1 /*M///////////////////////////////////////////////////////////////////////////////////////
\r
3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
\r
5 // By downloading, copying, installing or using the software you agree to this license.
\r
6 // If you do not agree to this license, do not download, install,
\r
7 // copy or use the software.
\r
10 // Intel License Agreement
\r
12 // Copyright (C) 2000, Intel Corporation, all rights reserved.
\r
13 // Third party copyrights are property of their respective owners.
\r
15 // Redistribution and use in source and binary forms, with or without modification,
\r
16 // are permitted provided that the following conditions are met:
\r
18 // * Redistribution's of source code must retain the above copyright notice,
\r
19 // this list of conditions and the following disclaimer.
\r
21 // * Redistribution's in binary form must reproduce the above copyright notice,
\r
22 // this list of conditions and the following disclaimer in the documentation
\r
23 // and/or other materials provided with the distribution.
\r
25 // * The name of Intel Corporation may not be used to endorse or promote products
\r
26 // derived from this software without specific prior written permission.
\r
28 // This software is provided by the copyright holders and contributors "as is" and
\r
29 // any express or implied warranties, including, but not limited to, the implied
\r
30 // warranties of merchantability and fitness for a particular purpose are disclaimed.
\r
31 // In no event shall the Intel Corporation or contributors be liable for any direct,
\r
32 // indirect, incidental, special, exemplary, or consequential damages
\r
33 // (including, but not limited to, procurement of substitute goods or services;
\r
34 // loss of use, data, or profits; or business interruption) however caused
\r
35 // and on any theory of liability, whether in contract, strict liability,
\r
36 // or tort (including negligence or otherwise) arising in any way out of
\r
37 // the use of this software, even if advised of the possibility of such damage.
\r
44 static const float ord_nan = FLT_MAX*0.5f;
\r
45 static const int min_block_size = 1 << 16;
\r
46 static const int block_size_delta = 1 << 10;
\r
48 CvDTreeTrainData::CvDTreeTrainData()
\r
50 var_idx = var_type = cat_count = cat_ofs = cat_map =
\r
51 priors = priors_mult = counts = buf = direction = split_buf = responses_copy = 0;
\r
52 pred_int_buf = resp_int_buf = cv_lables_buf = sample_idx_buf = 0;
\r
53 pred_float_buf = resp_float_buf = 0;
\r
54 tree_storage = temp_storage = 0;
\r
60 CvDTreeTrainData::CvDTreeTrainData( const CvMat* _train_data, int _tflag,
\r
61 const CvMat* _responses, const CvMat* _var_idx,
\r
62 const CvMat* _sample_idx, const CvMat* _var_type,
\r
63 const CvMat* _missing_mask, const CvDTreeParams& _params,
\r
64 bool _shared, bool _add_labels )
\r
66 var_idx = var_type = cat_count = cat_ofs = cat_map =
\r
67 priors = priors_mult = counts = buf = direction = split_buf = responses_copy = 0;
\r
69 pred_int_buf = resp_int_buf = cv_lables_buf = sample_idx_buf = 0;
\r
70 pred_float_buf = resp_float_buf = 0;
\r
72 tree_storage = temp_storage = 0;
\r
74 set_data( _train_data, _tflag, _responses, _var_idx, _sample_idx,
\r
75 _var_type, _missing_mask, _params, _shared, _add_labels );
\r
79 CvDTreeTrainData::~CvDTreeTrainData()
\r
85 bool CvDTreeTrainData::set_params( const CvDTreeParams& _params )
\r
89 CV_FUNCNAME( "CvDTreeTrainData::set_params" );
\r
96 if( params.max_categories < 2 )
\r
97 CV_ERROR( CV_StsOutOfRange, "params.max_categories should be >= 2" );
\r
98 params.max_categories = MIN( params.max_categories, 15 );
\r
100 if( params.max_depth < 0 )
\r
101 CV_ERROR( CV_StsOutOfRange, "params.max_depth should be >= 0" );
\r
102 params.max_depth = MIN( params.max_depth, 25 );
\r
104 params.min_sample_count = MAX(params.min_sample_count,1);
\r
106 if( params.cv_folds < 0 )
\r
107 CV_ERROR( CV_StsOutOfRange,
\r
108 "params.cv_folds should be =0 (the tree is not pruned) "
\r
109 "or n>0 (tree is pruned using n-fold cross-validation)" );
\r
111 if( params.cv_folds == 1 )
\r
112 params.cv_folds = 0;
\r
114 if( params.regression_accuracy < 0 )
\r
115 CV_ERROR( CV_StsOutOfRange, "params.regression_accuracy should be >= 0" );
\r
124 #define CV_CMP_NUM_PTR(a,b) (*(a) < *(b))
\r
125 static CV_IMPLEMENT_QSORT_EX( icvSortIntPtr, int*, CV_CMP_NUM_PTR, int )
\r
126 static CV_IMPLEMENT_QSORT_EX( icvSortDblPtr, double*, CV_CMP_NUM_PTR, int )
\r
128 #define CV_CMP_NUM_IDX(i,j) (aux[i] < aux[j])
\r
129 static CV_IMPLEMENT_QSORT_EX( icvSortIntAux, int, CV_CMP_NUM_IDX, const float* )
\r
130 static CV_IMPLEMENT_QSORT_EX( icvSortUShAux, unsigned short, CV_CMP_NUM_IDX, const float* )
\r
132 #define CV_CMP_PAIRS(a,b) (*((a).i) < *((b).i))
\r
133 static CV_IMPLEMENT_QSORT_EX( icvSortPairs, CvPair16u32s, CV_CMP_PAIRS, int )
\r
135 void CvDTreeTrainData::set_data( const CvMat* _train_data, int _tflag,
\r
136 const CvMat* _responses, const CvMat* _var_idx, const CvMat* _sample_idx,
\r
137 const CvMat* _var_type, const CvMat* _missing_mask, const CvDTreeParams& _params,
\r
138 bool _shared, bool _add_labels, bool _update_data )
\r
140 CvMat* sample_indices = 0;
\r
141 CvMat* var_type0 = 0;
\r
142 CvMat* tmp_map = 0;
\r
144 CvPair16u32s* pair16u32s_ptr = 0;
\r
145 CvDTreeTrainData* data = 0;
\r
148 unsigned short* udst = 0;
\r
151 CV_FUNCNAME( "CvDTreeTrainData::set_data" );
\r
155 int sample_all = 0, r_type = 0, cv_n;
\r
156 int total_c_count = 0;
\r
157 int tree_block_size, temp_block_size, max_split_size, nv_size, cv_size = 0;
\r
158 int ds_step, dv_step, ms_step = 0, mv_step = 0; // {data|mask}{sample|var}_step
\r
161 const int *sidx = 0, *vidx = 0;
\r
164 if( _update_data && data_root )
\r
166 data = new CvDTreeTrainData( _train_data, _tflag, _responses, _var_idx,
\r
167 _sample_idx, _var_type, _missing_mask, _params, _shared, _add_labels );
\r
169 // compare new and old train data
\r
170 if( !(data->var_count == var_count &&
\r
171 cvNorm( data->var_type, var_type, CV_C ) < FLT_EPSILON &&
\r
172 cvNorm( data->cat_count, cat_count, CV_C ) < FLT_EPSILON &&
\r
173 cvNorm( data->cat_map, cat_map, CV_C ) < FLT_EPSILON) )
\r
174 CV_ERROR( CV_StsBadArg,
\r
175 "The new training data must have the same types and the input and output variables "
\r
176 "and the same categories for categorical variables" );
\r
178 cvReleaseMat( &priors );
\r
179 cvReleaseMat( &priors_mult );
\r
180 cvReleaseMat( &buf );
\r
181 cvReleaseMat( &direction );
\r
182 cvReleaseMat( &split_buf );
\r
183 cvReleaseMemStorage( &temp_storage );
\r
185 priors = data->priors; data->priors = 0;
\r
186 priors_mult = data->priors_mult; data->priors_mult = 0;
\r
187 buf = data->buf; data->buf = 0;
\r
188 buf_count = data->buf_count; buf_size = data->buf_size;
\r
189 sample_count = data->sample_count;
\r
191 direction = data->direction; data->direction = 0;
\r
192 split_buf = data->split_buf; data->split_buf = 0;
\r
193 temp_storage = data->temp_storage; data->temp_storage = 0;
\r
194 nv_heap = data->nv_heap; cv_heap = data->cv_heap;
\r
196 data_root = new_node( 0, sample_count, 0, 0 );
\r
205 CV_CALL( set_params( _params ));
\r
207 // check parameter types and sizes
\r
208 CV_CALL( cvCheckTrainData( _train_data, _tflag, _missing_mask, &var_all, &sample_all ));
\r
210 train_data = _train_data;
\r
211 responses = _responses;
\r
213 if( _tflag == CV_ROW_SAMPLE )
\r
215 ds_step = _train_data->step/CV_ELEM_SIZE(_train_data->type);
\r
217 if( _missing_mask )
\r
218 ms_step = _missing_mask->step, mv_step = 1;
\r
222 dv_step = _train_data->step/CV_ELEM_SIZE(_train_data->type);
\r
224 if( _missing_mask )
\r
225 mv_step = _missing_mask->step, ms_step = 1;
\r
229 sample_count = sample_all;
\r
230 var_count = var_all;
\r
234 CV_CALL( sample_indices = cvPreprocessIndexArray( _sample_idx, sample_all ));
\r
235 sidx = sample_indices->data.i;
\r
236 sample_count = sample_indices->rows + sample_indices->cols - 1;
\r
241 CV_CALL( var_idx = cvPreprocessIndexArray( _var_idx, var_all ));
\r
242 vidx = var_idx->data.i;
\r
243 var_count = var_idx->rows + var_idx->cols - 1;
\r
246 is_buf_16u = false;
\r
247 if ( sample_count < 65536 )
\r
248 is_buf_16u = true;
\r
250 if( !CV_IS_MAT(_responses) ||
\r
251 (CV_MAT_TYPE(_responses->type) != CV_32SC1 &&
\r
252 CV_MAT_TYPE(_responses->type) != CV_32FC1) ||
\r
253 (_responses->rows != 1 && _responses->cols != 1) ||
\r
254 _responses->rows + _responses->cols - 1 != sample_all )
\r
255 CV_ERROR( CV_StsBadArg, "The array of _responses must be an integer or "
\r
256 "floating-point vector containing as many elements as "
\r
257 "the total number of samples in the training data matrix" );
\r
260 CV_CALL( var_type0 = cvPreprocessVarType( _var_type, var_idx, var_count, &r_type ));
\r
262 CV_CALL( var_type = cvCreateMat( 1, var_count+2, CV_32SC1 ));
\r
266 ord_var_count = -1;
\r
268 is_classifier = r_type == CV_VAR_CATEGORICAL;
\r
270 // step 0. calc the number of categorical vars
\r
271 for( vi = 0; vi < var_count; vi++ )
\r
273 var_type->data.i[vi] = var_type0->data.ptr[vi] == CV_VAR_CATEGORICAL ?
\r
274 cat_var_count++ : ord_var_count--;
\r
277 ord_var_count = ~ord_var_count;
\r
278 cv_n = params.cv_folds;
\r
279 // set the two last elements of var_type array to be able
\r
280 // to locate responses and cross-validation labels using
\r
281 // the corresponding get_* functions.
\r
282 var_type->data.i[var_count] = cat_var_count;
\r
283 var_type->data.i[var_count+1] = cat_var_count+1;
\r
285 // in case of single ordered predictor we need dummy cv_labels
\r
286 // for safe split_node_data() operation
\r
287 have_labels = cv_n > 0 || (ord_var_count == 1 && cat_var_count == 0) || _add_labels;
\r
289 work_var_count = var_count + (is_classifier ? 1 : 0) + (have_labels ? 1 : 0);
\r
290 buf_size = (work_var_count + 1)*sample_count;
\r
292 buf_count = shared ? 2 : 1;
\r
296 CV_CALL( buf = cvCreateMat( buf_count, buf_size, CV_16UC1 ));
\r
297 CV_CALL( pair16u32s_ptr = (CvPair16u32s*)cvAlloc( sample_count*sizeof(pair16u32s_ptr[0]) ));
\r
301 CV_CALL( buf = cvCreateMat( buf_count, buf_size, CV_32SC1 ));
\r
302 CV_CALL( int_ptr = (int**)cvAlloc( sample_count*sizeof(int_ptr[0]) ));
\r
305 size = is_classifier ? (cat_var_count+1) : cat_var_count;
\r
306 size = !size ? 1 : size;
\r
307 CV_CALL( cat_count = cvCreateMat( 1, size, CV_32SC1 ));
\r
308 CV_CALL( cat_ofs = cvCreateMat( 1, size, CV_32SC1 ));
\r
310 size = is_classifier ? (cat_var_count + 1)*params.max_categories : cat_var_count*params.max_categories;
\r
311 size = !size ? 1 : size;
\r
312 CV_CALL( cat_map = cvCreateMat( 1, size, CV_32SC1 ));
\r
314 // now calculate the maximum size of split,
\r
315 // create memory storage that will keep nodes and splits of the decision tree
\r
316 // allocate root node and the buffer for the whole training data
\r
317 max_split_size = cvAlign(sizeof(CvDTreeSplit) +
\r
318 (MAX(0,sample_count - 33)/32)*sizeof(int),sizeof(void*));
\r
319 tree_block_size = MAX((int)sizeof(CvDTreeNode)*8, max_split_size);
\r
320 tree_block_size = MAX(tree_block_size + block_size_delta, min_block_size);
\r
321 CV_CALL( tree_storage = cvCreateMemStorage( tree_block_size ));
\r
322 CV_CALL( node_heap = cvCreateSet( 0, sizeof(*node_heap), sizeof(CvDTreeNode), tree_storage ));
\r
324 nv_size = var_count*sizeof(int);
\r
325 nv_size = cvAlign(MAX( nv_size, (int)sizeof(CvSetElem) ), sizeof(void*));
\r
327 temp_block_size = nv_size;
\r
331 if( sample_count < cv_n*MAX(params.min_sample_count,10) )
\r
332 CV_ERROR( CV_StsOutOfRange,
\r
333 "The many folds in cross-validation for such a small dataset" );
\r
335 cv_size = cvAlign( cv_n*(sizeof(int) + sizeof(double)*2), sizeof(double) );
\r
336 temp_block_size = MAX(temp_block_size, cv_size);
\r
339 temp_block_size = MAX( temp_block_size + block_size_delta, min_block_size );
\r
340 CV_CALL( temp_storage = cvCreateMemStorage( temp_block_size ));
\r
341 CV_CALL( nv_heap = cvCreateSet( 0, sizeof(*nv_heap), nv_size, temp_storage ));
\r
343 CV_CALL( cv_heap = cvCreateSet( 0, sizeof(*cv_heap), cv_size, temp_storage ));
\r
345 CV_CALL( data_root = new_node( 0, sample_count, 0, 0 ));
\r
352 _fdst = (float*)cvAlloc(sample_count*sizeof(_fdst[0]));
\r
353 if (is_buf_16u && (cat_var_count || is_classifier))
\r
354 _idst = (int*)cvAlloc(sample_count*sizeof(_idst[0]));
\r
356 // transform the training data to convenient representation
\r
357 for( vi = 0; vi <= var_count; vi++ )
\r
360 const uchar* mask = 0;
\r
361 int m_step = 0, step;
\r
362 const int* idata = 0;
\r
363 const float* fdata = 0;
\r
366 if( vi < var_count ) // analyze i-th input variable
\r
368 int vi0 = vidx ? vidx[vi] : vi;
\r
369 ci = get_var_type(vi);
\r
370 step = ds_step; m_step = ms_step;
\r
371 if( CV_MAT_TYPE(_train_data->type) == CV_32SC1 )
\r
372 idata = _train_data->data.i + vi0*dv_step;
\r
374 fdata = _train_data->data.fl + vi0*dv_step;
\r
375 if( _missing_mask )
\r
376 mask = _missing_mask->data.ptr + vi0*mv_step;
\r
378 else // analyze _responses
\r
380 ci = cat_var_count;
\r
381 step = CV_IS_MAT_CONT(_responses->type) ?
\r
382 1 : _responses->step / CV_ELEM_SIZE(_responses->type);
\r
383 if( CV_MAT_TYPE(_responses->type) == CV_32SC1 )
\r
384 idata = _responses->data.i;
\r
386 fdata = _responses->data.fl;
\r
389 if( (vi < var_count && ci>=0) ||
\r
390 (vi == var_count && is_classifier) ) // process categorical variable or response
\r
392 int c_count, prev_label;
\r
396 udst = (unsigned short*)(buf->data.s + vi*sample_count);
\r
398 idst = buf->data.i + vi*sample_count;
\r
401 for( i = 0; i < sample_count; i++ )
\r
403 int val = INT_MAX, si = sidx ? sidx[i] : i;
\r
404 if( !mask || !mask[si*m_step] )
\r
407 val = idata[si*step];
\r
410 float t = fdata[si*step];
\r
414 sprintf( err, "%d-th value of %d-th (categorical) "
\r
415 "variable is not an integer", i, vi );
\r
416 CV_ERROR( CV_StsBadArg, err );
\r
420 if( val == INT_MAX )
\r
422 sprintf( err, "%d-th value of %d-th (categorical) "
\r
423 "variable is too large", i, vi );
\r
424 CV_ERROR( CV_StsBadArg, err );
\r
431 pair16u32s_ptr[i].u = udst + i;
\r
432 pair16u32s_ptr[i].i = _idst + i;
\r
437 int_ptr[i] = idst + i;
\r
441 c_count = num_valid > 0;
\r
445 icvSortPairs( pair16u32s_ptr, sample_count, 0 );
\r
446 // count the categories
\r
447 for( i = 1; i < num_valid; i++ )
\r
448 if (*pair16u32s_ptr[i].i != *pair16u32s_ptr[i-1].i)
\r
453 icvSortIntPtr( int_ptr, sample_count, 0 );
\r
454 // count the categories
\r
455 for( i = 1; i < num_valid; i++ )
\r
456 c_count += *int_ptr[i] != *int_ptr[i-1];
\r
460 max_c_count = MAX( max_c_count, c_count );
\r
461 cat_count->data.i[ci] = c_count;
\r
462 cat_ofs->data.i[ci] = total_c_count;
\r
464 // resize cat_map, if need
\r
465 if( cat_map->cols < total_c_count + c_count )
\r
468 CV_CALL( cat_map = cvCreateMat( 1,
\r
469 MAX(cat_map->cols*3/2,total_c_count+c_count), CV_32SC1 ));
\r
470 for( i = 0; i < total_c_count; i++ )
\r
471 cat_map->data.i[i] = tmp_map->data.i[i];
\r
472 cvReleaseMat( &tmp_map );
\r
475 c_map = cat_map->data.i + total_c_count;
\r
476 total_c_count += c_count;
\r
481 // compact the class indices and build the map
\r
482 prev_label = ~*pair16u32s_ptr[0].i;
\r
483 for( i = 0; i < num_valid; i++ )
\r
485 int cur_label = *pair16u32s_ptr[i].i;
\r
486 if( cur_label != prev_label )
\r
487 c_map[++c_count] = prev_label = cur_label;
\r
488 *pair16u32s_ptr[i].u = (unsigned short)c_count;
\r
490 // replace labels for missing values with -1
\r
491 for( ; i < sample_count; i++ )
\r
492 *pair16u32s_ptr[i].u = 65535;
\r
496 // compact the class indices and build the map
\r
497 prev_label = ~*int_ptr[0];
\r
498 for( i = 0; i < num_valid; i++ )
\r
500 int cur_label = *int_ptr[i];
\r
501 if( cur_label != prev_label )
\r
502 c_map[++c_count] = prev_label = cur_label;
\r
503 *int_ptr[i] = c_count;
\r
505 // replace labels for missing values with -1
\r
506 for( ; i < sample_count; i++ )
\r
510 else if( ci < 0 ) // process ordered variable
\r
513 udst = (unsigned short*)(buf->data.s + vi*sample_count);
\r
515 idst = buf->data.i + vi*sample_count;
\r
517 for( i = 0; i < sample_count; i++ )
\r
519 float val = ord_nan;
\r
520 int si = sidx ? sidx[i] : i;
\r
521 if( !mask || !mask[si*m_step] )
\r
524 val = (float)idata[si*step];
\r
526 val = fdata[si*step];
\r
528 if( fabs(val) >= ord_nan )
\r
530 sprintf( err, "%d-th value of %d-th (ordered) "
\r
531 "variable (=%g) is too large", i, vi, val );
\r
532 CV_ERROR( CV_StsBadArg, err );
\r
537 udst[i] = (unsigned short)i;
\r
539 idst[i] = i; // ïåðåÃåñòè âûøå â if( idata )
\r
544 icvSortUShAux( udst, num_valid, _fdst);
\r
546 icvSortIntAux( idst, /*or num_valid?\*/ sample_count, _fdst );
\r
549 if( vi < var_count )
\r
550 data_root->set_num_valid(vi, num_valid);
\r
553 // set sample labels
\r
555 udst = (unsigned short*)(buf->data.s + work_var_count*sample_count);
\r
557 idst = buf->data.i + work_var_count*sample_count;
\r
559 for (i = 0; i < sample_count; i++)
\r
562 udst[i] = sidx ? (unsigned short)sidx[i] : (unsigned short)i;
\r
564 idst[i] = sidx ? sidx[i] : i;
\r
569 unsigned short* udst = 0;
\r
575 udst = (unsigned short*)(buf->data.s + (get_work_var_count()-1)*sample_count);
\r
576 for( i = vi = 0; i < sample_count; i++ )
\r
578 udst[i] = (unsigned short)vi++;
\r
579 vi &= vi < cv_n ? -1 : 0;
\r
582 for( i = 0; i < sample_count; i++ )
\r
584 int a = cvRandInt(r) % sample_count;
\r
585 int b = cvRandInt(r) % sample_count;
\r
586 unsigned short unsh = (unsigned short)vi;
\r
587 CV_SWAP( udst[a], udst[b], unsh );
\r
592 idst = buf->data.i + (get_work_var_count()-1)*sample_count;
\r
593 for( i = vi = 0; i < sample_count; i++ )
\r
596 vi &= vi < cv_n ? -1 : 0;
\r
599 for( i = 0; i < sample_count; i++ )
\r
601 int a = cvRandInt(r) % sample_count;
\r
602 int b = cvRandInt(r) % sample_count;
\r
603 CV_SWAP( idst[a], idst[b], vi );
\r
609 cat_map->cols = MAX( total_c_count, 1 );
\r
611 max_split_size = cvAlign(sizeof(CvDTreeSplit) +
\r
612 (MAX(0,max_c_count - 33)/32)*sizeof(int),sizeof(void*));
\r
613 CV_CALL( split_heap = cvCreateSet( 0, sizeof(*split_heap), max_split_size, tree_storage ));
\r
615 have_priors = is_classifier && params.priors;
\r
616 if( is_classifier )
\r
618 int m = get_num_classes();
\r
620 CV_CALL( priors = cvCreateMat( 1, m, CV_64F ));
\r
621 for( i = 0; i < m; i++ )
\r
623 double val = have_priors ? params.priors[i] : 1.;
\r
625 CV_ERROR( CV_StsOutOfRange, "Every class weight should be positive" );
\r
626 priors->data.db[i] = val;
\r
630 // normalize weights
\r
632 cvScale( priors, priors, 1./sum );
\r
634 CV_CALL( priors_mult = cvCloneMat( priors ));
\r
635 CV_CALL( counts = cvCreateMat( 1, m, CV_32SC1 ));
\r
639 CV_CALL( direction = cvCreateMat( 1, sample_count, CV_8UC1 ));
\r
640 CV_CALL( split_buf = cvCreateMat( 1, sample_count, CV_32SC1 ));
\r
642 CV_CALL( pred_float_buf = (float*)cvAlloc(sample_count*sizeof(pred_float_buf[0])) );
\r
643 CV_CALL( pred_int_buf = (int*)cvAlloc(sample_count*sizeof(pred_int_buf[0])) );
\r
644 CV_CALL( resp_float_buf = (float*)cvAlloc(sample_count*sizeof(resp_float_buf[0])) );
\r
645 CV_CALL( resp_int_buf = (int*)cvAlloc(sample_count*sizeof(resp_int_buf[0])) );
\r
646 CV_CALL( cv_lables_buf = (int*)cvAlloc(sample_count*sizeof(cv_lables_buf[0])) );
\r
647 CV_CALL( sample_idx_buf = (int*)cvAlloc(sample_count*sizeof(sample_idx_buf[0])) );
\r
658 cvFree( &int_ptr );
\r
659 cvReleaseMat( &var_type0 );
\r
660 cvReleaseMat( &sample_indices );
\r
661 cvReleaseMat( &tmp_map );
\r
666 void CvDTreeTrainData::do_responses_copy()
\r
668 responses_copy = cvCreateMat( responses->rows, responses->cols, responses->type );
\r
669 cvCopy( responses, responses_copy);
\r
670 responses = responses_copy;
\r
673 CvDTreeNode* CvDTreeTrainData::subsample_data( const CvMat* _subsample_idx )
\r
675 CvDTreeNode* root = 0;
\r
676 CvMat* isubsample_idx = 0;
\r
677 CvMat* subsample_co = 0;
\r
679 CV_FUNCNAME( "CvDTreeTrainData::subsample_data" );
\r
684 CV_ERROR( CV_StsError, "No training data has been set" );
\r
686 if( _subsample_idx )
\r
687 CV_CALL( isubsample_idx = cvPreprocessIndexArray( _subsample_idx, sample_count ));
\r
689 if( !isubsample_idx )
\r
691 // make a copy of the root node
\r
694 root = new_node( 0, 1, 0, 0 );
\r
696 *root = *data_root;
\r
697 root->num_valid = temp.num_valid;
\r
698 if( root->num_valid )
\r
700 for( i = 0; i < var_count; i++ )
\r
701 root->num_valid[i] = data_root->num_valid[i];
\r
703 root->cv_Tn = temp.cv_Tn;
\r
704 root->cv_node_risk = temp.cv_node_risk;
\r
705 root->cv_node_error = temp.cv_node_error;
\r
709 int* sidx = isubsample_idx->data.i;
\r
710 // co - array of count/offset pairs (to handle duplicated values in _subsample_idx)
\r
711 int* co, cur_ofs = 0;
\r
713 int work_var_count = get_work_var_count();
\r
714 int count = isubsample_idx->rows + isubsample_idx->cols - 1;
\r
716 root = new_node( 0, count, 1, 0 );
\r
718 CV_CALL( subsample_co = cvCreateMat( 1, sample_count*2, CV_32SC1 ));
\r
719 cvZero( subsample_co );
\r
720 co = subsample_co->data.i;
\r
721 for( i = 0; i < count; i++ )
\r
723 for( i = 0; i < sample_count; i++ )
\r
727 co[i*2+1] = cur_ofs;
\r
728 cur_ofs += co[i*2];
\r
734 for( vi = 0; vi < work_var_count; vi++ )
\r
736 int ci = get_var_type(vi);
\r
738 if( ci >= 0 || vi >= var_count )
\r
740 int* src_buf = pred_int_buf;
\r
741 const int* src = 0;
\r
744 get_cat_var_data( data_root, vi, src_buf, &src );
\r
748 unsigned short* udst = (unsigned short*)(buf->data.s + root->buf_idx*buf->cols +
\r
749 vi*sample_count + root->offset);
\r
750 for( i = 0; i < count; i++ )
\r
752 int val = src[sidx[i]];
\r
753 udst[i] = (unsigned short)val;
\r
754 num_valid += val >= 0;
\r
759 int* idst = buf->data.i + root->buf_idx*buf->cols +
\r
760 vi*sample_count + root->offset;
\r
761 for( i = 0; i < count; i++ )
\r
763 int val = src[sidx[i]];
\r
765 num_valid += val >= 0;
\r
769 if( vi < var_count )
\r
770 root->set_num_valid(vi, num_valid);
\r
774 int *src_idx_buf = pred_int_buf;
\r
775 const int* src_idx = 0;
\r
776 float *src_val_buf = pred_float_buf;
\r
777 const float* src_val = 0;
\r
778 int j = 0, idx, count_i;
\r
779 int num_valid = data_root->get_num_valid(vi);
\r
781 get_ord_var_data( data_root, vi, src_val_buf, src_idx_buf, &src_val, &src_idx );
\r
784 unsigned short* udst_idx = (unsigned short*)(buf->data.s + root->buf_idx*buf->cols +
\r
785 vi*sample_count + data_root->offset);
\r
786 for( i = 0; i < num_valid; i++ )
\r
789 count_i = co[idx*2];
\r
791 for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
\r
792 udst_idx[j] = (unsigned short)cur_ofs;
\r
795 root->set_num_valid(vi, j);
\r
797 for( ; i < sample_count; i++ )
\r
800 count_i = co[idx*2];
\r
802 for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
\r
803 udst_idx[j] = (unsigned short)cur_ofs;
\r
808 int* idst_idx = buf->data.i + root->buf_idx*buf->cols +
\r
809 vi*sample_count + root->offset;
\r
810 for( i = 0; i < num_valid; i++ )
\r
813 count_i = co[idx*2];
\r
815 for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
\r
816 idst_idx[j] = cur_ofs;
\r
819 root->set_num_valid(vi, j);
\r
821 for( ; i < sample_count; i++ )
\r
824 count_i = co[idx*2];
\r
826 for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
\r
827 idst_idx[j] = cur_ofs;
\r
832 // sample indices subsampling
\r
833 int* sample_idx_src_buf = sample_idx_buf;
\r
834 const int* sample_idx_src = 0;
\r
835 get_sample_indices(data_root, sample_idx_src_buf, &sample_idx_src);
\r
838 unsigned short* sample_idx_dst = (unsigned short*)(buf->data.s + root->buf_idx*buf->cols +
\r
839 get_work_var_count()*sample_count + root->offset);
\r
840 for (i = 0; i < count; i++)
\r
841 sample_idx_dst[i] = (unsigned short)sample_idx_src[sidx[i]];
\r
845 int* sample_idx_dst = buf->data.i + root->buf_idx*buf->cols +
\r
846 get_work_var_count()*sample_count + root->offset;
\r
847 for (i = 0; i < count; i++)
\r
848 sample_idx_dst[i] = sample_idx_src[sidx[i]];
\r
854 cvReleaseMat( &isubsample_idx );
\r
855 cvReleaseMat( &subsample_co );
\r
861 void CvDTreeTrainData::get_vectors( const CvMat* _subsample_idx,
\r
862 float* values, uchar* missing,
\r
863 float* responses, bool get_class_idx )
\r
865 CvMat* subsample_idx = 0;
\r
866 CvMat* subsample_co = 0;
\r
868 CV_FUNCNAME( "CvDTreeTrainData::get_vectors" );
\r
872 int i, vi, total = sample_count, count = total, cur_ofs = 0;
\r
876 if( _subsample_idx )
\r
878 CV_CALL( subsample_idx = cvPreprocessIndexArray( _subsample_idx, sample_count ));
\r
879 sidx = subsample_idx->data.i;
\r
880 CV_CALL( subsample_co = cvCreateMat( 1, sample_count*2, CV_32SC1 ));
\r
881 co = subsample_co->data.i;
\r
882 cvZero( subsample_co );
\r
883 count = subsample_idx->cols + subsample_idx->rows - 1;
\r
884 for( i = 0; i < count; i++ )
\r
886 for( i = 0; i < total; i++ )
\r
888 int count_i = co[i*2];
\r
891 co[i*2+1] = cur_ofs*var_count;
\r
892 cur_ofs += count_i;
\r
898 memset( missing, 1, count*var_count );
\r
900 for( vi = 0; vi < var_count; vi++ )
\r
902 int ci = get_var_type(vi);
\r
903 if( ci >= 0 ) // categorical
\r
905 float* dst = values + vi;
\r
906 uchar* m = missing ? missing + vi : 0;
\r
907 int* src_buf = pred_int_buf;
\r
908 const int* src = 0;
\r
909 get_cat_var_data(data_root, vi, src_buf, &src);
\r
911 for( i = 0; i < count; i++, dst += var_count )
\r
913 int idx = sidx ? sidx[i] : i;
\r
914 int val = src[idx];
\r
918 *m = (!is_buf_16u && val < 0) || (is_buf_16u && (val == 65535));
\r
925 float* dst = values + vi;
\r
926 uchar* m = missing ? missing + vi : 0;
\r
927 int count1 = data_root->get_num_valid(vi);
\r
928 float *src_val_buf = pred_float_buf;
\r
929 const float *src_val = 0;
\r
930 int* src_idx_buf = pred_int_buf;
\r
931 const int* src_idx = 0;
\r
932 get_ord_var_data(data_root, vi, src_val_buf, src_idx_buf, &src_val, &src_idx);
\r
934 for( i = 0; i < count1; i++ )
\r
936 int idx = src_idx[i];
\r
940 count_i = co[idx*2];
\r
941 cur_ofs = co[idx*2+1];
\r
944 cur_ofs = idx*var_count;
\r
947 float val = src_val[i];
\r
948 for( ; count_i > 0; count_i--, cur_ofs += var_count )
\r
950 dst[cur_ofs] = val;
\r
962 if( is_classifier )
\r
964 int* src_buf = resp_int_buf;
\r
965 const int* src = 0;
\r
966 get_class_labels(data_root, src_buf, &src);
\r
967 for( i = 0; i < count; i++ )
\r
969 int idx = sidx ? sidx[i] : i;
\r
970 int val = get_class_idx ? src[idx] :
\r
971 cat_map->data.i[cat_ofs->data.i[cat_var_count]+src[idx]];
\r
972 responses[i] = (float)val;
\r
977 float *_values_buf = resp_float_buf;
\r
978 const float* _values = 0;
\r
979 get_ord_responses(data_root, _values_buf, &_values);
\r
980 for( i = 0; i < count; i++ )
\r
982 int idx = sidx ? sidx[i] : i;
\r
983 responses[i] = _values[idx];
\r
990 cvReleaseMat( &subsample_idx );
\r
991 cvReleaseMat( &subsample_co );
\r
995 CvDTreeNode* CvDTreeTrainData::new_node( CvDTreeNode* parent, int count,
\r
996 int storage_idx, int offset )
\r
998 CvDTreeNode* node = (CvDTreeNode*)cvSetNew( node_heap );
\r
1000 node->sample_count = count;
\r
1001 node->depth = parent ? parent->depth + 1 : 0;
\r
1002 node->parent = parent;
\r
1003 node->left = node->right = 0;
\r
1006 node->class_idx = 0;
\r
1009 node->buf_idx = storage_idx;
\r
1010 node->offset = offset;
\r
1012 node->num_valid = (int*)cvSetNew( nv_heap );
\r
1014 node->num_valid = 0;
\r
1015 node->alpha = node->node_risk = node->tree_risk = node->tree_error = 0.;
\r
1016 node->complexity = 0;
\r
1018 if( params.cv_folds > 0 && cv_heap )
\r
1020 int cv_n = params.cv_folds;
\r
1021 node->Tn = INT_MAX;
\r
1022 node->cv_Tn = (int*)cvSetNew( cv_heap );
\r
1023 node->cv_node_risk = (double*)cvAlignPtr(node->cv_Tn + cv_n, sizeof(double));
\r
1024 node->cv_node_error = node->cv_node_risk + cv_n;
\r
1030 node->cv_node_risk = 0;
\r
1031 node->cv_node_error = 0;
\r
1038 CvDTreeSplit* CvDTreeTrainData::new_split_ord( int vi, float cmp_val,
\r
1039 int split_point, int inversed, float quality )
\r
1041 CvDTreeSplit* split = (CvDTreeSplit*)cvSetNew( split_heap );
\r
1042 split->var_idx = vi;
\r
1043 split->condensed_idx = INT_MIN;
\r
1044 split->ord.c = cmp_val;
\r
1045 split->ord.split_point = split_point;
\r
1046 split->inversed = inversed;
\r
1047 split->quality = quality;
\r
1054 CvDTreeSplit* CvDTreeTrainData::new_split_cat( int vi, float quality )
\r
1056 CvDTreeSplit* split = (CvDTreeSplit*)cvSetNew( split_heap );
\r
1057 int i, n = (max_c_count + 31)/32;
\r
1059 split->var_idx = vi;
\r
1060 split->condensed_idx = INT_MIN;
\r
1061 split->inversed = 0;
\r
1062 split->quality = quality;
\r
1063 for( i = 0; i < n; i++ )
\r
1064 split->subset[i] = 0;
\r
1071 void CvDTreeTrainData::free_node( CvDTreeNode* node )
\r
1073 CvDTreeSplit* split = node->split;
\r
1074 free_node_data( node );
\r
1077 CvDTreeSplit* next = split->next;
\r
1078 cvSetRemoveByPtr( split_heap, split );
\r
1082 cvSetRemoveByPtr( node_heap, node );
\r
1086 void CvDTreeTrainData::free_node_data( CvDTreeNode* node )
\r
1088 if( node->num_valid )
\r
1090 cvSetRemoveByPtr( nv_heap, node->num_valid );
\r
1091 node->num_valid = 0;
\r
1093 // do not free cv_* fields, as all the cross-validation related data is released at once.
\r
1097 void CvDTreeTrainData::free_train_data()
\r
1099 cvReleaseMat( &counts );
\r
1100 cvReleaseMat( &buf );
\r
1101 cvReleaseMat( &direction );
\r
1102 cvReleaseMat( &split_buf );
\r
1103 cvReleaseMemStorage( &temp_storage );
\r
1104 cvReleaseMat( &responses_copy );
\r
1105 cvFree( &pred_float_buf );
\r
1106 cvFree( &pred_int_buf );
\r
1107 cvFree( &resp_float_buf );
\r
1108 cvFree( &resp_int_buf );
\r
1109 cvFree( &cv_lables_buf );
\r
1110 cvFree( &sample_idx_buf );
\r
1112 cv_heap = nv_heap = 0;
\r
1116 void CvDTreeTrainData::clear()
\r
1118 free_train_data();
\r
1120 cvReleaseMemStorage( &tree_storage );
\r
1122 cvReleaseMat( &var_idx );
\r
1123 cvReleaseMat( &var_type );
\r
1124 cvReleaseMat( &cat_count );
\r
1125 cvReleaseMat( &cat_ofs );
\r
1126 cvReleaseMat( &cat_map );
\r
1127 cvReleaseMat( &priors );
\r
1128 cvReleaseMat( &priors_mult );
\r
1130 node_heap = split_heap = 0;
\r
1132 sample_count = var_all = var_count = max_c_count = ord_var_count = cat_var_count = 0;
\r
1133 have_labels = have_priors = is_classifier = false;
\r
1135 buf_count = buf_size = 0;
\r
1144 int CvDTreeTrainData::get_num_classes() const
\r
1146 return is_classifier ? cat_count->data.i[cat_var_count] : 0;
\r
1150 int CvDTreeTrainData::get_var_type(int vi) const
\r
1152 return var_type->data.i[vi];
\r
1155 int CvDTreeTrainData::get_ord_var_data( CvDTreeNode* n, int vi, float* ord_values_buf, int* indices_buf, const float** ord_values, const int** indices )
\r
1157 int vidx = var_idx ? var_idx->data.i[vi] : vi;
\r
1158 int node_sample_count = n->sample_count;
\r
1159 int* sample_indices_buf = sample_idx_buf;
\r
1160 const int* sample_indices = 0;
\r
1161 int td_step = train_data->step/CV_ELEM_SIZE(train_data->type);
\r
1163 get_sample_indices(n, sample_indices_buf, &sample_indices);
\r
1166 *indices = buf->data.i + n->buf_idx*buf->cols +
\r
1167 vi*sample_count + n->offset;
\r
1169 const unsigned short* short_indices = (const unsigned short*)(buf->data.s + n->buf_idx*buf->cols +
\r
1170 vi*sample_count + n->offset );
\r
1171 for( int i = 0; i < node_sample_count; i++ )
\r
1172 indices_buf[i] = short_indices[i];
\r
1173 *indices = indices_buf;
\r
1176 if( tflag == CV_ROW_SAMPLE )
\r
1178 for( int i = 0; i < node_sample_count &&
\r
1179 ((((*indices)[i] >= 0) && !is_buf_16u) || (((*indices)[i] != 65535) && is_buf_16u)); i++ )
\r
1181 int idx = (*indices)[i];
\r
1182 idx = sample_indices[idx];
\r
1183 ord_values_buf[i] = *(train_data->data.fl + idx * td_step + vidx);
\r
1187 for( int i = 0; i < node_sample_count &&
\r
1188 ((((*indices)[i] >= 0) && !is_buf_16u) || (((*indices)[i] != 65535) && is_buf_16u)); i++ )
\r
1190 int idx = (*indices)[i];
\r
1191 idx = sample_indices[idx];
\r
1192 ord_values_buf[i] = *(train_data->data.fl + vidx* td_step + idx);
\r
1195 *ord_values = ord_values_buf;
\r
1196 return 0; //TODO: return the number of non-missing values
\r
1200 void CvDTreeTrainData::get_class_labels( CvDTreeNode* n, int* labels_buf, const int** labels )
\r
1202 if (is_classifier)
\r
1203 get_cat_var_data( n, var_count, labels_buf, labels );
\r
1206 void CvDTreeTrainData::get_sample_indices( CvDTreeNode* n, int* indices_buf, const int** indices )
\r
1208 get_cat_var_data( n, get_work_var_count(), indices_buf, indices );
\r
1211 void CvDTreeTrainData::get_ord_responses( CvDTreeNode* n, float* values_buf, const float** values)
\r
1213 int sample_count = n->sample_count;
\r
1214 int* indices_buf = sample_idx_buf;
\r
1215 const int* indices = 0;
\r
1217 int r_step = responses->step/CV_ELEM_SIZE(responses->type);
\r
1219 get_sample_indices(n, indices_buf, &indices);
\r
1222 for( int i = 0; i < sample_count &&
\r
1223 (((indices[i] >= 0) && !is_buf_16u) || ((indices[i] != 65535) && is_buf_16u)); i++ )
\r
1225 int idx = indices[i];
\r
1226 values_buf[i] = *(responses->data.fl + idx * r_step);
\r
1229 *values = values_buf;
\r
1233 void CvDTreeTrainData::get_cv_labels( CvDTreeNode* n, int* labels_buf, const int** labels )
\r
1236 get_cat_var_data( n, get_work_var_count()- 1, labels_buf, labels );
\r
1240 int CvDTreeTrainData::get_cat_var_data( CvDTreeNode* n, int vi, int* cat_values_buf, const int** cat_values )
\r
1243 *cat_values = buf->data.i + n->buf_idx*buf->cols +
\r
1244 vi*sample_count + n->offset;
\r
1246 const unsigned short* short_values = (const unsigned short*)(buf->data.s + n->buf_idx*buf->cols +
\r
1247 vi*sample_count + n->offset);
\r
1248 for( int i = 0; i < n->sample_count; i++ )
\r
1249 cat_values_buf[i] = short_values[i];
\r
1250 *cat_values = cat_values_buf;
\r
1253 return 0; //TODO: return the number of non-missing values
\r
1257 int CvDTreeTrainData::get_child_buf_idx( CvDTreeNode* n )
\r
1259 int idx = n->buf_idx + 1;
\r
1260 if( idx >= buf_count )
\r
1261 idx = shared ? 1 : 0;
\r
1266 void CvDTreeTrainData::write_params( CvFileStorage* fs )
\r
1268 CV_FUNCNAME( "CvDTreeTrainData::write_params" );
\r
1272 int vi, vcount = var_count;
\r
1274 cvWriteInt( fs, "is_classifier", is_classifier ? 1 : 0 );
\r
1275 cvWriteInt( fs, "var_all", var_all );
\r
1276 cvWriteInt( fs, "var_count", var_count );
\r
1277 cvWriteInt( fs, "ord_var_count", ord_var_count );
\r
1278 cvWriteInt( fs, "cat_var_count", cat_var_count );
\r
1280 cvStartWriteStruct( fs, "training_params", CV_NODE_MAP );
\r
1281 cvWriteInt( fs, "use_surrogates", params.use_surrogates ? 1 : 0 );
\r
1283 if( is_classifier )
\r
1285 cvWriteInt( fs, "max_categories", params.max_categories );
\r
1289 cvWriteReal( fs, "regression_accuracy", params.regression_accuracy );
\r
1292 cvWriteInt( fs, "max_depth", params.max_depth );
\r
1293 cvWriteInt( fs, "min_sample_count", params.min_sample_count );
\r
1294 cvWriteInt( fs, "cross_validation_folds", params.cv_folds );
\r
1296 if( params.cv_folds > 1 )
\r
1298 cvWriteInt( fs, "use_1se_rule", params.use_1se_rule ? 1 : 0 );
\r
1299 cvWriteInt( fs, "truncate_pruned_tree", params.truncate_pruned_tree ? 1 : 0 );
\r
1303 cvWrite( fs, "priors", priors );
\r
1305 cvEndWriteStruct( fs );
\r
1308 cvWrite( fs, "var_idx", var_idx );
\r
1310 cvStartWriteStruct( fs, "var_type", CV_NODE_SEQ+CV_NODE_FLOW );
\r
1312 for( vi = 0; vi < vcount; vi++ )
\r
1313 cvWriteInt( fs, 0, var_type->data.i[vi] >= 0 );
\r
1315 cvEndWriteStruct( fs );
\r
1317 if( cat_count && (cat_var_count > 0 || is_classifier) )
\r
1319 CV_ASSERT( cat_count != 0 );
\r
1320 cvWrite( fs, "cat_count", cat_count );
\r
1321 cvWrite( fs, "cat_map", cat_map );
\r
1328 void CvDTreeTrainData::read_params( CvFileStorage* fs, CvFileNode* node )
\r
1330 CV_FUNCNAME( "CvDTreeTrainData::read_params" );
\r
1334 CvFileNode *tparams_node, *vartype_node;
\r
1335 CvSeqReader reader;
\r
1336 int vi, max_split_size, tree_block_size;
\r
1338 is_classifier = (cvReadIntByName( fs, node, "is_classifier" ) != 0);
\r
1339 var_all = cvReadIntByName( fs, node, "var_all" );
\r
1340 var_count = cvReadIntByName( fs, node, "var_count", var_all );
\r
1341 cat_var_count = cvReadIntByName( fs, node, "cat_var_count" );
\r
1342 ord_var_count = cvReadIntByName( fs, node, "ord_var_count" );
\r
1344 tparams_node = cvGetFileNodeByName( fs, node, "training_params" );
\r
1346 if( tparams_node ) // training parameters are not necessary
\r
1348 params.use_surrogates = cvReadIntByName( fs, tparams_node, "use_surrogates", 1 ) != 0;
\r
1350 if( is_classifier )
\r
1352 params.max_categories = cvReadIntByName( fs, tparams_node, "max_categories" );
\r
1356 params.regression_accuracy =
\r
1357 (float)cvReadRealByName( fs, tparams_node, "regression_accuracy" );
\r
1360 params.max_depth = cvReadIntByName( fs, tparams_node, "max_depth" );
\r
1361 params.min_sample_count = cvReadIntByName( fs, tparams_node, "min_sample_count" );
\r
1362 params.cv_folds = cvReadIntByName( fs, tparams_node, "cross_validation_folds" );
\r
1364 if( params.cv_folds > 1 )
\r
1366 params.use_1se_rule = cvReadIntByName( fs, tparams_node, "use_1se_rule" ) != 0;
\r
1367 params.truncate_pruned_tree =
\r
1368 cvReadIntByName( fs, tparams_node, "truncate_pruned_tree" ) != 0;
\r
1371 priors = (CvMat*)cvReadByName( fs, tparams_node, "priors" );
\r
1374 if( !CV_IS_MAT(priors) )
\r
1375 CV_ERROR( CV_StsParseError, "priors must stored as a matrix" );
\r
1376 priors_mult = cvCloneMat( priors );
\r
1380 CV_CALL( var_idx = (CvMat*)cvReadByName( fs, node, "var_idx" ));
\r
1383 if( !CV_IS_MAT(var_idx) ||
\r
1384 (var_idx->cols != 1 && var_idx->rows != 1) ||
\r
1385 var_idx->cols + var_idx->rows - 1 != var_count ||
\r
1386 CV_MAT_TYPE(var_idx->type) != CV_32SC1 )
\r
1387 CV_ERROR( CV_StsParseError,
\r
1388 "var_idx (if exist) must be valid 1d integer vector containing <var_count> elements" );
\r
1390 for( vi = 0; vi < var_count; vi++ )
\r
1391 if( (unsigned)var_idx->data.i[vi] >= (unsigned)var_all )
\r
1392 CV_ERROR( CV_StsOutOfRange, "some of var_idx elements are out of range" );
\r
1395 ////// read var type
\r
1396 CV_CALL( var_type = cvCreateMat( 1, var_count + 2, CV_32SC1 ));
\r
1398 cat_var_count = 0;
\r
1399 ord_var_count = -1;
\r
1400 vartype_node = cvGetFileNodeByName( fs, node, "var_type" );
\r
1402 if( vartype_node && CV_NODE_TYPE(vartype_node->tag) == CV_NODE_INT && var_count == 1 )
\r
1403 var_type->data.i[0] = vartype_node->data.i ? cat_var_count++ : ord_var_count--;
\r
1406 if( !vartype_node || CV_NODE_TYPE(vartype_node->tag) != CV_NODE_SEQ ||
\r
1407 vartype_node->data.seq->total != var_count )
\r
1408 CV_ERROR( CV_StsParseError, "var_type must exist and be a sequence of 0's and 1's" );
\r
1410 cvStartReadSeq( vartype_node->data.seq, &reader );
\r
1412 for( vi = 0; vi < var_count; vi++ )
\r
1414 CvFileNode* n = (CvFileNode*)reader.ptr;
\r
1415 if( CV_NODE_TYPE(n->tag) != CV_NODE_INT || (n->data.i & ~1) )
\r
1416 CV_ERROR( CV_StsParseError, "var_type must exist and be a sequence of 0's and 1's" );
\r
1417 var_type->data.i[vi] = n->data.i ? cat_var_count++ : ord_var_count--;
\r
1418 CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
\r
1421 var_type->data.i[var_count] = cat_var_count;
\r
1423 ord_var_count = ~ord_var_count;
\r
1424 if( cat_var_count != cat_var_count || ord_var_count != ord_var_count )
\r
1425 CV_ERROR( CV_StsParseError, "var_type is inconsistent with cat_var_count and ord_var_count" );
\r
1428 if( cat_var_count > 0 || is_classifier )
\r
1430 int ccount, total_c_count = 0;
\r
1431 CV_CALL( cat_count = (CvMat*)cvReadByName( fs, node, "cat_count" ));
\r
1432 CV_CALL( cat_map = (CvMat*)cvReadByName( fs, node, "cat_map" ));
\r
1434 if( !CV_IS_MAT(cat_count) || !CV_IS_MAT(cat_map) ||
\r
1435 (cat_count->cols != 1 && cat_count->rows != 1) ||
\r
1436 CV_MAT_TYPE(cat_count->type) != CV_32SC1 ||
\r
1437 cat_count->cols + cat_count->rows - 1 != cat_var_count + is_classifier ||
\r
1438 (cat_map->cols != 1 && cat_map->rows != 1) ||
\r
1439 CV_MAT_TYPE(cat_map->type) != CV_32SC1 )
\r
1440 CV_ERROR( CV_StsParseError,
\r
1441 "Both cat_count and cat_map must exist and be valid 1d integer vectors of an appropriate size" );
\r
1443 ccount = cat_var_count + is_classifier;
\r
1445 CV_CALL( cat_ofs = cvCreateMat( 1, ccount + 1, CV_32SC1 ));
\r
1446 cat_ofs->data.i[0] = 0;
\r
1449 for( vi = 0; vi < ccount; vi++ )
\r
1451 int val = cat_count->data.i[vi];
\r
1453 CV_ERROR( CV_StsOutOfRange, "some of cat_count elements are out of range" );
\r
1454 max_c_count = MAX( max_c_count, val );
\r
1455 cat_ofs->data.i[vi+1] = total_c_count += val;
\r
1458 if( cat_map->cols + cat_map->rows - 1 != total_c_count )
\r
1459 CV_ERROR( CV_StsBadSize,
\r
1460 "cat_map vector length is not equal to the total number of categories in all categorical vars" );
\r
1463 max_split_size = cvAlign(sizeof(CvDTreeSplit) +
\r
1464 (MAX(0,max_c_count - 33)/32)*sizeof(int),sizeof(void*));
\r
1466 tree_block_size = MAX((int)sizeof(CvDTreeNode)*8, max_split_size);
\r
1467 tree_block_size = MAX(tree_block_size + block_size_delta, min_block_size);
\r
1468 CV_CALL( tree_storage = cvCreateMemStorage( tree_block_size ));
\r
1469 CV_CALL( node_heap = cvCreateSet( 0, sizeof(node_heap[0]),
\r
1470 sizeof(CvDTreeNode), tree_storage ));
\r
1471 CV_CALL( split_heap = cvCreateSet( 0, sizeof(split_heap[0]),
\r
1472 max_split_size, tree_storage ));
\r
1476 /////////////////////// Decision Tree /////////////////////////
\r
1478 CvDTree::CvDTree()
\r
1481 var_importance = 0;
\r
1482 default_model_name = "my_tree";
\r
1488 void CvDTree::clear()
\r
1490 cvReleaseMat( &var_importance );
\r
1493 if( !data->shared )
\r
1500 pruned_tree_idx = -1;
\r
1504 CvDTree::~CvDTree()
\r
1510 const CvDTreeNode* CvDTree::get_root() const
\r
1516 int CvDTree::get_pruned_tree_idx() const
\r
1518 return pruned_tree_idx;
\r
1522 CvDTreeTrainData* CvDTree::get_data()
\r
1528 bool CvDTree::train( const CvMat* _train_data, int _tflag,
\r
1529 const CvMat* _responses, const CvMat* _var_idx,
\r
1530 const CvMat* _sample_idx, const CvMat* _var_type,
\r
1531 const CvMat* _missing_mask, CvDTreeParams _params )
\r
1533 bool result = false;
\r
1535 CV_FUNCNAME( "CvDTree::train" );
\r
1540 data = new CvDTreeTrainData( _train_data, _tflag, _responses,
\r
1541 _var_idx, _sample_idx, _var_type,
\r
1542 _missing_mask, _params, false );
\r
1543 CV_CALL( result = do_train(0) );
\r
1550 bool CvDTree::train( CvMLData* _data, CvDTreeParams _params )
\r
1552 bool result = false;
\r
1554 CV_FUNCNAME( "CvDTree::train" );
\r
1558 const CvMat* values = _data->get_values();
1559 const CvMat* response = _data->get_response();
1560 const CvMat* missing = _data->get_missing();
1561 const CvMat* var_types = _data->get_var_types();
1562 const CvMat* train_sidx = _data->get_train_sample_idx();
1563 const CvMat* var_idx = _data->get_var_idx();
\r
1565 CV_CALL( result = train( values, CV_ROW_SAMPLE, response, var_idx,
\r
1566 train_sidx, var_types, missing, _params ) );
\r
1573 bool CvDTree::train( CvDTreeTrainData* _data, const CvMat* _subsample_idx )
\r
1575 bool result = false;
\r
1577 CV_FUNCNAME( "CvDTree::train" );
\r
1583 data->shared = true;
\r
1584 CV_CALL( result = do_train(_subsample_idx));
\r
1592 bool CvDTree::do_train( const CvMat* _subsample_idx )
\r
1594 bool result = false;
\r
1596 CV_FUNCNAME( "CvDTree::do_train" );
\r
1600 root = data->subsample_data( _subsample_idx );
\r
1602 CV_CALL( try_split_node(root));
\r
1604 if( data->params.cv_folds > 0 )
\r
1605 CV_CALL( prune_cv());
\r
1607 if( !data->shared )
\r
1608 data->free_train_data();
\r
1618 void CvDTree::try_split_node( CvDTreeNode* node )
\r
1620 CvDTreeSplit* best_split = 0;
\r
1621 int i, n = node->sample_count, vi;
\r
1622 bool can_split = true;
\r
1623 double quality_scale;
\r
1625 calc_node_value( node );
\r
1627 if( node->sample_count <= data->params.min_sample_count ||
\r
1628 node->depth >= data->params.max_depth )
\r
1629 can_split = false;
\r
1631 if( can_split && data->is_classifier )
\r
1633 // check if we have a "pure" node,
\r
1634 // we assume that cls_count is filled by calc_node_value()
\r
1635 int* cls_count = data->counts->data.i;
\r
1636 int nz = 0, m = data->get_num_classes();
\r
1637 for( i = 0; i < m; i++ )
\r
1638 nz += cls_count[i] != 0;
\r
1639 if( nz == 1 ) // there is only one class
\r
1640 can_split = false;
\r
1642 else if( can_split )
\r
1644 if( sqrt(node->node_risk)/n < data->params.regression_accuracy )
\r
1645 can_split = false;
\r
1650 best_split = find_best_split(node);
\r
1651 // TODO: check the split quality ...
\r
1652 node->split = best_split;
\r
1655 if( !can_split || !best_split )
\r
1657 data->free_node_data(node);
\r
1661 quality_scale = calc_node_dir( node );
\r
1663 if( data->params.use_surrogates )
\r
1665 // find all the surrogate splits
\r
1666 // and sort them by their similarity to the primary one
\r
1667 for( vi = 0; vi < data->var_count; vi++ )
\r
1669 CvDTreeSplit* split;
\r
1670 int ci = data->get_var_type(vi);
\r
1672 if( vi == best_split->var_idx )
\r
1676 split = find_surrogate_split_cat( node, vi );
\r
1678 split = find_surrogate_split_ord( node, vi );
\r
1682 // insert the split
\r
1683 CvDTreeSplit* prev_split = node->split;
\r
1684 split->quality = (float)(split->quality*quality_scale);
\r
1686 while( prev_split->next &&
\r
1687 prev_split->next->quality > split->quality )
\r
1688 prev_split = prev_split->next;
\r
1689 split->next = prev_split->next;
\r
1690 prev_split->next = split;
\r
1695 split_node_data( node );
\r
1696 try_split_node( node->left );
\r
1697 try_split_node( node->right );
\r
1701 // calculate direction (left(-1),right(1),missing(0))
\r
1702 // for each sample using the best split
\r
1703 // the function returns scale coefficients for surrogate split quality factors.
\r
1704 // the scale is applied to normalize surrogate split quality relatively to the
\r
1705 // best (primary) split quality. That is, if a surrogate split is absolutely
\r
1706 // identical to the primary split, its quality will be set to the maximum value =
\r
1707 // quality of the primary split; otherwise, it will be lower.
\r
1708 // besides, the function compute node->maxlr,
\r
1709 // minimum possible quality (w/o considering the above mentioned scale)
\r
1710 // for a surrogate split. Surrogate splits with quality less than node->maxlr
\r
1711 // are not discarded.
\r
1712 double CvDTree::calc_node_dir( CvDTreeNode* node )
\r
1714 char* dir = (char*)data->direction->data.ptr;
\r
1715 int i, n = node->sample_count, vi = node->split->var_idx;
\r
1718 assert( !node->split->inversed );
\r
1720 if( data->get_var_type(vi) >= 0 ) // split on categorical var
\r
1722 int* labels_buf = data->pred_int_buf;
\r
1723 const int* labels = 0;
\r
1724 const int* subset = node->split->subset;
\r
1725 data->get_cat_var_data( node, vi, labels_buf, &labels );
\r
1726 if( !data->have_priors )
\r
1728 int sum = 0, sum_abs = 0;
\r
1730 for( i = 0; i < n; i++ )
\r
1732 int idx = labels[i];
\r
1733 int d = ( ((idx >= 0)&&(!data->is_buf_16u)) || ((idx != 65535)&&(data->is_buf_16u)) ) ?
\r
1734 CV_DTREE_CAT_DIR(idx,subset) : 0;
\r
1735 sum += d; sum_abs += d & 1;
\r
1739 R = (sum_abs + sum) >> 1;
\r
1740 L = (sum_abs - sum) >> 1;
\r
1744 const double* priors = data->priors_mult->data.db;
\r
1745 double sum = 0, sum_abs = 0;
\r
1746 int *responses_buf = data->resp_int_buf;
\r
1747 const int* responses;
\r
1748 data->get_class_labels(node, responses_buf, &responses);
\r
1750 for( i = 0; i < n; i++ )
\r
1752 int idx = labels[i];
\r
1753 double w = priors[responses[i]];
\r
1754 int d = idx >= 0 ? CV_DTREE_CAT_DIR(idx,subset) : 0;
\r
1755 sum += d*w; sum_abs += (d & 1)*w;
\r
1759 R = (sum_abs + sum) * 0.5;
\r
1760 L = (sum_abs - sum) * 0.5;
\r
1763 else // split on ordered var
\r
1765 int split_point = node->split->ord.split_point;
\r
1766 int n1 = node->get_num_valid(vi);
\r
1768 float* val_buf = data->pred_float_buf;
\r
1769 const float* val = 0;
\r
1770 int* sorted_buf = data->pred_int_buf;
\r
1771 const int* sorted = 0;
\r
1772 data->get_ord_var_data( node, vi, val_buf, sorted_buf, &val, &sorted);
\r
1774 assert( 0 <= split_point && split_point < n1-1 );
\r
1776 if( !data->have_priors )
\r
1778 for( i = 0; i <= split_point; i++ )
\r
1779 dir[sorted[i]] = (char)-1;
\r
1780 for( ; i < n1; i++ )
\r
1781 dir[sorted[i]] = (char)1;
\r
1782 for( ; i < n; i++ )
\r
1783 dir[sorted[i]] = (char)0;
\r
1785 L = split_point-1;
\r
1786 R = n1 - split_point + 1;
\r
1790 const double* priors = data->priors_mult->data.db;
\r
1791 int* responses_buf = data->resp_int_buf;
\r
1792 const int* responses = 0;
\r
1793 data->get_class_labels(node, responses_buf, &responses);
\r
1796 for( i = 0; i <= split_point; i++ )
\r
1798 int idx = sorted[i];
\r
1799 double w = priors[responses[idx]];
\r
1800 dir[idx] = (char)-1;
\r
1804 for( ; i < n1; i++ )
\r
1806 int idx = sorted[i];
\r
1807 double w = priors[responses[idx]];
\r
1808 dir[idx] = (char)1;
\r
1812 for( ; i < n; i++ )
\r
1813 dir[sorted[i]] = (char)0;
\r
1817 node->maxlr = MAX( L, R );
\r
1818 return node->split->quality/(L + R);
\r
1822 CvDTreeSplit* CvDTree::find_best_split( CvDTreeNode* node )
\r
1825 CvDTreeSplit *best_split = 0, *split = 0, *t;
\r
1827 for( vi = 0; vi < data->var_count; vi++ )
\r
1829 int ci = data->get_var_type(vi);
\r
1830 if( node->get_num_valid(vi) <= 1 )
\r
1833 if( data->is_classifier )
\r
1836 split = find_split_cat_class( node, vi );
\r
1838 split = find_split_ord_class( node, vi );
\r
1843 split = find_split_cat_reg( node, vi );
\r
1845 split = find_split_ord_reg( node, vi );
\r
1850 if( !best_split || best_split->quality < split->quality )
\r
1851 CV_SWAP( best_split, split, t );
\r
1853 cvSetRemoveByPtr( data->split_heap, split );
\r
1857 return best_split;
\r
1861 CvDTreeSplit* CvDTree::find_split_ord_class( CvDTreeNode* node, int vi )
\r
1863 const float epsilon = FLT_EPSILON*2;
\r
1864 int n = node->sample_count;
\r
1865 int n1 = node->get_num_valid(vi);
\r
1866 int m = data->get_num_classes();
\r
1868 float* values_buf = data->pred_float_buf;
\r
1869 const float* values = 0;
\r
1870 int* indices_buf = data->pred_int_buf;
\r
1871 const int* indices = 0;
\r
1872 data->get_ord_var_data( node, vi, values_buf, indices_buf, &values, &indices );
\r
1873 int* responses_buf = data->resp_int_buf;
\r
1874 const int* responses = 0;
\r
1875 data->get_class_labels( node, responses_buf, &responses );
\r
1877 const int* rc0 = data->counts->data.i;
\r
1878 int* lc = (int*)cvStackAlloc(m*sizeof(lc[0]));
\r
1879 int* rc = (int*)cvStackAlloc(m*sizeof(rc[0]));
\r
1880 int i, best_i = -1;
\r
1881 double lsum2 = 0, rsum2 = 0, best_val = 0;
\r
1882 const double* priors = data->have_priors ? data->priors_mult->data.db : 0;
\r
1884 // init arrays of class instance counters on both sides of the split
\r
1885 for( i = 0; i < m; i++ )
\r
1891 // compensate for missing values
\r
1892 for( i = n1; i < n; i++ )
\r
1894 rc[responses[indices[i]]]--;
\r
1899 int L = 0, R = n1;
\r
1901 for( i = 0; i < m; i++ )
\r
1902 rsum2 += (double)rc[i]*rc[i];
\r
1904 for( i = 0; i < n1 - 1; i++ )
\r
1906 int idx = responses[indices[i]];
\r
1909 lv = lc[idx]; rv = rc[idx];
\r
1910 lsum2 += lv*2 + 1;
\r
1911 rsum2 -= rv*2 - 1;
\r
1912 lc[idx] = lv + 1; rc[idx] = rv - 1;
\r
1914 if( values[i] + epsilon < values[i+1] )
\r
1916 double val = (lsum2*R + rsum2*L)/((double)L*R);
\r
1917 if( best_val < val )
\r
1927 double L = 0, R = 0;
\r
1928 for( i = 0; i < m; i++ )
\r
1930 double wv = rc[i]*priors[i];
\r
1935 for( i = 0; i < n1 - 1; i++ )
\r
1937 int idx = responses[indices[i]];
\r
1939 double p = priors[idx], p2 = p*p;
\r
1941 lv = lc[idx]; rv = rc[idx];
\r
1942 lsum2 += p2*(lv*2 + 1);
\r
1943 rsum2 -= p2*(rv*2 - 1);
\r
1944 lc[idx] = lv + 1; rc[idx] = rv - 1;
\r
1946 if( values[i] + epsilon < values[i+1] )
\r
1948 double val = (lsum2*R + rsum2*L)/((double)L*R);
\r
1949 if( best_val < val )
\r
1958 return best_i >= 0 ? data->new_split_ord( vi,
\r
1959 (values[best_i] + values[best_i+1])*0.5f, best_i,
\r
1960 0, (float)best_val ) : 0;
\r
1964 void CvDTree::cluster_categories( const int* vectors, int n, int m,
\r
1965 int* csums, int k, int* labels )
\r
1967 // TODO: consider adding priors (class weights) and sample weights to the clustering algorithm
\r
1968 int iters = 0, max_iters = 100;
\r
1970 double* buf = (double*)cvStackAlloc( (n + k)*sizeof(buf[0]) );
\r
1971 double *v_weights = buf, *c_weights = buf + k;
\r
1972 bool modified = true;
\r
1973 CvRNG* r = &data->rng;
\r
1975 // assign labels randomly
\r
1976 for( i = idx = 0; i < n; i++ )
\r
1979 const int* v = vectors + i*m;
\r
1980 labels[i] = idx++;
\r
1981 idx &= idx < k ? -1 : 0;
\r
1983 // compute weight of each vector
\r
1984 for( j = 0; j < m; j++ )
\r
1986 v_weights[i] = sum ? 1./sum : 0.;
\r
1989 for( i = 0; i < n; i++ )
\r
1991 int i1 = cvRandInt(r) % n;
\r
1992 int i2 = cvRandInt(r) % n;
\r
1993 CV_SWAP( labels[i1], labels[i2], j );
\r
1996 for( iters = 0; iters <= max_iters; iters++ )
\r
1998 // calculate csums
\r
1999 for( i = 0; i < k; i++ )
\r
2001 for( j = 0; j < m; j++ )
\r
2002 csums[i*m + j] = 0;
\r
2005 for( i = 0; i < n; i++ )
\r
2007 const int* v = vectors + i*m;
\r
2008 int* s = csums + labels[i]*m;
\r
2009 for( j = 0; j < m; j++ )
\r
2013 // exit the loop here, when we have up-to-date csums
\r
2014 if( iters == max_iters || !modified )
\r
2019 // calculate weight of each cluster
\r
2020 for( i = 0; i < k; i++ )
\r
2022 const int* s = csums + i*m;
\r
2024 for( j = 0; j < m; j++ )
\r
2026 c_weights[i] = sum ? 1./sum : 0;
\r
2029 // now for each vector determine the closest cluster
\r
2030 for( i = 0; i < n; i++ )
\r
2032 const int* v = vectors + i*m;
\r
2033 double alpha = v_weights[i];
\r
2034 double min_dist2 = DBL_MAX;
\r
2037 for( idx = 0; idx < k; idx++ )
\r
2039 const int* s = csums + idx*m;
\r
2040 double dist2 = 0., beta = c_weights[idx];
\r
2041 for( j = 0; j < m; j++ )
\r
2043 double t = v[j]*alpha - s[j]*beta;
\r
2046 if( min_dist2 > dist2 )
\r
2048 min_dist2 = dist2;
\r
2053 if( min_idx != labels[i] )
\r
2055 labels[i] = min_idx;
\r
2061 CvDTreeSplit* CvDTree::find_split_cat_class( CvDTreeNode* node, int vi )
\r
2063 CvDTreeSplit* split = 0;
\r
2064 int ci = data->get_var_type(vi);
\r
2065 int n = node->sample_count;
\r
2066 int m = data->get_num_classes();
\r
2067 int _mi = data->cat_count->data.i[ci], mi = _mi;
\r
2069 int* labels_buf = data->pred_int_buf;
\r
2070 const int* labels = 0;
\r
2071 data->get_cat_var_data(node, vi, labels_buf, &labels);
\r
2072 int *responses_buf = data->resp_int_buf;
\r
2073 const int* responses = 0;
\r
2074 data->get_class_labels(node, responses_buf, &responses);
\r
2076 int* lc = (int*)cvStackAlloc(m*sizeof(lc[0]));
\r
2077 int* rc = (int*)cvStackAlloc(m*sizeof(rc[0]));
\r
2078 int* _cjk = (int*)cvStackAlloc(m*(mi+1)*sizeof(_cjk[0]))+m, *cjk = _cjk;
\r
2079 double* c_weights = (double*)cvStackAlloc( mi*sizeof(c_weights[0]) );
\r
2080 int* cluster_labels = 0;
\r
2081 int** int_ptr = 0;
\r
2083 double L = 0, R = 0;
\r
2084 double best_val = 0;
\r
2085 int prevcode = 0, best_subset = -1, subset_i, subset_n, subtract = 0;
\r
2086 const double* priors = data->priors_mult->data.db;
\r
2088 // init array of counters:
\r
2089 // c_{jk} - number of samples that have vi-th input variable = j and response = k.
\r
2090 for( j = -1; j < mi; j++ )
\r
2091 for( k = 0; k < m; k++ )
\r
2094 for( i = 0; i < n; i++ )
\r
2096 j = ( labels[i] == 65535 && data->is_buf_16u) ? -1 : labels[i];
\r
2103 if( mi > data->params.max_categories )
\r
2105 mi = MIN(data->params.max_categories, n);
\r
2107 cluster_labels = (int*)cvStackAlloc(mi*sizeof(cluster_labels[0]));
\r
2108 cluster_categories( _cjk, _mi, m, cjk, mi, cluster_labels );
\r
2111 subset_n = 1 << mi;
\r
2116 int_ptr = (int**)cvStackAlloc( mi*sizeof(int_ptr[0]) );
\r
2117 for( j = 0; j < mi; j++ )
\r
2118 int_ptr[j] = cjk + j*2 + 1;
\r
2119 icvSortIntPtr( int_ptr, mi, 0 );
\r
2124 for( k = 0; k < m; k++ )
\r
2127 for( j = 0; j < mi; j++ )
\r
2128 sum += cjk[j*m + k];
\r
2133 for( j = 0; j < mi; j++ )
\r
2136 for( k = 0; k < m; k++ )
\r
2137 sum += cjk[j*m + k]*priors[k];
\r
2138 c_weights[j] = sum;
\r
2139 R += c_weights[j];
\r
2142 for( ; subset_i < subset_n; subset_i++ )
\r
2146 double lsum2 = 0, rsum2 = 0;
\r
2149 idx = (int)(int_ptr[subset_i] - cjk)/2;
\r
2152 int graycode = (subset_i>>1)^subset_i;
\r
2153 int diff = graycode ^ prevcode;
\r
2155 // determine index of the changed bit.
\r
2157 idx = diff >= (1 << 16) ? 16 : 0;
\r
2158 u.f = (float)(((diff >> 16) | diff) & 65535);
\r
2159 idx += (u.i >> 23) - 127;
\r
2160 subtract = graycode < prevcode;
\r
2161 prevcode = graycode;
\r
2164 crow = cjk + idx*m;
\r
2165 weight = c_weights[idx];
\r
2166 if( weight < FLT_EPSILON )
\r
2171 for( k = 0; k < m; k++ )
\r
2174 int lval = lc[k] + t;
\r
2175 int rval = rc[k] - t;
\r
2176 double p = priors[k], p2 = p*p;
\r
2177 lsum2 += p2*lval*lval;
\r
2178 rsum2 += p2*rval*rval;
\r
2179 lc[k] = lval; rc[k] = rval;
\r
2186 for( k = 0; k < m; k++ )
\r
2189 int lval = lc[k] - t;
\r
2190 int rval = rc[k] + t;
\r
2191 double p = priors[k], p2 = p*p;
\r
2192 lsum2 += p2*lval*lval;
\r
2193 rsum2 += p2*rval*rval;
\r
2194 lc[k] = lval; rc[k] = rval;
\r
2200 if( L > FLT_EPSILON && R > FLT_EPSILON )
\r
2202 double val = (lsum2*R + rsum2*L)/((double)L*R);
\r
2203 if( best_val < val )
\r
2206 best_subset = subset_i;
\r
2211 if( best_subset < 0 )
\r
2214 split = data->new_split_cat( vi, (float)best_val );
\r
2218 for( i = 0; i <= best_subset; i++ )
\r
2220 idx = (int)(int_ptr[i] - cjk) >> 1;
\r
2221 split->subset[idx >> 5] |= 1 << (idx & 31);
\r
2226 for( i = 0; i < _mi; i++ )
\r
2228 idx = cluster_labels ? cluster_labels[i] : i;
\r
2229 if( best_subset & (1 << idx) )
\r
2230 split->subset[i >> 5] |= 1 << (i & 31);
\r
2238 CvDTreeSplit* CvDTree::find_split_ord_reg( CvDTreeNode* node, int vi )
\r
2240 const float epsilon = FLT_EPSILON*2;
\r
2241 int n = node->sample_count;
\r
2242 int n1 = node->get_num_valid(vi);
\r
2244 float* values_buf = data->pred_float_buf;
\r
2245 const float* values = 0;
\r
2246 int* indices_buf = data->pred_int_buf;
\r
2247 const int* indices = 0;
\r
2248 data->get_ord_var_data( node, vi, values_buf, indices_buf, &values, &indices );
\r
2249 float* responses_buf = data->resp_float_buf;
\r
2250 const float* responses = 0;
\r
2251 data->get_ord_responses( node, responses_buf, &responses );
\r
2253 int i, best_i = -1;
\r
2254 double best_val = 0, lsum = 0, rsum = node->value*n;
\r
2255 int L = 0, R = n1;
\r
2257 // compensate for missing values
\r
2258 for( i = n1; i < n; i++ )
\r
2259 rsum -= responses[indices[i]];
\r
2261 // find the optimal split
\r
2262 for( i = 0; i < n1 - 1; i++ )
\r
2264 float t = responses[indices[i]];
\r
2269 if( values[i] + epsilon < values[i+1] )
\r
2271 double val = (lsum*lsum*R + rsum*rsum*L)/((double)L*R);
\r
2272 if( best_val < val )
\r
2280 return best_i >= 0 ? data->new_split_ord( vi,
\r
2281 (values[best_i] + values[best_i+1])*0.5f, best_i,
\r
2282 0, (float)best_val ) : 0;
\r
2286 CvDTreeSplit* CvDTree::find_split_cat_reg( CvDTreeNode* node, int vi )
\r
2288 CvDTreeSplit* split;
\r
2289 int ci = data->get_var_type(vi);
\r
2290 int n = node->sample_count;
\r
2291 int mi = data->cat_count->data.i[ci];
\r
2292 int* labels_buf = data->pred_int_buf;
\r
2293 const int* labels = 0;
\r
2294 float* responses_buf = data->resp_float_buf;
\r
2295 const float* responses = 0;
\r
2296 data->get_cat_var_data(node, vi, labels_buf, &labels);
\r
2297 data->get_ord_responses(node, responses_buf, &responses);
\r
2299 double* sum = (double*)cvStackAlloc( (mi+1)*sizeof(sum[0]) ) + 1;
\r
2300 int* counts = (int*)cvStackAlloc( (mi+1)*sizeof(counts[0]) ) + 1;
\r
2301 double** sum_ptr = (double**)cvStackAlloc( (mi+1)*sizeof(sum_ptr[0]) );
\r
2302 int i, L = 0, R = 0;
\r
2303 double best_val = 0, lsum = 0, rsum = 0;
\r
2304 int best_subset = -1, subset_i;
\r
2306 for( i = -1; i < mi; i++ )
\r
2307 sum[i] = counts[i] = 0;
\r
2309 // calculate sum response and weight of each category of the input var
\r
2310 for( i = 0; i < n; i++ )
\r
2312 int idx = ( (labels[i] == 65535) && data->is_buf_16u ) ? -1 : labels[i];
\r
2313 double s = sum[idx] + responses[i];
\r
2314 int nc = counts[idx] + 1;
\r
2319 // calculate average response in each category
\r
2320 for( i = 0; i < mi; i++ )
\r
2324 sum[i] /= MAX(counts[i],1);
\r
2325 sum_ptr[i] = sum + i;
\r
2328 icvSortDblPtr( sum_ptr, mi, 0 );
\r
2330 // revert back to unnormalized sums
\r
2331 // (there should be a very little loss of accuracy)
\r
2332 for( i = 0; i < mi; i++ )
\r
2333 sum[i] *= counts[i];
\r
2335 for( subset_i = 0; subset_i < mi-1; subset_i++ )
\r
2337 int idx = (int)(sum_ptr[subset_i] - sum);
\r
2338 int ni = counts[idx];
\r
2342 double s = sum[idx];
\r
2343 lsum += s; L += ni;
\r
2344 rsum -= s; R -= ni;
\r
2348 double val = (lsum*lsum*R + rsum*rsum*L)/((double)L*R);
\r
2349 if( best_val < val )
\r
2352 best_subset = subset_i;
\r
2358 if( best_subset < 0 )
\r
2361 split = data->new_split_cat( vi, (float)best_val );
\r
2362 for( i = 0; i <= best_subset; i++ )
\r
2364 int idx = (int)(sum_ptr[i] - sum);
\r
2365 split->subset[idx >> 5] |= 1 << (idx & 31);
\r
2371 CvDTreeSplit* CvDTree::find_surrogate_split_ord( CvDTreeNode* node, int vi )
\r
2373 const float epsilon = FLT_EPSILON*2;
\r
2374 const char* dir = (char*)data->direction->data.ptr;
\r
2375 int n1 = node->get_num_valid(vi);
\r
2376 float* values_buf = data->pred_float_buf;
\r
2377 const float* values = 0;
\r
2378 int* indices_buf = data->pred_int_buf;
\r
2379 const int* indices = 0;
\r
2380 data->get_ord_var_data( node, vi, values_buf, indices_buf, &values, &indices );
\r
2381 // LL - number of samples that both the primary and the surrogate splits send to the left
\r
2382 // LR - ... primary split sends to the left and the surrogate split sends to the right
\r
2383 // RL - ... primary split sends to the right and the surrogate split sends to the left
\r
2384 // RR - ... both send to the right
\r
2385 int i, best_i = -1, best_inversed = 0;
\r
2388 if( !data->have_priors )
\r
2390 int LL = 0, RL = 0, LR, RR;
\r
2391 int worst_val = cvFloor(node->maxlr), _best_val = worst_val;
\r
2392 int sum = 0, sum_abs = 0;
\r
2394 for( i = 0; i < n1; i++ )
\r
2396 int d = dir[indices[i]];
\r
2397 sum += d; sum_abs += d & 1;
\r
2400 // sum_abs = R + L; sum = R - L
\r
2401 RR = (sum_abs + sum) >> 1;
\r
2402 LR = (sum_abs - sum) >> 1;
\r
2404 // initially all the samples are sent to the right by the surrogate split,
\r
2405 // LR of them are sent to the left by primary split, and RR - to the right.
\r
2406 // now iteratively compute LL, LR, RL and RR for every possible surrogate split value.
\r
2407 for( i = 0; i < n1 - 1; i++ )
\r
2409 int d = dir[indices[i]];
\r
2414 if( LL + RR > _best_val && values[i] + epsilon < values[i+1] )
\r
2416 best_val = LL + RR;
\r
2417 best_i = i; best_inversed = 0;
\r
2423 if( RL + LR > _best_val && values[i] + epsilon < values[i+1] )
\r
2425 best_val = RL + LR;
\r
2426 best_i = i; best_inversed = 1;
\r
2430 best_val = _best_val;
\r
2434 double LL = 0, RL = 0, LR, RR;
\r
2435 double worst_val = node->maxlr;
\r
2436 double sum = 0, sum_abs = 0;
\r
2437 const double* priors = data->priors_mult->data.db;
\r
2438 int* responses_buf = data->resp_int_buf;
\r
2439 const int* responses = 0;
\r
2440 data->get_class_labels(node, responses_buf, &responses);
\r
2441 best_val = worst_val;
\r
2443 for( i = 0; i < n1; i++ )
\r
2445 int idx = indices[i];
\r
2446 double w = priors[responses[idx]];
\r
2448 sum += d*w; sum_abs += (d & 1)*w;
\r
2451 // sum_abs = R + L; sum = R - L
\r
2452 RR = (sum_abs + sum)*0.5;
\r
2453 LR = (sum_abs - sum)*0.5;
\r
2455 // initially all the samples are sent to the right by the surrogate split,
\r
2456 // LR of them are sent to the left by primary split, and RR - to the right.
\r
2457 // now iteratively compute LL, LR, RL and RR for every possible surrogate split value.
\r
2458 for( i = 0; i < n1 - 1; i++ )
\r
2460 int idx = indices[i];
\r
2461 double w = priors[responses[idx]];
\r
2467 if( LL + RR > best_val && values[i] + epsilon < values[i+1] )
\r
2469 best_val = LL + RR;
\r
2470 best_i = i; best_inversed = 0;
\r
2476 if( RL + LR > best_val && values[i] + epsilon < values[i+1] )
\r
2478 best_val = RL + LR;
\r
2479 best_i = i; best_inversed = 1;
\r
2485 return best_i >= 0 && best_val > node->maxlr ? data->new_split_ord( vi,
\r
2486 (values[best_i] + values[best_i+1])*0.5f, best_i,
\r
2487 best_inversed, (float)best_val ) : 0;
\r
2491 CvDTreeSplit* CvDTree::find_surrogate_split_cat( CvDTreeNode* node, int vi )
\r
2493 const char* dir = (char*)data->direction->data.ptr;
\r
2494 int n = node->sample_count;
\r
2495 int* labels_buf = data->pred_int_buf;
\r
2496 const int* labels = 0;
\r
2497 data->get_cat_var_data(node, vi, labels_buf, &labels);
\r
2498 // LL - number of samples that both the primary and the surrogate splits send to the left
\r
2499 // LR - ... primary split sends to the left and the surrogate split sends to the right
\r
2500 // RL - ... primary split sends to the right and the surrogate split sends to the left
\r
2501 // RR - ... both send to the right
\r
2502 CvDTreeSplit* split = data->new_split_cat( vi, 0 );
\r
2503 int i, mi = data->cat_count->data.i[data->get_var_type(vi)], l_win = 0;
\r
2504 double best_val = 0;
\r
2505 double* lc = (double*)cvStackAlloc( (mi+1)*2*sizeof(lc[0]) ) + 1;
\r
2506 double* rc = lc + mi + 1;
\r
2508 for( i = -1; i < mi; i++ )
\r
2509 lc[i] = rc[i] = 0;
\r
2511 // for each category calculate the weight of samples
\r
2512 // sent to the left (lc) and to the right (rc) by the primary split
\r
2513 if( !data->have_priors )
\r
2515 int* _lc = (int*)cvStackAlloc((mi+2)*2*sizeof(_lc[0])) + 1;
\r
2516 int* _rc = _lc + mi + 1;
\r
2518 for( i = -1; i < mi; i++ )
\r
2519 _lc[i] = _rc[i] = 0;
\r
2521 for( i = 0; i < n; i++ )
\r
2523 int idx = ( (labels[i] == 65535) && (data->is_buf_16u) ) ? -1 : labels[i];
\r
2525 int sum = _lc[idx] + d;
\r
2526 int sum_abs = _rc[idx] + (d & 1);
\r
2527 _lc[idx] = sum; _rc[idx] = sum_abs;
\r
2530 for( i = 0; i < mi; i++ )
\r
2533 int sum_abs = _rc[i];
\r
2534 lc[i] = (sum_abs - sum) >> 1;
\r
2535 rc[i] = (sum_abs + sum) >> 1;
\r
2540 const double* priors = data->priors_mult->data.db;
\r
2541 int* responses_buf = data->resp_int_buf;
\r
2542 const int* responses = 0;
\r
2543 data->get_class_labels(node, responses_buf, &responses);
\r
2545 for( i = 0; i < n; i++ )
\r
2547 int idx = ( (labels[i] == 65535) && (data->is_buf_16u) ) ? -1 : labels[i];
\r
2548 double w = priors[responses[i]];
\r
2550 double sum = lc[idx] + d*w;
\r
2551 double sum_abs = rc[idx] + (d & 1)*w;
\r
2552 lc[idx] = sum; rc[idx] = sum_abs;
\r
2555 for( i = 0; i < mi; i++ )
\r
2557 double sum = lc[i];
\r
2558 double sum_abs = rc[i];
\r
2559 lc[i] = (sum_abs - sum) * 0.5;
\r
2560 rc[i] = (sum_abs + sum) * 0.5;
\r
2564 // 2. now form the split.
\r
2565 // in each category send all the samples to the same direction as majority
\r
2566 for( i = 0; i < mi; i++ )
\r
2568 double lval = lc[i], rval = rc[i];
\r
2571 split->subset[i >> 5] |= 1 << (i & 31);
\r
2579 split->quality = (float)best_val;
\r
2580 if( split->quality <= node->maxlr || l_win == 0 || l_win == mi )
\r
2581 cvSetRemoveByPtr( data->split_heap, split ), split = 0;
\r
2587 void CvDTree::calc_node_value( CvDTreeNode* node )
\r
2589 int i, j, k, n = node->sample_count, cv_n = data->params.cv_folds;
\r
2590 int* cv_labels_buf = data->cv_lables_buf;
\r
2591 const int* cv_labels = 0;
\r
2592 data->get_cv_labels(node, cv_labels_buf, &cv_labels);
\r
2594 if( data->is_classifier )
\r
2596 // in case of classification tree:
\r
2597 // * node value is the label of the class that has the largest weight in the node.
\r
2598 // * node risk is the weighted number of misclassified samples,
\r
2599 // * j-th cross-validation fold value and risk are calculated as above,
\r
2600 // but using the samples with cv_labels(*)!=j.
\r
2601 // * j-th cross-validation fold error is calculated as the weighted number of
\r
2602 // misclassified samples with cv_labels(*)==j.
\r
2604 // compute the number of instances of each class
\r
2605 int* cls_count = data->counts->data.i;
\r
2606 int* responses_buf = data->resp_int_buf;
\r
2607 const int* responses = 0;
\r
2608 data->get_class_labels(node, responses_buf, &responses);
\r
2609 int m = data->get_num_classes();
\r
2610 int* cv_cls_count = (int*)cvStackAlloc(m*cv_n*sizeof(cv_cls_count[0]));
\r
2611 double max_val = -1, total_weight = 0;
\r
2613 double* priors = data->priors_mult->data.db;
\r
2615 for( k = 0; k < m; k++ )
\r
2620 for( i = 0; i < n; i++ )
\r
2621 cls_count[responses[i]]++;
\r
2625 for( j = 0; j < cv_n; j++ )
\r
2626 for( k = 0; k < m; k++ )
\r
2627 cv_cls_count[j*m + k] = 0;
\r
2629 for( i = 0; i < n; i++ )
\r
2631 j = cv_labels[i]; k = responses[i];
\r
2632 cv_cls_count[j*m + k]++;
\r
2635 for( j = 0; j < cv_n; j++ )
\r
2636 for( k = 0; k < m; k++ )
\r
2637 cls_count[k] += cv_cls_count[j*m + k];
\r
2640 if( data->have_priors && node->parent == 0 )
\r
2642 // compute priors_mult from priors, take the sample ratio into account.
\r
2644 for( k = 0; k < m; k++ )
\r
2646 int n_k = cls_count[k];
\r
2647 priors[k] = data->priors->data.db[k]*(n_k ? 1./n_k : 0.);
\r
2651 for( k = 0; k < m; k++ )
\r
2655 for( k = 0; k < m; k++ )
\r
2657 double val = cls_count[k]*priors[k];
\r
2658 total_weight += val;
\r
2659 if( max_val < val )
\r
2666 node->class_idx = max_k;
\r
2667 node->value = data->cat_map->data.i[
\r
2668 data->cat_ofs->data.i[data->cat_var_count] + max_k];
\r
2669 node->node_risk = total_weight - max_val;
\r
2671 for( j = 0; j < cv_n; j++ )
\r
2673 double sum_k = 0, sum = 0, max_val_k = 0;
\r
2674 max_val = -1; max_k = -1;
\r
2676 for( k = 0; k < m; k++ )
\r
2678 double w = priors[k];
\r
2679 double val_k = cv_cls_count[j*m + k]*w;
\r
2680 double val = cls_count[k]*w - val_k;
\r
2683 if( max_val < val )
\r
2686 max_val_k = val_k;
\r
2691 node->cv_Tn[j] = INT_MAX;
\r
2692 node->cv_node_risk[j] = sum - max_val;
\r
2693 node->cv_node_error[j] = sum_k - max_val_k;
\r
2698 // in case of regression tree:
\r
2699 // * node value is 1/n*sum_i(Y_i), where Y_i is i-th response,
\r
2700 // n is the number of samples in the node.
\r
2701 // * node risk is the sum of squared errors: sum_i((Y_i - <node_value>)^2)
\r
2702 // * j-th cross-validation fold value and risk are calculated as above,
\r
2703 // but using the samples with cv_labels(*)!=j.
\r
2704 // * j-th cross-validation fold error is calculated
\r
2705 // using samples with cv_labels(*)==j as the test subset:
\r
2706 // error_j = sum_(i,cv_labels(i)==j)((Y_i - <node_value_j>)^2),
\r
2707 // where node_value_j is the node value calculated
\r
2708 // as described in the previous bullet, and summation is done
\r
2709 // over the samples with cv_labels(*)==j.
\r
2711 double sum = 0, sum2 = 0;
\r
2712 float* values_buf = data->resp_float_buf;
\r
2713 const float* values = 0;
\r
2714 data->get_ord_responses(node, values_buf, &values);
\r
2715 double *cv_sum = 0, *cv_sum2 = 0;
\r
2716 int* cv_count = 0;
\r
2720 for( i = 0; i < n; i++ )
\r
2722 double t = values[i];
\r
2729 cv_sum = (double*)cvStackAlloc( cv_n*sizeof(cv_sum[0]) );
\r
2730 cv_sum2 = (double*)cvStackAlloc( cv_n*sizeof(cv_sum2[0]) );
\r
2731 cv_count = (int*)cvStackAlloc( cv_n*sizeof(cv_count[0]) );
\r
2733 for( j = 0; j < cv_n; j++ )
\r
2735 cv_sum[j] = cv_sum2[j] = 0.;
\r
2739 for( i = 0; i < n; i++ )
\r
2742 double t = values[i];
\r
2743 double s = cv_sum[j] + t;
\r
2744 double s2 = cv_sum2[j] + t*t;
\r
2745 int nc = cv_count[j] + 1;
\r
2751 for( j = 0; j < cv_n; j++ )
\r
2754 sum2 += cv_sum2[j];
\r
2758 node->node_risk = sum2 - (sum/n)*sum;
\r
2759 node->value = sum/n;
\r
2761 for( j = 0; j < cv_n; j++ )
\r
2763 double s = cv_sum[j], si = sum - s;
\r
2764 double s2 = cv_sum2[j], s2i = sum2 - s2;
\r
2765 int c = cv_count[j], ci = n - c;
\r
2766 double r = si/MAX(ci,1);
\r
2767 node->cv_node_risk[j] = s2i - r*r*ci;
\r
2768 node->cv_node_error[j] = s2 - 2*r*s + c*r*r;
\r
2769 node->cv_Tn[j] = INT_MAX;
\r
2775 void CvDTree::complete_node_dir( CvDTreeNode* node )
\r
2777 int vi, i, n = node->sample_count, nl, nr, d0 = 0, d1 = -1;
\r
2778 int nz = n - node->get_num_valid(node->split->var_idx);
\r
2779 char* dir = (char*)data->direction->data.ptr;
\r
2781 // try to complete direction using surrogate splits
\r
2782 if( nz && data->params.use_surrogates )
\r
2784 CvDTreeSplit* split = node->split->next;
\r
2785 for( ; split != 0 && nz; split = split->next )
\r
2787 int inversed_mask = split->inversed ? -1 : 0;
\r
2788 vi = split->var_idx;
\r
2790 if( data->get_var_type(vi) >= 0 ) // split on categorical var
\r
2792 int* labels_buf = data->pred_int_buf;
\r
2793 const int* labels = 0;
\r
2794 data->get_cat_var_data(node, vi, labels_buf, &labels);
\r
2795 const int* subset = split->subset;
\r
2797 for( i = 0; i < n; i++ )
\r
2799 int idx = labels[i];
\r
2800 if( !dir[i] && ( ((idx >= 0)&&(!data->is_buf_16u)) || ((idx != 65535)&&(data->is_buf_16u)) ))
\r
2803 int d = CV_DTREE_CAT_DIR(idx,subset);
\r
2804 dir[i] = (char)((d ^ inversed_mask) - inversed_mask);
\r
2810 else // split on ordered var
\r
2812 float* values_buf = data->pred_float_buf;
\r
2813 const float* values = 0;
\r
2814 int* indices_buf = data->pred_int_buf;
\r
2815 const int* indices = 0;
\r
2816 data->get_ord_var_data( node, vi, values_buf, indices_buf, &values, &indices );
\r
2817 int split_point = split->ord.split_point;
\r
2818 int n1 = node->get_num_valid(vi);
\r
2820 assert( 0 <= split_point && split_point < n-1 );
\r
2822 for( i = 0; i < n1; i++ )
\r
2824 int idx = indices[i];
\r
2827 int d = i <= split_point ? -1 : 1;
\r
2828 dir[idx] = (char)((d ^ inversed_mask) - inversed_mask);
\r
2837 // find the default direction for the rest
\r
2840 for( i = nr = 0; i < n; i++ )
\r
2843 d0 = nl > nr ? -1 : nr > nl;
\r
2846 // make sure that every sample is directed either to the left or to the right
\r
2847 for( i = 0; i < n; i++ )
\r
2857 dir[i] = (char)d; // remap (-1,1) to (0,1)
\r
2862 void CvDTree::split_node_data( CvDTreeNode* node )
\r
2864 int vi, i, n = node->sample_count, nl, nr, scount = data->sample_count;
\r
2865 char* dir = (char*)data->direction->data.ptr;
\r
2866 CvDTreeNode *left = 0, *right = 0;
\r
2867 int* new_idx = data->split_buf->data.i;
\r
2868 int new_buf_idx = data->get_child_buf_idx( node );
\r
2869 int work_var_count = data->get_work_var_count();
\r
2870 CvMat* buf = data->buf;
\r
2871 int* temp_buf = (int*)cvStackAlloc(n*sizeof(temp_buf[0]));
\r
2873 complete_node_dir(node);
\r
2875 for( i = nl = nr = 0; i < n; i++ )
\r
2878 // initialize new indices for splitting ordered variables
\r
2879 new_idx[i] = (nl & (d-1)) | (nr & -d); // d ? ri : li
\r
2885 bool split_input_data;
\r
2886 node->left = left = data->new_node( node, nl, new_buf_idx, node->offset );
\r
2887 node->right = right = data->new_node( node, nr, new_buf_idx, node->offset + nl );
\r
2889 split_input_data = node->depth + 1 < data->params.max_depth &&
\r
2890 (node->left->sample_count > data->params.min_sample_count ||
\r
2891 node->right->sample_count > data->params.min_sample_count);
\r
2893 // split ordered variables, keep both halves sorted.
\r
2894 for( vi = 0; vi < data->var_count; vi++ )
\r
2896 int ci = data->get_var_type(vi);
\r
2897 int n1 = node->get_num_valid(vi);
\r
2898 int *src_idx_buf = data->pred_int_buf;
\r
2899 const int* src_idx = 0;
\r
2900 float *src_val_buf = data->pred_float_buf;
\r
2901 const float* src_val = 0;
\r
2903 if( ci >= 0 || !split_input_data )
\r
2906 data->get_ord_var_data(node, vi, src_val_buf, src_idx_buf, &src_val, &src_idx);
\r
2908 for(i = 0; i < n; i++)
\r
2909 temp_buf[i] = src_idx[i];
\r
2911 if (data->is_buf_16u)
\r
2913 unsigned short *ldst, *rdst, *ldst0, *rdst0;
\r
2914 //unsigned short tl, tr;
\r
2915 ldst0 = ldst = (unsigned short*)(buf->data.s + left->buf_idx*buf->cols +
\r
2916 vi*scount + left->offset);
\r
2917 rdst0 = rdst = (unsigned short*)(ldst + nl);
\r
2920 for( i = 0; i < n1; i++ )
\r
2922 int idx = temp_buf[i];
\r
2924 idx = new_idx[idx];
\r
2927 *rdst = (unsigned short)idx;
\r
2932 *ldst = (unsigned short)idx;
\r
2937 left->set_num_valid(vi, (int)(ldst - ldst0));
\r
2938 right->set_num_valid(vi, (int)(rdst - rdst0));
\r
2941 for( ; i < n; i++ )
\r
2943 int idx = temp_buf[i];
\r
2945 idx = new_idx[idx];
\r
2948 *rdst = (unsigned short)idx;
\r
2953 *ldst = (unsigned short)idx;
\r
2960 int *ldst0, *ldst, *rdst0, *rdst;
\r
2961 ldst0 = ldst = buf->data.i + left->buf_idx*buf->cols +
\r
2962 vi*scount + left->offset;
\r
2963 rdst0 = rdst = buf->data.i + right->buf_idx*buf->cols +
\r
2964 vi*scount + right->offset;
\r
2967 for( i = 0; i < n1; i++ )
\r
2969 int idx = temp_buf[i];
\r
2971 idx = new_idx[idx];
\r
2984 left->set_num_valid(vi, (int)(ldst - ldst0));
\r
2985 right->set_num_valid(vi, (int)(rdst - rdst0));
\r
2988 for( ; i < n; i++ )
\r
2990 int idx = temp_buf[i];
\r
2992 idx = new_idx[idx];
\r
3007 // split categorical vars, responses and cv_labels using new_idx relocation table
\r
3008 for( vi = 0; vi < work_var_count; vi++ )
\r
3010 int ci = data->get_var_type(vi);
\r
3011 int n1 = node->get_num_valid(vi), nr1 = 0;
\r
3013 if( ci < 0 || (vi < data->var_count && !split_input_data) )
\r
3016 int *src_lbls_buf = data->pred_int_buf;
\r
3017 const int* src_lbls = 0;
\r
3018 data->get_cat_var_data(node, vi, src_lbls_buf, &src_lbls);
\r
3020 for(i = 0; i < n; i++)
\r
3021 temp_buf[i] = src_lbls[i];
\r
3023 if (data->is_buf_16u)
\r
3025 unsigned short *ldst = (unsigned short *)(buf->data.s + left->buf_idx*buf->cols +
\r
3026 vi*scount + left->offset);
\r
3027 unsigned short *rdst = (unsigned short *)(buf->data.s + right->buf_idx*buf->cols +
\r
3028 vi*scount + right->offset);
\r
3030 for( i = 0; i < n; i++ )
\r
3033 int idx = temp_buf[i];
\r
3036 *rdst = (unsigned short)idx;
\r
3038 nr1 += (idx != 65535 )&d;
\r
3042 *ldst = (unsigned short)idx;
\r
3047 if( vi < data->var_count )
\r
3049 left->set_num_valid(vi, n1 - nr1);
\r
3050 right->set_num_valid(vi, nr1);
\r
3055 int *ldst = buf->data.i + left->buf_idx*buf->cols +
\r
3056 vi*scount + left->offset;
\r
3057 int *rdst = buf->data.i + right->buf_idx*buf->cols +
\r
3058 vi*scount + right->offset;
\r
3060 for( i = 0; i < n; i++ )
\r
3063 int idx = temp_buf[i];
\r
3068 nr1 += (idx >= 0)&d;
\r
3078 if( vi < data->var_count )
\r
3080 left->set_num_valid(vi, n1 - nr1);
\r
3081 right->set_num_valid(vi, nr1);
\r
3087 // split sample indices
\r
3088 int *sample_idx_src_buf = data->sample_idx_buf;
\r
3089 const int* sample_idx_src = 0;
\r
3090 data->get_sample_indices(node, sample_idx_src_buf, &sample_idx_src);
\r
3092 for(i = 0; i < n; i++)
\r
3093 temp_buf[i] = sample_idx_src[i];
\r
3095 int pos = data->get_work_var_count();
\r
3096 if (data->is_buf_16u)
\r
3098 unsigned short* ldst = (unsigned short*)(buf->data.s + left->buf_idx*buf->cols +
\r
3099 pos*scount + left->offset);
\r
3100 unsigned short* rdst = (unsigned short*)(buf->data.s + right->buf_idx*buf->cols +
\r
3101 pos*scount + right->offset);
\r
3102 for (i = 0; i < n; i++)
\r
3105 unsigned short idx = (unsigned short)temp_buf[i];
\r
3120 int* ldst = buf->data.i + left->buf_idx*buf->cols +
\r
3121 pos*scount + left->offset;
\r
3122 int* rdst = buf->data.i + right->buf_idx*buf->cols +
\r
3123 pos*scount + right->offset;
\r
3124 for (i = 0; i < n; i++)
\r
3127 int idx = temp_buf[i];
\r
3141 // deallocate the parent node data that is not needed anymore
\r
3142 data->free_node_data(node);
\r
3145 float CvDTree::calc_error( CvMLData* _data, int type )
\r
3148 const CvMat* values = _data->get_values();
3149 const CvMat* response = _data->get_response();
3150 const CvMat* missing = _data->get_missing();
3151 const CvMat* sample_idx = (type == CV_TEST_ERROR) ? _data->get_test_sample_idx() : _data->get_train_sample_idx();
\r
3152 const CvMat* var_types = _data->get_var_types();
\r
3153 int* sidx = sample_idx ? sample_idx->data.i : 0;
\r
3154 int r_step = CV_IS_MAT_CONT(response->type) ?
\r
3155 1 : response->step / CV_ELEM_SIZE(response->type);
\r
3156 bool is_classifier = var_types->data.ptr[var_types->cols-1] == CV_VAR_CATEGORICAL;
\r
3157 int sample_count = sample_idx ? sample_idx->cols : 0;
\r
3158 sample_count = (type == CV_TRAIN_ERROR && sample_count == 0) ? values->rows : sample_count;
\r
3159 if ( is_classifier )
\r
3161 for( int i = 0; i < sample_count; i++ )
\r
3163 CvMat sample, miss;
\r
3164 int si = sidx ? sidx[i] : i;
\r
3165 cvGetRow( values, &sample, si );
\r
3167 cvGetRow( missing, &miss, si );
\r
3168 float r = (float)predict( &sample, missing ? &miss : 0 )->value;
\r
3169 int d = fabs((double)r - response->data.fl[si*r_step]) <= FLT_EPSILON ? 0 : 1;
\r
3172 err = sample_count ? err / (float)sample_count * 100 : -FLT_MAX;
\r
3176 for( int i = 0; i < sample_count; i++ )
\r
3178 CvMat sample, miss;
\r
3179 int si = sidx ? sidx[i] : i;
\r
3180 cvGetRow( values, &sample, si );
\r
3182 cvGetRow( missing, &miss, si );
\r
3183 float r = (float)predict( &sample, missing ? &miss : 0 )->value;
\r
3184 float d = r - response->data.fl[si*r_step];
\r
3187 err = sample_count ? err / (float)sample_count : -FLT_MAX;
\r
3192 void CvDTree::prune_cv()
\r
3196 CvMat* err_jk = 0;
\r
3198 // 1. build tree sequence for each cv fold, calculate error_{Tj,beta_k}.
\r
3199 // 2. choose the best tree index (if need, apply 1SE rule).
\r
3200 // 3. store the best index and cut the branches.
\r
3202 CV_FUNCNAME( "CvDTree::prune_cv" );
\r
3206 int ti, j, tree_count = 0, cv_n = data->params.cv_folds, n = root->sample_count;
\r
3207 // currently, 1SE for regression is not implemented
\r
3208 bool use_1se = data->params.use_1se_rule != 0 && data->is_classifier;
\r
3210 double min_err = 0, min_err_se = 0;
\r
3213 CV_CALL( ab = cvCreateMat( 1, 256, CV_64F ));
\r
3215 // build the main tree sequence, calculate alpha's
\r
3216 for(;;tree_count++)
\r
3218 double min_alpha = update_tree_rnc(tree_count, -1);
\r
3219 if( cut_tree(tree_count, -1, min_alpha) )
\r
3222 if( ab->cols <= tree_count )
\r
3224 CV_CALL( temp = cvCreateMat( 1, ab->cols*3/2, CV_64F ));
\r
3225 for( ti = 0; ti < ab->cols; ti++ )
\r
3226 temp->data.db[ti] = ab->data.db[ti];
\r
3227 cvReleaseMat( &ab );
\r
3232 ab->data.db[tree_count] = min_alpha;
\r
3235 ab->data.db[0] = 0.;
\r
3237 if( tree_count > 0 )
\r
3239 for( ti = 1; ti < tree_count-1; ti++ )
\r
3240 ab->data.db[ti] = sqrt(ab->data.db[ti]*ab->data.db[ti+1]);
\r
3241 ab->data.db[tree_count-1] = DBL_MAX*0.5;
\r
3243 CV_CALL( err_jk = cvCreateMat( cv_n, tree_count, CV_64F ));
\r
3244 err = err_jk->data.db;
\r
3246 for( j = 0; j < cv_n; j++ )
\r
3248 int tj = 0, tk = 0;
\r
3249 for( ; tk < tree_count; tj++ )
\r
3251 double min_alpha = update_tree_rnc(tj, j);
\r
3252 if( cut_tree(tj, j, min_alpha) )
\r
3253 min_alpha = DBL_MAX;
\r
3255 for( ; tk < tree_count; tk++ )
\r
3257 if( ab->data.db[tk] > min_alpha )
\r
3259 err[j*tree_count + tk] = root->tree_error;
\r
3264 for( ti = 0; ti < tree_count; ti++ )
\r
3266 double sum_err = 0;
\r
3267 for( j = 0; j < cv_n; j++ )
\r
3268 sum_err += err[j*tree_count + ti];
\r
3269 if( ti == 0 || sum_err < min_err )
\r
3271 min_err = sum_err;
\r
3274 min_err_se = sqrt( sum_err*(n - sum_err) );
\r
3276 else if( sum_err < min_err + min_err_se )
\r
3281 pruned_tree_idx = min_idx;
\r
3282 free_prune_data(data->params.truncate_pruned_tree != 0);
\r
3286 cvReleaseMat( &err_jk );
\r
3287 cvReleaseMat( &ab );
\r
3288 cvReleaseMat( &temp );
\r
3292 double CvDTree::update_tree_rnc( int T, int fold )
\r
3294 CvDTreeNode* node = root;
\r
3295 double min_alpha = DBL_MAX;
\r
3299 CvDTreeNode* parent;
\r
3302 int t = fold >= 0 ? node->cv_Tn[fold] : node->Tn;
\r
3303 if( t <= T || !node->left )
\r
3305 node->complexity = 1;
\r
3306 node->tree_risk = node->node_risk;
\r
3307 node->tree_error = 0.;
\r
3310 node->tree_risk = node->cv_node_risk[fold];
\r
3311 node->tree_error = node->cv_node_error[fold];
\r
3315 node = node->left;
\r
3318 for( parent = node->parent; parent && parent->right == node;
\r
3319 node = parent, parent = parent->parent )
\r
3321 parent->complexity += node->complexity;
\r
3322 parent->tree_risk += node->tree_risk;
\r
3323 parent->tree_error += node->tree_error;
\r
3325 parent->alpha = ((fold >= 0 ? parent->cv_node_risk[fold] : parent->node_risk)
\r
3326 - parent->tree_risk)/(parent->complexity - 1);
\r
3327 min_alpha = MIN( min_alpha, parent->alpha );
\r
3333 parent->complexity = node->complexity;
\r
3334 parent->tree_risk = node->tree_risk;
\r
3335 parent->tree_error = node->tree_error;
\r
3336 node = parent->right;
\r
3343 int CvDTree::cut_tree( int T, int fold, double min_alpha )
\r
3345 CvDTreeNode* node = root;
\r
3351 CvDTreeNode* parent;
\r
3354 int t = fold >= 0 ? node->cv_Tn[fold] : node->Tn;
\r
3355 if( t <= T || !node->left )
\r
3357 if( node->alpha <= min_alpha + FLT_EPSILON )
\r
3360 node->cv_Tn[fold] = T;
\r
3363 if( node == root )
\r
3367 node = node->left;
\r
3370 for( parent = node->parent; parent && parent->right == node;
\r
3371 node = parent, parent = parent->parent )
\r
3377 node = parent->right;
\r
3384 void CvDTree::free_prune_data(bool cut_tree)
\r
3386 CvDTreeNode* node = root;
\r
3390 CvDTreeNode* parent;
\r
3393 // do not call cvSetRemoveByPtr( cv_heap, node->cv_Tn )
\r
3394 // as we will clear the whole cross-validation heap at the end
\r
3396 node->cv_node_error = node->cv_node_risk = 0;
\r
3399 node = node->left;
\r
3402 for( parent = node->parent; parent && parent->right == node;
\r
3403 node = parent, parent = parent->parent )
\r
3405 if( cut_tree && parent->Tn <= pruned_tree_idx )
\r
3407 data->free_node( parent->left );
\r
3408 data->free_node( parent->right );
\r
3409 parent->left = parent->right = 0;
\r
3416 node = parent->right;
\r
3419 if( data->cv_heap )
\r
3420 cvClearSet( data->cv_heap );
\r
3424 void CvDTree::free_tree()
\r
3426 if( root && data && data->shared )
\r
3428 pruned_tree_idx = INT_MIN;
\r
3429 free_prune_data(true);
\r
3430 data->free_node(root);
\r
3435 CvDTreeNode* CvDTree::predict( const CvMat* _sample,
\r
3436 const CvMat* _missing, bool preprocessed_input ) const
\r
3438 CvDTreeNode* result = 0;
\r
3441 CV_FUNCNAME( "CvDTree::predict" );
\r
3445 int i, step, mstep = 0;
\r
3446 const float* sample;
\r
3447 const uchar* m = 0;
\r
3448 CvDTreeNode* node = root;
\r
3455 CV_ERROR( CV_StsError, "The tree has not been trained yet" );
\r
3457 if( !CV_IS_MAT(_sample) || CV_MAT_TYPE(_sample->type) != CV_32FC1 ||
\r
3458 (_sample->cols != 1 && _sample->rows != 1) ||
\r
3459 (_sample->cols + _sample->rows - 1 != data->var_all && !preprocessed_input) ||
\r
3460 (_sample->cols + _sample->rows - 1 != data->var_count && preprocessed_input) )
\r
3461 CV_ERROR( CV_StsBadArg,
\r
3462 "the input sample must be 1d floating-point vector with the same "
\r
3463 "number of elements as the total number of variables used for training" );
\r
3465 sample = _sample->data.fl;
\r
3466 step = CV_IS_MAT_CONT(_sample->type) ? 1 : _sample->step/sizeof(sample[0]);
\r
3468 if( data->cat_count && !preprocessed_input ) // cache for categorical variables
\r
3470 int n = data->cat_count->cols;
\r
3471 catbuf = (int*)cvStackAlloc(n*sizeof(catbuf[0]));
\r
3472 for( i = 0; i < n; i++ )
\r
3478 if( !CV_IS_MAT(_missing) || !CV_IS_MASK_ARR(_missing) ||
\r
3479 !CV_ARE_SIZES_EQ(_missing, _sample) )
\r
3480 CV_ERROR( CV_StsBadArg,
\r
3481 "the missing data mask must be 8-bit vector of the same size as input sample" );
\r
3482 m = _missing->data.ptr;
\r
3483 mstep = CV_IS_MAT_CONT(_missing->type) ? 1 : _missing->step/sizeof(m[0]);
\r
3486 vtype = data->var_type->data.i;
\r
3487 vidx = data->var_idx && !preprocessed_input ? data->var_idx->data.i : 0;
\r
3488 cmap = data->cat_map ? data->cat_map->data.i : 0;
\r
3489 cofs = data->cat_ofs ? data->cat_ofs->data.i : 0;
\r
3491 while( node->Tn > pruned_tree_idx && node->left )
\r
3493 CvDTreeSplit* split = node->split;
\r
3495 for( ; !dir && split != 0; split = split->next )
\r
3497 int vi = split->var_idx;
\r
3498 int ci = vtype[vi];
\r
3499 i = vidx ? vidx[vi] : vi;
\r
3500 float val = sample[i*step];
\r
3501 if( m && m[i*mstep] )
\r
3503 if( ci < 0 ) // ordered
\r
3504 dir = val <= split->ord.c ? -1 : 1;
\r
3505 else // categorical
\r
3508 if( preprocessed_input )
\r
3515 int a = c = cofs[ci];
\r
3516 int b = (ci+1 >= data->cat_ofs->cols) ? data->cat_map->cols : cofs[ci+1];
\r
3518 int ival = cvRound(val);
\r
3520 CV_ERROR( CV_StsBadArg,
\r
3521 "one of input categorical variable is not an integer" );
\r
3528 if( ival < cmap[c] )
\r
3530 else if( ival > cmap[c] )
\r
3536 if( c < 0 || ival != cmap[c] )
\r
3539 catbuf[ci] = c -= cofs[ci];
\r
3542 c = ( (c == 65535) && data->is_buf_16u ) ? -1 : c;
\r
3543 dir = CV_DTREE_CAT_DIR(c, split->subset);
\r
3546 if( split->inversed )
\r
3552 double diff = node->right->sample_count - node->left->sample_count;
\r
3553 dir = diff < 0 ? -1 : 1;
\r
3555 node = dir < 0 ? node->left : node->right;
\r
3566 const CvMat* CvDTree::get_var_importance()
\r
3568 if( !var_importance )
\r
3570 CvDTreeNode* node = root;
\r
3571 double* importance;
\r
3574 var_importance = cvCreateMat( 1, data->var_count, CV_64F );
\r
3575 cvZero( var_importance );
\r
3576 importance = var_importance->data.db;
\r
3580 CvDTreeNode* parent;
\r
3581 for( ;; node = node->left )
\r
3583 CvDTreeSplit* split = node->split;
\r
3585 if( !node->left || node->Tn <= pruned_tree_idx )
\r
3588 for( ; split != 0; split = split->next )
\r
3589 importance[split->var_idx] += split->quality;
\r
3592 for( parent = node->parent; parent && parent->right == node;
\r
3593 node = parent, parent = parent->parent )
\r
3599 node = parent->right;
\r
3602 cvNormalize( var_importance, var_importance, 1., 0, CV_L1 );
\r
3605 return var_importance;
\r
3609 void CvDTree::write_split( CvFileStorage* fs, CvDTreeSplit* split )
\r
3613 cvStartWriteStruct( fs, 0, CV_NODE_MAP + CV_NODE_FLOW );
\r
3614 cvWriteInt( fs, "var", split->var_idx );
\r
3615 cvWriteReal( fs, "quality", split->quality );
\r
3617 ci = data->get_var_type(split->var_idx);
\r
3618 if( ci >= 0 ) // split on a categorical var
\r
3620 int i, n = data->cat_count->data.i[ci], to_right = 0, default_dir;
\r
3621 for( i = 0; i < n; i++ )
\r
3622 to_right += CV_DTREE_CAT_DIR(i,split->subset) > 0;
\r
3624 // ad-hoc rule when to use inverse categorical split notation
\r
3625 // to achieve more compact and clear representation
\r
3626 default_dir = to_right <= 1 || to_right <= MIN(3, n/2) || to_right <= n/3 ? -1 : 1;
\r
3628 cvStartWriteStruct( fs, default_dir*(split->inversed ? -1 : 1) > 0 ?
\r
3629 "in" : "not_in", CV_NODE_SEQ+CV_NODE_FLOW );
\r
3631 for( i = 0; i < n; i++ )
\r
3633 int dir = CV_DTREE_CAT_DIR(i,split->subset);
\r
3634 if( dir*default_dir < 0 )
\r
3635 cvWriteInt( fs, 0, i );
\r
3637 cvEndWriteStruct( fs );
\r
3640 cvWriteReal( fs, !split->inversed ? "le" : "gt", split->ord.c );
\r
3642 cvEndWriteStruct( fs );
\r
3646 void CvDTree::write_node( CvFileStorage* fs, CvDTreeNode* node )
\r
3648 CvDTreeSplit* split;
\r
3650 cvStartWriteStruct( fs, 0, CV_NODE_MAP );
\r
3652 cvWriteInt( fs, "depth", node->depth );
\r
3653 cvWriteInt( fs, "sample_count", node->sample_count );
\r
3654 cvWriteReal( fs, "value", node->value );
\r
3656 if( data->is_classifier )
\r
3657 cvWriteInt( fs, "norm_class_idx", node->class_idx );
\r
3659 cvWriteInt( fs, "Tn", node->Tn );
\r
3660 cvWriteInt( fs, "complexity", node->complexity );
\r
3661 cvWriteReal( fs, "alpha", node->alpha );
\r
3662 cvWriteReal( fs, "node_risk", node->node_risk );
\r
3663 cvWriteReal( fs, "tree_risk", node->tree_risk );
\r
3664 cvWriteReal( fs, "tree_error", node->tree_error );
\r
3668 cvStartWriteStruct( fs, "splits", CV_NODE_SEQ );
\r
3670 for( split = node->split; split != 0; split = split->next )
\r
3671 write_split( fs, split );
\r
3673 cvEndWriteStruct( fs );
\r
3676 cvEndWriteStruct( fs );
\r
3680 void CvDTree::write_tree_nodes( CvFileStorage* fs )
\r
3682 //CV_FUNCNAME( "CvDTree::write_tree_nodes" );
\r
3686 CvDTreeNode* node = root;
\r
3688 // traverse the tree and save all the nodes in depth-first order
\r
3691 CvDTreeNode* parent;
\r
3694 write_node( fs, node );
\r
3697 node = node->left;
\r
3700 for( parent = node->parent; parent && parent->right == node;
\r
3701 node = parent, parent = parent->parent )
\r
3707 node = parent->right;
\r
3714 void CvDTree::write( CvFileStorage* fs, const char* name )
\r
3716 //CV_FUNCNAME( "CvDTree::write" );
\r
3720 cvStartWriteStruct( fs, name, CV_NODE_MAP, CV_TYPE_NAME_ML_TREE );
\r
3722 get_var_importance();
\r
3723 data->write_params( fs );
\r
3724 if( var_importance )
\r
3725 cvWrite( fs, "var_importance", var_importance );
\r
3728 cvEndWriteStruct( fs );
\r
3734 void CvDTree::write( CvFileStorage* fs )
\r
3736 //CV_FUNCNAME( "CvDTree::write" );
\r
3740 cvWriteInt( fs, "best_tree_idx", pruned_tree_idx );
\r
3742 cvStartWriteStruct( fs, "nodes", CV_NODE_SEQ );
\r
3743 write_tree_nodes( fs );
\r
3744 cvEndWriteStruct( fs );
\r
3750 CvDTreeSplit* CvDTree::read_split( CvFileStorage* fs, CvFileNode* fnode )
\r
3752 CvDTreeSplit* split = 0;
\r
3754 CV_FUNCNAME( "CvDTree::read_split" );
\r
3760 if( !fnode || CV_NODE_TYPE(fnode->tag) != CV_NODE_MAP )
\r
3761 CV_ERROR( CV_StsParseError, "some of the splits are not stored properly" );
\r
3763 vi = cvReadIntByName( fs, fnode, "var", -1 );
\r
3764 if( (unsigned)vi >= (unsigned)data->var_count )
\r
3765 CV_ERROR( CV_StsOutOfRange, "Split variable index is out of range" );
\r
3767 ci = data->get_var_type(vi);
\r
3768 if( ci >= 0 ) // split on categorical var
\r
3770 int i, n = data->cat_count->data.i[ci], inversed = 0, val;
\r
3771 CvSeqReader reader;
\r
3772 CvFileNode* inseq;
\r
3773 split = data->new_split_cat( vi, 0 );
\r
3774 inseq = cvGetFileNodeByName( fs, fnode, "in" );
\r
3777 inseq = cvGetFileNodeByName( fs, fnode, "not_in" );
\r
3781 (CV_NODE_TYPE(inseq->tag) != CV_NODE_SEQ && CV_NODE_TYPE(inseq->tag) != CV_NODE_INT))
\r
3782 CV_ERROR( CV_StsParseError,
\r
3783 "Either 'in' or 'not_in' tags should be inside a categorical split data" );
\r
3785 if( CV_NODE_TYPE(inseq->tag) == CV_NODE_INT )
\r
3787 val = inseq->data.i;
\r
3788 if( (unsigned)val >= (unsigned)n )
\r
3789 CV_ERROR( CV_StsOutOfRange, "some of in/not_in elements are out of range" );
\r
3791 split->subset[val >> 5] |= 1 << (val & 31);
\r
3795 cvStartReadSeq( inseq->data.seq, &reader );
\r
3797 for( i = 0; i < reader.seq->total; i++ )
\r
3799 CvFileNode* inode = (CvFileNode*)reader.ptr;
\r
3800 val = inode->data.i;
\r
3801 if( CV_NODE_TYPE(inode->tag) != CV_NODE_INT || (unsigned)val >= (unsigned)n )
\r
3802 CV_ERROR( CV_StsOutOfRange, "some of in/not_in elements are out of range" );
\r
3804 split->subset[val >> 5] |= 1 << (val & 31);
\r
3805 CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
\r
3809 // for categorical splits we do not use inversed splits,
\r
3810 // instead we inverse the variable set in the split
\r
3812 for( i = 0; i < (n + 31) >> 5; i++ )
\r
3813 split->subset[i] ^= -1;
\r
3817 CvFileNode* cmp_node;
\r
3818 split = data->new_split_ord( vi, 0, 0, 0, 0 );
\r
3820 cmp_node = cvGetFileNodeByName( fs, fnode, "le" );
\r
3823 cmp_node = cvGetFileNodeByName( fs, fnode, "gt" );
\r
3824 split->inversed = 1;
\r
3827 split->ord.c = (float)cvReadReal( cmp_node );
\r
3830 split->quality = (float)cvReadRealByName( fs, fnode, "quality" );
\r
3838 CvDTreeNode* CvDTree::read_node( CvFileStorage* fs, CvFileNode* fnode, CvDTreeNode* parent )
\r
3840 CvDTreeNode* node = 0;
\r
3842 CV_FUNCNAME( "CvDTree::read_node" );
\r
3846 CvFileNode* splits;
\r
3849 if( !fnode || CV_NODE_TYPE(fnode->tag) != CV_NODE_MAP )
\r
3850 CV_ERROR( CV_StsParseError, "some of the tree elements are not stored properly" );
\r
3852 CV_CALL( node = data->new_node( parent, 0, 0, 0 ));
\r
3853 depth = cvReadIntByName( fs, fnode, "depth", -1 );
\r
3854 if( depth != node->depth )
\r
3855 CV_ERROR( CV_StsParseError, "incorrect node depth" );
\r
3857 node->sample_count = cvReadIntByName( fs, fnode, "sample_count" );
\r
3858 node->value = cvReadRealByName( fs, fnode, "value" );
\r
3859 if( data->is_classifier )
\r
3860 node->class_idx = cvReadIntByName( fs, fnode, "norm_class_idx" );
\r
3862 node->Tn = cvReadIntByName( fs, fnode, "Tn" );
\r
3863 node->complexity = cvReadIntByName( fs, fnode, "complexity" );
\r
3864 node->alpha = cvReadRealByName( fs, fnode, "alpha" );
\r
3865 node->node_risk = cvReadRealByName( fs, fnode, "node_risk" );
\r
3866 node->tree_risk = cvReadRealByName( fs, fnode, "tree_risk" );
\r
3867 node->tree_error = cvReadRealByName( fs, fnode, "tree_error" );
\r
3869 splits = cvGetFileNodeByName( fs, fnode, "splits" );
\r
3872 CvSeqReader reader;
\r
3873 CvDTreeSplit* last_split = 0;
\r
3875 if( CV_NODE_TYPE(splits->tag) != CV_NODE_SEQ )
\r
3876 CV_ERROR( CV_StsParseError, "splits tag must stored as a sequence" );
\r
3878 cvStartReadSeq( splits->data.seq, &reader );
\r
3879 for( i = 0; i < reader.seq->total; i++ )
\r
3881 CvDTreeSplit* split;
\r
3882 CV_CALL( split = read_split( fs, (CvFileNode*)reader.ptr ));
\r
3884 node->split = last_split = split;
\r
3886 last_split = last_split->next = split;
\r
3888 CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
\r
3898 void CvDTree::read_tree_nodes( CvFileStorage* fs, CvFileNode* fnode )
\r
3900 CV_FUNCNAME( "CvDTree::read_tree_nodes" );
\r
3904 CvSeqReader reader;
\r
3905 CvDTreeNode _root;
\r
3906 CvDTreeNode* parent = &_root;
\r
3908 parent->left = parent->right = parent->parent = 0;
\r
3910 cvStartReadSeq( fnode->data.seq, &reader );
\r
3912 for( i = 0; i < reader.seq->total; i++ )
\r
3914 CvDTreeNode* node;
\r
3916 CV_CALL( node = read_node( fs, (CvFileNode*)reader.ptr, parent != &_root ? parent : 0 ));
\r
3917 if( !parent->left )
\r
3918 parent->left = node;
\r
3920 parent->right = node;
\r
3925 while( parent && parent->right )
\r
3926 parent = parent->parent;
\r
3929 CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
\r
3932 root = _root.left;
\r
3938 void CvDTree::read( CvFileStorage* fs, CvFileNode* fnode )
\r
3940 CvDTreeTrainData* _data = new CvDTreeTrainData();
\r
3941 _data->read_params( fs, fnode );
\r
3943 read( fs, fnode, _data );
\r
3944 get_var_importance();
\r
3948 // a special entry point for reading weak decision trees from the tree ensembles
\r
3949 void CvDTree::read( CvFileStorage* fs, CvFileNode* node, CvDTreeTrainData* _data )
\r
3951 CV_FUNCNAME( "CvDTree::read" );
\r
3955 CvFileNode* tree_nodes;
\r
3960 tree_nodes = cvGetFileNodeByName( fs, node, "nodes" );
\r
3961 if( !tree_nodes || CV_NODE_TYPE(tree_nodes->tag) != CV_NODE_SEQ )
\r
3962 CV_ERROR( CV_StsParseError, "nodes tag is missing" );
\r
3964 pruned_tree_idx = cvReadIntByName( fs, node, "best_tree_idx", -1 );
\r
3965 read_tree_nodes( fs, tree_nodes );
\r
3970 /* End of file. */
\r