRealByName( fs, tparams_node, "dw_plus", 0 );
1445             params.rp_dw_minus = cvReadRealByName( fs, tparams_node, "dw_minus", 0 );
1446             params.rp_dw_min = cvReadRealByName( fs, tparams_node, "dw_min", 0 );
1447             params.rp_dw_max = cvReadRealByName( fs, tparams_node, "dw_max", 0 );
1448         }
1449 
1450         tcrit_node = cvGetFileNodeByName( fs, tparams_node, "term_criteria" );
1451         if( tcrit_node )
1452         {
1453             params.term_crit.epsilon = cvReadRealByName( fs, tcrit_node, "epsilon", -1 );
1454             params.term_crit.max_iter = cvReadIntByName( fs, tcrit_node, "iterations", -1 );
1455             params.term_crit.type = (params.term_crit.epsilon >= 0 ? CV_TERMCRIT_EPS : 0) +
1456                                    (params.term_crit.max_iter >= 0 ? CV_TERMCRIT_ITER : 0);
1457         }
1458     }
1459 
1460     __END__;
1461 }
1462 
1463 
read(CvFileStorage * fs,CvFileNode * node)1464 void CvANN_MLP::read( CvFileStorage* fs, CvFileNode* node )
1465 {
1466     CvMat* _layer_sizes = 0;
1467 
1468     CV_FUNCNAME( "CvANN_MLP::read" );
1469 
1470     __BEGIN__;
1471 
1472     CvFileNode* w;
1473     CvSeqReader reader;
1474     int i, l_count;
1475 
1476     _layer_sizes = (CvMat*)cvReadByName( fs, node, "layer_sizes" );
1477     CV_CALL( create( _layer_sizes, SIGMOID_SYM, 0, 0 ));
1478     l_count = layer_sizes->cols;
1479 
1480     CV_CALL( read_params( fs, node ));
1481 
1482     w = cvGetFileNodeByName( fs, node, "input_scale" );
1483     if( !w || CV_NODE_TYPE(w->tag) != CV_NODE_SEQ ||
1484         w->data.seq->total != layer_sizes->data.i[0]*2 )
1485         CV_ERROR( CV_StsParseError, "input_scale tag is not found or is invalid" );
1486 
1487     CV_CALL( cvReadRawData( fs, w, weights[0], "d" ));
1488 
1489     w = cvGetFileNodeByName( fs, node, "output_scale" );
1490     if( !w || CV_NODE_TYPE(w->tag) != CV_NODE_SEQ ||
1491         w->data.seq->total != layer_sizes->data.i[l_count-1]*2 )
1492         CV_ERROR( CV_StsParseError, "output_scale tag is not found or is invalid" );
1493 
1494     CV_CALL( cvReadRawData( fs, w, weights[l_count], "d" ));
1495 
1496     w = cvGetFileNodeByName( fs, node, "inv_output_scale" );
1497     if( !w || CV_NODE_TYPE(w->tag) != CV_NODE_SEQ ||
1498         w->data.seq->total != layer_sizes->data.i[l_count-1]*2 )
1499         CV_ERROR( CV_StsParseError, "inv_output_scale tag is not found or is invalid" );
1500 
1501     CV_CALL( cvReadRawData( fs, w, weights[l_count+1], "d" ));
1502 
1503     w = cvGetFileNodeByName( fs, node, "weights" );
1504     if( !w || CV_NODE_TYPE(w->tag) != CV_NODE_SEQ ||
1505         w->data.seq->total != l_count - 1 )
1506         CV_ERROR( CV_StsParseError, "weights tag is not found or is invalid" );
1507 
1508     cvStartReadSeq( w->data.seq, &reader );
1509 
1510     for( i = 1; i < l_count; i++ )
1511     {
1512         w = (CvFileNode*)reader.ptr;
1513         CV_CALL( cvReadRawData( fs, w, weights[i], "d" ));
1514         CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
1515     }
1516 
1517     __END__;
1518 }
1519 
1520 /* End of file. */
1521