1 /*M///////////////////////////////////////////////////////////////////////////////////////
3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
5 // By downloading, copying, installing or using the software you agree to this license.
6 // If you do not agree to this license, do not download, install,
7 // copy or use the software.
10 // Intel License Agreement
12 // Copyright (C) 2000, Intel Corporation, all rights reserved.
13 // Third party copyrights are property of their respective owners.
15 // Redistribution and use in source and binary forms, with or without modification,
16 // are permitted provided that the following conditions are met:
18 // * Redistribution's of source code must retain the above copyright notice,
19 // this list of conditions and the following disclaimer.
21 // * Redistribution's in binary form must reproduce the above copyright notice,
22 // this list of conditions and the following disclaimer in the documentation
23 // and/or other materials provided with the distribution.
25 // * The name of Intel Corporation may not be used to endorse or promote products
26 // derived from this software without specific prior written permission.
28 // This software is provided by the copyright holders and contributors "as is" and
29 // any express or implied warranties, including, but not limited to, the implied
30 // warranties of merchantability and fitness for a particular purpose are disclaimed.
31 // In no event shall the Intel Corporation or contributors be liable for any direct,
32 // indirect, incidental, special, exemplary, or consequential damages
33 // (including, but not limited to, procurement of substitute goods or services;
34 // loss of use, data, or profits; or business interruption) however caused
35 // and on any theory of liability, whether in contract, strict liability,
36 // or tort (including negligence or otherwise) arising in any way out of
37 // the use of this software, even if advised of the possibility of such damage.
43 static const float ord_nan = FLT_MAX*0.5f;
44 static const int min_block_size = 1 << 16;
45 static const int block_size_delta = 1 << 10;
47 CvDTreeTrainData::CvDTreeTrainData()
49 var_idx = var_type = cat_count = cat_ofs = cat_map =
50 priors = counts = buf = direction = split_buf = 0;
51 tree_storage = temp_storage = 0;
57 CvDTreeTrainData::CvDTreeTrainData( const CvMat* _train_data, int _tflag,
58 const CvMat* _responses, const CvMat* _var_idx,
59 const CvMat* _sample_idx, const CvMat* _var_type,
60 const CvMat* _missing_mask,
61 CvDTreeParams _params, bool _shared, bool _add_weights )
63 var_idx = var_type = cat_count = cat_ofs = cat_map =
64 priors = counts = buf = direction = split_buf = 0;
65 tree_storage = temp_storage = 0;
67 set_data( _train_data, _tflag, _responses, _var_idx, _sample_idx,
68 _var_type, _missing_mask, _params, _shared, _add_weights );
72 CvDTreeTrainData::~CvDTreeTrainData()
78 bool CvDTreeTrainData::set_params( const CvDTreeParams& _params )
82 CV_FUNCNAME( "CvDTreeTrainData::set_params" );
89 if( params.max_categories < 2 )
90 CV_ERROR( CV_StsOutOfRange, "params.max_categories should be >= 2" );
91 params.max_categories = MIN( params.max_categories, 15 );
93 if( params.max_depth < 0 )
94 CV_ERROR( CV_StsOutOfRange, "params.max_depth should be >= 0" );
95 params.max_depth = MIN( params.max_depth, 25 );
97 params.min_sample_count = MAX(params.min_sample_count,1);
99 if( params.cv_folds < 0 )
100 CV_ERROR( CV_StsOutOfRange,
101 "params.cv_folds should be =0 (the tree is not pruned) "
102 "or n>0 (tree is pruned using n-fold cross-validation)" );
104 if( params.cv_folds == 1 )
107 if( params.regression_accuracy < 0 )
108 CV_ERROR( CV_StsOutOfRange, "params.regression_accuracy should be >= 0" );
118 #define CV_CMP_NUM_PTR(a,b) (*(a) < *(b))
119 static CV_IMPLEMENT_QSORT_EX( icvSortIntPtr, int*, CV_CMP_NUM_PTR, int )
120 static CV_IMPLEMENT_QSORT_EX( icvSortDblPtr, double*, CV_CMP_NUM_PTR, int )
122 #define CV_CMP_PAIRS(a,b) ((a).val < (b).val)
123 static CV_IMPLEMENT_QSORT_EX( icvSortPairs, CvPair32s32f, CV_CMP_PAIRS, int )
125 void CvDTreeTrainData::set_data( const CvMat* _train_data, int _tflag,
126 const CvMat* _responses, const CvMat* _var_idx, const CvMat* _sample_idx,
127 const CvMat* _var_type, const CvMat* _missing_mask, CvDTreeParams _params,
128 bool _shared, bool _add_weights )
130 CvMat* sample_idx = 0;
131 CvMat* var_type0 = 0;
135 CV_FUNCNAME( "CvDTreeTrainData::set_data" );
139 int sample_all = 0, r_type = 0, cv_n;
140 int total_c_count = 0;
141 int tree_block_size, temp_block_size, max_split_size, nv_size, cv_size = 0;
142 int ds_step, dv_step, ms_step = 0, mv_step = 0; // {data|mask}{sample|var}_step
145 const int *sidx = 0, *vidx = 0;
152 CV_CALL( set_params( _params ));
154 // check parameter types and sizes
155 CV_CALL( cvCheckTrainData( _train_data, _tflag, _missing_mask, &var_all, &sample_all ));
156 if( _tflag == CV_ROW_SAMPLE )
158 ds_step = _train_data->step/CV_ELEM_SIZE(_train_data->type);
161 ms_step = _missing_mask->step, mv_step = 1;
165 dv_step = _train_data->step/CV_ELEM_SIZE(_train_data->type);
168 mv_step = _missing_mask->step, ms_step = 1;
171 sample_count = sample_all;
176 CV_CALL( sample_idx = cvPreprocessIndexArray( _sample_idx, sample_all ));
177 sidx = sample_idx->data.i;
178 sample_count = sample_idx->rows + sample_idx->cols - 1;
183 CV_CALL( var_idx = cvPreprocessIndexArray( _var_idx, var_all ));
184 vidx = var_idx->data.i;
185 var_count = var_idx->rows + var_idx->cols - 1;
188 if( !CV_IS_MAT(_responses) ||
189 (CV_MAT_TYPE(_responses->type) != CV_32SC1 &&
190 CV_MAT_TYPE(_responses->type) != CV_32FC1) ||
191 _responses->rows != 1 && _responses->cols != 1 ||
192 _responses->rows + _responses->cols - 1 != sample_all )
193 CV_ERROR( CV_StsBadArg, "The array of _responses must be an integer or "
194 "floating-point vector containing as many elements as "
195 "the total number of samples in the training data matrix" );
197 CV_CALL( var_type0 = cvPreprocessVarType( _var_type, var_idx, var_all, &r_type ));
198 CV_CALL( var_type = cvCreateMat( 1, var_count+2, CV_32SC1 ));
203 is_classifier = r_type == CV_VAR_CATEGORICAL;
205 // step 0. calc the number of categorical vars
206 for( vi = 0; vi < var_count; vi++ )
208 var_type->data.i[vi] = var_type0->data.ptr[vi] == CV_VAR_CATEGORICAL ?
209 cat_var_count++ : ord_var_count--;
212 ord_var_count = ~ord_var_count;
213 cv_n = params.cv_folds;
214 // set the two last elements of var_type array to be able
215 // to locate responses and cross-validation labels using
216 // the corresponding get_* functions.
217 var_type->data.i[var_count] = cat_var_count;
218 var_type->data.i[var_count+1] = cat_var_count+1;
220 // in case of single ordered predictor we need dummy cv_labels
221 // for safe split_node_data() operation
222 have_cv_labels = cv_n > 0 || ord_var_count == 1 && cat_var_count == 0;
223 have_weights = _add_weights;
225 buf_size = (ord_var_count*2 + cat_var_count + 1 +
226 (have_cv_labels ? 1 : 0) + (have_weights ? 1 : 0))*sample_count + 2;
228 buf_count = shared ? 3 : 2;
229 CV_CALL( buf = cvCreateMat( buf_count, buf_size, CV_32SC1 ));
230 CV_CALL( cat_count = cvCreateMat( 1, cat_var_count+1, CV_32SC1 ));
231 CV_CALL( cat_ofs = cvCreateMat( 1, cat_count->cols+1, CV_32SC1 ));
232 CV_CALL( cat_map = cvCreateMat( 1, cat_count->cols*10 + 128, CV_32SC1 ));
234 // now calculate the maximum size of split,
235 // create memory storage that will keep nodes and splits of the decision tree
236 // allocate root node and the buffer for the whole training data
237 max_split_size = cvAlign(sizeof(CvDTreeSplit) +
238 (MAX(0,sample_count - 33)/32)*sizeof(int),sizeof(void*));
239 tree_block_size = MAX((int)sizeof(CvDTreeNode)*8, max_split_size);
240 tree_block_size = MAX(tree_block_size + block_size_delta, min_block_size);
241 CV_CALL( tree_storage = cvCreateMemStorage( tree_block_size ));
242 CV_CALL( node_heap = cvCreateSet( 0, sizeof(*node_heap), sizeof(CvDTreeNode), tree_storage ));
244 temp_block_size = nv_size = var_count*sizeof(int);
247 if( sample_count < cv_n*MAX(params.min_sample_count,10) )
248 CV_ERROR( CV_StsOutOfRange,
249 "The many folds in cross-validation for such a small dataset" );
251 cv_size = cvAlign( cv_n*(sizeof(int) + sizeof(double)*2), sizeof(double) );
252 temp_block_size = MAX(temp_block_size, cv_size);
255 temp_block_size = MAX( temp_block_size + block_size_delta, min_block_size );
256 CV_CALL( temp_storage = cvCreateMemStorage( temp_block_size ));
257 CV_CALL( nv_heap = cvCreateSet( 0, sizeof(*nv_heap), nv_size, temp_storage ));
259 CV_CALL( cv_heap = cvCreateSet( 0, sizeof(*cv_heap), cv_size, temp_storage ));
261 CV_CALL( data_root = new_node( 0, sample_count, 0, 0 ));
262 CV_CALL( int_ptr = (int**)cvAlloc( sample_count*sizeof(int_ptr[0]) ));
266 // transform the training data to convenient representation
267 for( vi = 0; vi <= var_count; vi++ )
270 const uchar* mask = 0;
271 int m_step = 0, step;
272 const int* idata = 0;
273 const float* fdata = 0;
276 if( vi < var_count ) // analyze i-th input variable
278 int vi0 = vidx ? vidx[vi] : vi;
279 ci = get_var_type(vi);
280 step = ds_step; m_step = ms_step;
281 if( CV_MAT_TYPE(_train_data->type) == CV_32SC1 )
282 idata = _train_data->data.i + vi0*dv_step;
284 fdata = _train_data->data.fl + vi0*dv_step;
286 mask = _missing_mask->data.ptr + vi0*mv_step;
288 else // analyze _responses
291 step = CV_IS_MAT_CONT(_responses->type) ?
292 1 : _responses->step / CV_ELEM_SIZE(_responses->type);
293 if( CV_MAT_TYPE(_responses->type) == CV_32SC1 )
294 idata = _responses->data.i;
296 fdata = _responses->data.fl;
299 if( vi < var_count && ci >= 0 ||
300 vi == var_count && is_classifier ) // process categorical variable or response
302 int c_count, prev_label;
303 int* c_map, *dst = get_cat_var_data( data_root, vi );
306 for( i = 0; i < sample_count; i++ )
308 int val = INT_MAX, si = sidx ? sidx[i] : i;
309 if( !mask || !mask[si*m_step] )
312 val = idata[si*step];
315 float t = fdata[si*step];
319 sprintf( err, "%d-th value of %d-th (categorical) "
320 "variable is not an integer", i, vi );
321 CV_ERROR( CV_StsBadArg, err );
327 sprintf( err, "%d-th value of %d-th (categorical) "
328 "variable is too large", i, vi );
329 CV_ERROR( CV_StsBadArg, err );
334 int_ptr[i] = dst + i;
337 // sort all the values, including the missing measurements
338 // that should all move to the end
339 icvSortIntPtr( int_ptr, sample_count, 0 );
340 //qsort( int_ptr, sample_count, sizeof(int_ptr[0]), icvCmpIntPtr );
342 c_count = num_valid > 0;
344 // count the categories
345 for( i = 1; i < num_valid; i++ )
346 c_count += *int_ptr[i] != *int_ptr[i-1];
349 max_c_count = MAX( max_c_count, c_count );
350 cat_count->data.i[ci] = c_count;
351 cat_ofs->data.i[ci] = total_c_count;
353 // resize cat_map, if need
354 if( cat_map->cols < total_c_count + c_count )
357 CV_CALL( cat_map = cvCreateMat( 1,
358 MAX(cat_map->cols*3/2,total_c_count+c_count), CV_32SC1 ));
359 for( i = 0; i < total_c_count; i++ )
360 cat_map->data.i[i] = tmp_map->data.i[i];
361 cvReleaseMat( &tmp_map );
364 c_map = cat_map->data.i + total_c_count;
365 total_c_count += c_count;
367 // compact the class indices and build the map
368 prev_label = ~*int_ptr[0];
371 for( i = 0; i < num_valid; i++ )
373 int cur_label = *int_ptr[i];
374 if( cur_label != prev_label )
375 c_map[++c_count] = prev_label = cur_label;
376 *int_ptr[i] = c_count;
379 // replace labels for missing values with -1
380 for( ; i < sample_count; i++ )
383 else if( ci < 0 ) // process ordered variable
385 CvPair32s32f* dst = get_ord_var_data( data_root, vi );
387 for( i = 0; i < sample_count; i++ )
390 int si = sidx ? sidx[i] : i;
391 if( !mask || !mask[si*m_step] )
394 val = (float)idata[si*step];
396 val = fdata[si*step];
398 if( fabs(val) >= ord_nan )
400 sprintf( err, "%d-th value of %d-th (ordered) "
401 "variable (=%g) is too large", i, vi, val );
402 CV_ERROR( CV_StsBadArg, err );
410 icvSortPairs( dst, sample_count, 0 );
412 else // special case: process ordered response,
413 // it will be stored similarly to categorical vars (i.e. no pairs)
415 float* dst = get_ord_responses( data_root );
417 for( i = 0; i < sample_count; i++ )
420 int si = sidx ? sidx[i] : i;
422 val = (float)idata[si*step];
424 val = fdata[si*step];
426 if( fabs(val) >= ord_nan )
428 sprintf( err, "%d-th value of %d-th (ordered) "
429 "variable (=%g) is out of range", i, vi, val );
430 CV_ERROR( CV_StsBadArg, err );
435 cat_count->data.i[cat_var_count] = 0;
436 cat_ofs->data.i[cat_var_count] = total_c_count;
437 num_valid = sample_count;
441 data_root->set_num_valid(vi, num_valid);
446 int* dst = get_cv_labels(data_root);
449 for( i = vi = 0; i < sample_count; i++ )
452 vi &= vi < cv_n ? -1 : 0;
455 for( i = 0; i < sample_count; i++ )
457 int a = cvRandInt(r) % sample_count;
458 int b = cvRandInt(r) % sample_count;
459 CV_SWAP( dst[a], dst[b], vi );
463 cat_map->cols = MAX( total_c_count, 1 );
465 max_split_size = cvAlign(sizeof(CvDTreeSplit) +
466 (MAX(0,max_c_count - 33)/32)*sizeof(int),sizeof(void*));
467 CV_CALL( split_heap = cvCreateSet( 0, sizeof(*split_heap), max_split_size, tree_storage ));
469 have_priors = is_classifier && params.priors;
472 int m = get_num_classes(), rows = 4;
474 CV_CALL( priors = cvCreateMat( 1, m, CV_64F ));
475 for( i = 0; i < m; i++ )
477 double val = have_priors ? params.priors[i] : 1.;
479 CV_ERROR( CV_StsOutOfRange, "Every class weight should be positive" );
480 priors->data.db[i] = val;
485 cvScale( priors, priors, 1./sum );
487 if( cat_var_count > 0 || params.cv_folds > 0 )
489 // need storage for cjk (see find_split_cat_gini) and risks/errors
490 rows += MAX( max_c_count, params.cv_folds ) + 1;
491 // add buffer for k-means clustering
492 if( m > 2 && max_c_count > params.max_categories )
493 rows += params.max_categories + (max_c_count+m-1)/m;
496 CV_CALL( counts = cvCreateMat( rows, m, CV_32SC2 ));
499 CV_CALL( direction = cvCreateMat( 1, sample_count, CV_8UC1 ));
500 CV_CALL( split_buf = cvCreateMat( 1, sample_count, CV_32SC1 ));
505 cvReleaseMat( &sample_idx );
506 cvReleaseMat( &var_type0 );
507 cvReleaseMat( &tmp_map );
511 CvDTreeNode* CvDTreeTrainData::subsample_data( const CvMat* _subsample_idx )
513 CvDTreeNode* root = 0;
514 CvMat* isubsample_idx = 0;
515 CvMat* subsample_co = 0;
517 CV_FUNCNAME( "CvDTreeTrainData::subsample_data" );
522 CV_ERROR( CV_StsError, "No training data has been set" );
525 CV_CALL( isubsample_idx = cvPreprocessIndexArray( _subsample_idx, sample_count ));
527 if( !isubsample_idx )
529 // make a copy of the root node
532 root = new_node( 0, 1, 0, 0 );
535 root->num_valid = temp.num_valid;
536 if( root->num_valid )
538 for( i = 0; i < var_count; i++ )
539 root->num_valid[i] = data_root->num_valid[i];
541 root->cv_Tn = temp.cv_Tn;
542 root->cv_node_risk = temp.cv_node_risk;
543 root->cv_node_error = temp.cv_node_error;
547 int* sidx = isubsample_idx->data.i;
548 // co - array of count/offset pairs (to handle duplicated values in _subsample_idx)
549 int* co, cur_ofs = 0;
550 int vi, i, total = data_root->sample_count;
551 int count = isubsample_idx->rows + isubsample_idx->cols - 1;
552 root = new_node( 0, count, 1, 0 );
554 CV_CALL( subsample_co = cvCreateMat( 1, total*2, CV_32SC1 ));
555 cvZero( subsample_co );
556 co = subsample_co->data.i;
557 for( i = 0; i < count; i++ )
559 for( i = 0; i < total; i++ )
570 for( vi = 0; vi <= var_count + (have_cv_labels ? 1 : 0); vi++ )
572 int ci = get_var_type(vi);
574 if( ci >= 0 || vi >= var_count )
576 const int* src = get_cat_var_data( data_root, vi );
577 int* dst = get_cat_var_data( root, vi );
580 for( i = 0; i < count; i++ )
582 int val = src[sidx[i]];
584 num_valid += val >= 0;
588 root->set_num_valid(vi, num_valid);
592 const CvPair32s32f* src = get_ord_var_data( data_root, vi );
593 CvPair32s32f* dst = get_ord_var_data( root, vi );
594 int j = 0, idx, count_i;
595 int num_valid = data_root->get_num_valid(vi);
597 for( i = 0; i < num_valid; i++ )
603 float val = src[i].val;
604 for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
612 root->set_num_valid(vi, j);
614 for( ; i < total; i++ )
620 float val = src[i].val;
621 for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
634 cvReleaseMat( &isubsample_idx );
635 cvReleaseMat( &subsample_co );
641 void CvDTreeTrainData::get_vectors( const CvMat* _subsample_idx,
642 float* values, uchar* missing,
643 float* responses, bool get_class_idx )
645 CvMat* subsample_idx = 0;
646 CvMat* subsample_co = 0;
648 CV_FUNCNAME( "CvDTreeTrainData::get_vectors" );
652 int i, vi, total = sample_count, count = total, cur_ofs = 0;
658 CV_CALL( subsample_idx = cvPreprocessIndexArray( _subsample_idx, sample_count ));
659 sidx = subsample_idx->data.i;
660 CV_CALL( subsample_co = cvCreateMat( 1, sample_count*2, CV_32SC1 ));
661 co = subsample_co->data.i;
662 cvZero( subsample_co );
663 count = subsample_idx->cols + subsample_idx->rows - 1;
664 for( i = 0; i < count; i++ )
666 for( i = 0; i < total; i++ )
668 int count_i = co[i*2];
671 co[i*2+1] = cur_ofs*var_count;
677 memset( missing, 1, count*var_count );
679 for( vi = 0; vi < var_count; vi++ )
681 int ci = get_var_type(vi);
682 if( ci >= 0 ) // categorical
684 float* dst = values + vi;
685 uchar* m = missing + vi;
686 const int* src = get_cat_var_data(data_root, vi);
688 for( i = 0; i < count; i++, dst += var_count, m += var_count )
690 int idx = sidx ? sidx[i] : i;
698 float* dst = values + vi;
699 uchar* m = missing + vi;
700 const CvPair32s32f* src = get_ord_var_data(data_root, vi);
701 int count1 = data_root->get_num_valid(vi);
703 for( i = 0; i < count1; i++ )
710 cur_ofs = co[idx*2+1];
713 cur_ofs = idx*var_count;
716 float val = src[i].val;
717 for( ; count_i > 0; count_i--, cur_ofs += var_count )
730 const int* src = get_class_labels(data_root);
731 for( i = 0; i < count; i++ )
733 int idx = sidx ? sidx[i] : i;
734 int val = get_class_idx ? src[idx] :
735 cat_map->data.i[cat_ofs->data.i[cat_var_count]+src[idx]];
736 responses[i] = (float)val;
741 const float* src = get_ord_responses(data_root);
742 for( i = 0; i < count; i++ )
744 int idx = sidx ? sidx[i] : i;
745 responses[i] = src[idx];
751 cvReleaseMat( &subsample_idx );
752 cvReleaseMat( &subsample_co );
756 CvDTreeNode* CvDTreeTrainData::new_node( CvDTreeNode* parent, int count,
757 int storage_idx, int offset )
759 CvDTreeNode* node = (CvDTreeNode*)cvSetNew( node_heap );
761 node->sample_count = count;
762 node->depth = parent ? parent->depth + 1 : 0;
763 node->parent = parent;
764 node->left = node->right = 0;
770 node->buf_idx = storage_idx;
771 node->offset = offset;
773 node->num_valid = (int*)cvSetNew( nv_heap );
776 node->alpha = node->node_risk = node->tree_risk = node->tree_error = 0.;
777 node->complexity = 0;
779 if( params.cv_folds > 0 && cv_heap )
781 int cv_n = params.cv_folds;
783 node->cv_Tn = (int*)cvSetNew( cv_heap );
784 node->cv_node_risk = (double*)cvAlignPtr(node->cv_Tn + cv_n, sizeof(double));
785 node->cv_node_error = node->cv_node_risk + cv_n;
791 node->cv_node_risk = 0;
792 node->cv_node_error = 0;
799 CvDTreeSplit* CvDTreeTrainData::new_split_ord( int vi, float cmp_val,
800 int split_point, int inversed, float quality )
802 CvDTreeSplit* split = (CvDTreeSplit*)cvSetNew( split_heap );
804 split->ord.c = cmp_val;
805 split->ord.split_point = split_point;
806 split->inversed = inversed;
807 split->quality = quality;
814 CvDTreeSplit* CvDTreeTrainData::new_split_cat( int vi, float quality )
816 CvDTreeSplit* split = (CvDTreeSplit*)cvSetNew( split_heap );
817 int i, n = (max_c_count + 31)/32;
821 split->quality = quality;
822 for( i = 0; i < n; i++ )
823 split->subset[i] = 0;
830 void CvDTreeTrainData::free_node( CvDTreeNode* node )
832 CvDTreeSplit* split = node->split;
833 free_node_data( node );
836 CvDTreeSplit* next = split->next;
837 cvSetRemoveByPtr( split_heap, split );
841 cvSetRemoveByPtr( node_heap, node );
845 void CvDTreeTrainData::free_node_data( CvDTreeNode* node )
847 if( node->num_valid )
849 cvSetRemoveByPtr( nv_heap, node->num_valid );
852 // do not free cv_* fields, as all the cross-validation related data is released at once.
856 void CvDTreeTrainData::free_train_data()
858 cvReleaseMat( &counts );
859 cvReleaseMat( &buf );
860 cvReleaseMat( &direction );
861 cvReleaseMat( &split_buf );
862 cvReleaseMemStorage( &temp_storage );
863 cv_heap = nv_heap = 0;
867 void CvDTreeTrainData::clear()
871 cvReleaseMemStorage( &tree_storage );
873 cvReleaseMat( &var_idx );
874 cvReleaseMat( &var_type );
875 cvReleaseMat( &cat_count );
876 cvReleaseMat( &cat_ofs );
877 cvReleaseMat( &cat_map );
878 cvReleaseMat( &priors );
880 node_heap = split_heap = 0;
882 sample_count = var_all = var_count = max_c_count = ord_var_count = cat_var_count = 0;
883 have_cv_labels = have_priors = is_classifier = false;
885 buf_count = buf_size = 0;
894 int CvDTreeTrainData::get_num_classes() const
896 return is_classifier ? cat_count->data.i[cat_var_count] : 0;
900 int CvDTreeTrainData::get_var_type(int vi) const
902 return var_type->data.i[vi];
906 CvPair32s32f* CvDTreeTrainData::get_ord_var_data( CvDTreeNode* n, int vi )
908 int oi = ~get_var_type(vi);
909 assert( 0 <= oi && oi < ord_var_count );
910 return (CvPair32s32f*)(buf->data.i + n->buf_idx*buf->cols +
911 n->offset + oi*n->sample_count*2);
915 int* CvDTreeTrainData::get_class_labels( CvDTreeNode* n )
917 return get_cat_var_data( n, var_count );
921 float* CvDTreeTrainData::get_ord_responses( CvDTreeNode* n )
923 return (float*)get_cat_var_data( n, var_count );
927 int* CvDTreeTrainData::get_cv_labels( CvDTreeNode* n )
929 return params.cv_folds > 0 ? get_cat_var_data( n, var_count + 1 ) : 0;
933 int* CvDTreeTrainData::get_cat_var_data( CvDTreeNode* n, int vi )
935 int ci = get_var_type(vi);
936 assert( 0 <= ci && ci <= cat_var_count + 1 );
937 return buf->data.i + n->buf_idx*buf->cols + n->offset +
938 (ord_var_count*2 + ci)*n->sample_count;
942 float* CvDTreeTrainData::get_weights( CvDTreeNode* n )
944 return have_weights ?
945 (float*)get_cat_var_data( n, var_count + 1 + (params.cv_folds > 0) ) : 0;
949 int CvDTreeTrainData::get_child_buf_idx( CvDTreeNode* n )
951 int idx = n->buf_idx + 1;
952 if( idx >= buf_count )
953 idx = shared ? 1 : 0;
958 /////////////////////// Decision Tree /////////////////////////
964 default_model_name = "my_tree";
970 void CvDTree::clear()
972 cvReleaseMat( &var_importance );
982 pruned_tree_idx = -1;
992 const CvDTreeNode* CvDTree::get_root() const
998 int CvDTree::get_pruned_tree_idx() const
1000 return pruned_tree_idx;
1004 CvDTreeTrainData* CvDTree::get_data()
1010 bool CvDTree::train( const CvMat* _train_data, int _tflag,
1011 const CvMat* _responses, const CvMat* _var_idx,
1012 const CvMat* _sample_idx, const CvMat* _var_type,
1013 const CvMat* _missing_mask, CvDTreeParams _params )
1015 bool result = false;
1017 CV_FUNCNAME( "CvDTree::train" );
1022 data = new CvDTreeTrainData( _train_data, _tflag, _responses,
1023 _var_idx, _sample_idx, _var_type,
1024 _missing_mask, _params, false );
1025 CV_CALL( result = do_train(0));
1033 bool CvDTree::train( CvDTreeTrainData* _data, const CvMat* _subsample_idx )
1035 bool result = false;
1037 CV_FUNCNAME( "CvDTree::train" );
1043 data->shared = true;
1044 CV_CALL( result = do_train(_subsample_idx));
1052 bool CvDTree::do_train( const CvMat* _subsample_idx )
1054 bool result = false;
1056 CV_FUNCNAME( "CvDTree::do_train" );
1060 root = data->subsample_data( _subsample_idx );
1062 CV_CALL( try_split_node(root));
1064 if( data->params.cv_folds > 0 )
1065 CV_CALL( prune_cv());
1068 data->free_train_data();
1078 #define DTREE_CAT_DIR(idx,subset) \
1079 (2*((subset[(idx)>>5]&(1 << ((idx) & 31)))==0)-1)
1081 void CvDTree::try_split_node( CvDTreeNode* node )
1083 CvDTreeSplit* best_split = 0;
1084 int i, n = node->sample_count, vi;
1085 bool can_split = true;
1086 double quality_scale;
1088 calc_node_value( node );
1090 if( node->sample_count <= data->params.min_sample_count ||
1091 node->depth >= data->params.max_depth )
1094 if( can_split && data->is_classifier )
1096 // check if we have a "pure" node,
1097 // we assume that cls_count is filled by calc_node_value()
1098 int* cls_count = data->counts->data.i;
1099 int nz = 0, m = data->get_num_classes();
1100 for( i = 0; i < m; i++ )
1101 nz += cls_count[i] != 0;
1102 if( nz == 1 ) // there is only one class
1105 else if( can_split )
1107 const float* responses = data->get_ord_responses( node );
1108 float diff = responses[n-1] - responses[0];
1109 if( diff < data->params.regression_accuracy )
1115 best_split = find_best_split(node);
1116 // TODO: check the split quality ...
1117 node->split = best_split;
1120 if( !can_split || !best_split )
1122 data->free_node_data(node);
1126 quality_scale = calc_node_dir( node );
1128 if( data->params.use_surrogates )
1130 // find all the surrogate splits
1131 // and sort them by their similarity to the primary one
1132 for( vi = 0; vi < data->var_count; vi++ )
1134 CvDTreeSplit* split;
1135 int ci = data->get_var_type(vi);
1137 if( vi == best_split->var_idx )
1141 split = find_surrogate_split_cat( node, vi );
1143 split = find_surrogate_split_ord( node, vi );
1148 CvDTreeSplit* prev_split = node->split;
1149 split->quality = (float)(split->quality*quality_scale);
1151 while( prev_split->next &&
1152 prev_split->next->quality > split->quality )
1153 prev_split = prev_split->next;
1154 split->next = prev_split->next;
1155 prev_split->next = split;
1160 split_node_data( node );
1161 try_split_node( node->left );
1162 try_split_node( node->right );
1166 // calculate direction (left(-1),right(1),missing(0))
1167 // for each sample using the best split
1168 // the function returns scale coefficients for surrogate split quality factors.
1169 // the scale is applied to normalize surrogate split quality relatively to the
1170 // best (primary) split quality. That is, if a surrogate split is absolutely
1171 // identical to the primary split, its quality will be set to the maximum value =
1172 // quality of the primary split; otherwise, it will be lower.
1173 // besides, the function compute node->maxlr,
1174 // minimum possible quality (w/o considering the above mentioned scale)
1175 // for a surrogate split. Surrogate splits with quality less than node->maxlr
1176 // are not discarded.
1177 double CvDTree::calc_node_dir( CvDTreeNode* node )
1179 char* dir = (char*)data->direction->data.ptr;
1180 int i, n = node->sample_count, vi = node->split->var_idx;
1183 assert( !node->split->inversed );
1185 if( data->get_var_type(vi) >= 0 ) // split on categorical var
1187 const int* labels = data->get_cat_var_data(node,vi);
1188 const int* subset = node->split->subset;
1190 if( !data->have_priors )
1192 int sum = 0, sum_abs = 0;
1194 for( i = 0; i < n; i++ )
1196 int idx = labels[i];
1197 int d = idx >= 0 ? DTREE_CAT_DIR(idx,subset) : 0;
1198 sum += d; sum_abs += d & 1;
1202 R = (sum_abs + sum) >> 1;
1203 L = (sum_abs - sum) >> 1;
1207 const int* responses = data->get_class_labels(node);
1208 const double* priors = data->priors->data.db;
1209 double sum = 0, sum_abs = 0;
1211 for( i = 0; i < n; i++ )
1213 int idx = labels[i];
1214 double w = priors[responses[i]];
1215 int d = idx >= 0 ? DTREE_CAT_DIR(idx,subset) : 0;
1216 sum += d*w; sum_abs += (d & 1)*w;
1220 R = (sum_abs + sum) * 0.5;
1221 L = (sum_abs - sum) * 0.5;
1224 else // split on ordered var
1226 const CvPair32s32f* sorted = data->get_ord_var_data(node,vi);
1227 int split_point = node->split->ord.split_point;
1228 int n1 = node->get_num_valid(vi);
1230 assert( 0 <= split_point && split_point < n1-1 );
1232 if( !data->have_priors )
1234 for( i = 0; i <= split_point; i++ )
1235 dir[sorted[i].i] = (char)-1;
1236 for( ; i < n1; i++ )
1237 dir[sorted[i].i] = (char)1;
1239 dir[sorted[i].i] = (char)0;
1242 R = n1 - split_point + 1;
1246 const int* responses = data->get_class_labels(node);
1247 const double* priors = data->priors->data.db;
1250 for( i = 0; i <= split_point; i++ )
1252 int idx = sorted[i].i;
1253 double w = priors[responses[idx]];
1254 dir[idx] = (char)-1;
1258 for( ; i < n1; i++ )
1260 int idx = sorted[i].i;
1261 double w = priors[responses[idx]];
1267 dir[sorted[i].i] = (char)0;
1271 node->maxlr = MAX( L, R );
1272 return node->split->quality/(L + R);
1276 CvDTreeSplit* CvDTree::find_best_split( CvDTreeNode* node )
1279 CvDTreeSplit *best_split = 0, *split = 0, *t;
1281 for( vi = 0; vi < data->var_count; vi++ )
1283 int ci = data->get_var_type(vi);
1284 if( node->get_num_valid(vi) <= 1 )
1287 if( data->is_classifier )
1290 split = find_split_cat_class( node, vi );
1292 split = find_split_ord_class( node, vi );
1297 split = find_split_cat_reg( node, vi );
1299 split = find_split_ord_reg( node, vi );
1304 if( !best_split || best_split->quality < split->quality )
1305 CV_SWAP( best_split, split, t );
1307 cvSetRemoveByPtr( data->split_heap, split );
1315 CvDTreeSplit* CvDTree::find_split_ord_class( CvDTreeNode* node, int vi )
1317 const float epsilon = FLT_EPSILON*2;
1318 const CvPair32s32f* sorted = data->get_ord_var_data(node, vi);
1319 const int* responses = data->get_class_labels(node);
1320 int n = node->sample_count;
1321 int n1 = node->get_num_valid(vi);
1322 int m = data->get_num_classes();
1323 const int* rc0 = data->counts->data.i;
1324 int* lc = (int*)(rc0 + m);
1327 double lsum2 = 0, rsum2 = 0, best_val = 0;
1328 const double* priors = data->have_priors ? data->priors->data.db : 0;
1330 // init arrays of class instance counters on both sides of the split
1331 for( i = 0; i < m; i++ )
1337 // compensate for missing values
1338 for( i = n1; i < n; i++ )
1339 rc[responses[sorted[i].i]]--;
1345 for( i = 0; i < m; i++ )
1346 rsum2 += (double)rc[i]*rc[i];
1348 for( i = 0; i < n1 - 1; i++ )
1350 int idx = responses[sorted[i].i];
1353 lv = lc[idx]; rv = rc[idx];
1356 lc[idx] = lv + 1; rc[idx] = rv - 1;
1358 if( sorted[i].val + epsilon < sorted[i+1].val )
1360 double val = lsum2/L + rsum2/R;
1361 if( best_val < val )
1371 double L = 0, R = 0;
1372 for( i = 0; i < m; i++ )
1374 double wv = rc[i]*priors[i];
1379 for( i = 0; i < n1 - 1; i++ )
1381 int idx = responses[sorted[i].i];
1383 double p = priors[idx], p2 = p*p;
1385 lv = lc[idx]; rv = rc[idx];
1386 lsum2 += p2*(lv*2 + 1);
1387 rsum2 -= p2*(rv*2 - 1);
1388 lc[idx] = lv + 1; rc[idx] = rv - 1;
1390 if( sorted[i].val + epsilon < sorted[i+1].val )
1392 double val = lsum2/L + rsum2/R;
1393 if( best_val < val )
1402 return best_i >= 0 ? data->new_split_ord( vi,
1403 (sorted[best_i].val + sorted[best_i+1].val)*0.5f, best_i,
1404 0, (float)best_val ) : 0;
1408 void CvDTree::cluster_categories( const int* vectors, int n, int m,
1409 int* csums, int k, int* labels )
1411 // TODO: consider adding priors (class weights) and sample weights to the clustering algorithm
1412 int iters = 0, max_iters = 100;
1414 double* buf = (double*)cvStackAlloc( (n + k)*sizeof(buf[0]) );
1415 double *v_weights = buf, *c_weights = buf + k;
1416 bool modified = true;
1417 CvRNG* r = &data->rng;
1419 // assign labels randomly
1420 for( i = idx = 0; i < n; i++ )
1423 const int* v = vectors + i*m;
1425 idx &= idx < k ? -1 : 0;
1427 // compute weight of each vector
1428 for( j = 0; j < m; j++ )
1430 v_weights[i] = sum ? 1./sum : 0.;
1433 for( i = 0; i < n; i++ )
1435 int i1 = cvRandInt(r) % n;
1436 int i2 = cvRandInt(r) % n;
1437 CV_SWAP( labels[i1], labels[i2], j );
1440 for( iters = 0; iters <= max_iters; iters++ )
1443 for( i = 0; i < k; i++ )
1445 for( j = 0; j < m; j++ )
1449 for( i = 0; i < n; i++ )
1451 const int* v = vectors + i*m;
1452 int* s = csums + labels[i]*m;
1453 for( j = 0; j < m; j++ )
1457 // exit the loop here, when we have up-to-date csums
1458 if( iters == max_iters || !modified )
1463 // calculate weight of each cluster
1464 for( i = 0; i < k; i++ )
1466 const int* s = csums + i*m;
1468 for( j = 0; j < m; j++ )
1470 c_weights[i] = sum ? 1./sum : 0;
1473 // now for each vector determine the closest cluster
1474 for( i = 0; i < n; i++ )
1476 const int* v = vectors + i*m;
1477 double alpha = v_weights[i];
1478 double min_dist2 = DBL_MAX;
1481 for( idx = 0; idx < k; idx++ )
1483 const int* s = csums + idx*m;
1484 double dist2 = 0., beta = c_weights[idx];
1485 for( j = 0; j < m; j++ )
1487 double t = v[j]*alpha - s[j]*beta;
1490 if( min_dist2 > dist2 )
1497 if( min_idx != labels[i] )
1499 labels[i] = min_idx;
1505 CvDTreeSplit* CvDTree::find_split_cat_class( CvDTreeNode* node, int vi )
1507 CvDTreeSplit* split;
1508 const int* labels = data->get_cat_var_data(node, vi);
1509 const int* responses = data->get_class_labels(node);
1510 int ci = data->get_var_type(vi);
1511 int n = node->sample_count;
1512 int m = data->get_num_classes();
1513 int _mi = data->cat_count->data.i[ci], mi = _mi;
1514 const int* rc0 = data->counts->data.i;
1515 int* lc = (int*)(rc0 + m);
1517 int* _cjk = rc + m*2, *cjk = _cjk;
1518 double* c_weights = (double*)cvStackAlloc( mi*sizeof(c_weights[0]) );
1519 int* cluster_labels = 0;
1522 double L = 0, R = 0;
1523 double best_val = 0;
1524 int prevcode = 0, best_subset = -1, subset_i, subset_n, subtract = 0;
1525 const double* priors = data->priors->data.db;
1527 // init array of counters:
1528 // c_{jk} - number of samples that have vi-th input variable = j and response = k.
1529 for( j = -1; j < mi; j++ )
1530 for( k = 0; k < m; k++ )
1533 for( i = 0; i < n; i++ )
1542 if( mi > data->params.max_categories )
1544 mi = MIN(data->params.max_categories, n);
1546 cluster_labels = cjk + mi*m;
1547 cluster_categories( _cjk, _mi, m, cjk, mi, cluster_labels );
1555 int_ptr = (int**)cvStackAlloc( mi*sizeof(int_ptr[0]) );
1556 for( j = 0; j < mi; j++ )
1557 int_ptr[j] = cjk + j*2 + 1;
1558 icvSortIntPtr( int_ptr, mi, 0 );
1563 for( k = 0; k < m; k++ )
1566 for( j = 0; j < mi; j++ )
1567 sum += cjk[j*m + k];
1572 for( j = 0; j < mi; j++ )
1575 for( k = 0; k < m; k++ )
1576 sum += cjk[j*m + k]*priors[k];
1581 for( ; subset_i < subset_n; subset_i++ )
1585 double lsum2 = 0, rsum2 = 0;
1588 idx = (int)(int_ptr[subset_i] - cjk)/2;
1591 int graycode = (subset_i>>1)^subset_i;
1592 int diff = graycode ^ prevcode;
1594 // determine index of the changed bit.
1596 idx = diff >= (1 << 16) ? 16 : 0;
1597 u.f = (float)(((diff >> 16) | diff) & 65535);
1598 idx += (u.i >> 23) - 127;
1599 subtract = graycode < prevcode;
1600 prevcode = graycode;
1604 weight = c_weights[idx];
1605 if( weight < FLT_EPSILON )
1610 for( k = 0; k < m; k++ )
1613 int lval = lc[k] + t;
1614 int rval = rc[k] - t;
1615 double p = priors[k], p2 = p*p;
1616 lsum2 += p2*lval*lval;
1617 rsum2 += p2*rval*rval;
1618 lc[k] = lval; rc[k] = rval;
1625 for( k = 0; k < m; k++ )
1628 int lval = lc[k] - t;
1629 int rval = rc[k] + t;
1630 double p = priors[k], p2 = p*p;
1631 lsum2 += p2*lval*lval;
1632 rsum2 += p2*rval*rval;
1633 lc[k] = lval; rc[k] = rval;
1639 if( L > FLT_EPSILON && R > FLT_EPSILON )
1641 double val = lsum2/L + rsum2/R;
1642 if( best_val < val )
1645 best_subset = subset_i;
1650 if( best_subset < 0 )
1653 split = data->new_split_cat( vi, (float)best_val );
1657 for( i = 0; i <= best_subset; i++ )
1659 idx = (int)(int_ptr[i] - cjk) >> 1;
1660 split->subset[idx >> 5] |= 1 << (idx & 31);
1665 for( i = 0; i < _mi; i++ )
1667 idx = cluster_labels ? cluster_labels[i] : i;
1668 if( best_subset & (1 << idx) )
1669 split->subset[i >> 5] |= 1 << (i & 31);
1677 CvDTreeSplit* CvDTree::find_split_ord_reg( CvDTreeNode* node, int vi )
1679 const float epsilon = FLT_EPSILON*2;
1680 const CvPair32s32f* sorted = data->get_ord_var_data(node, vi);
1681 const float* responses = data->get_ord_responses(node);
1682 int n = node->sample_count;
1683 int n1 = node->get_num_valid(vi);
1685 double best_val = 0, lsum = 0, rsum = node->value*n;
1688 // compensate for missing values
1689 for( i = n1; i < n; i++ )
1690 rsum -= responses[sorted[i].i];
1692 // find the optimal split
1693 for( i = 0; i < n1 - 1; i++ )
1695 float t = responses[sorted[i].i];
1700 if( sorted[i].val + epsilon < sorted[i+1].val )
1702 double val = lsum*lsum/L + rsum*rsum/R;
1703 if( best_val < val )
1711 return best_i >= 0 ? data->new_split_ord( vi,
1712 (sorted[best_i].val + sorted[best_i+1].val)*0.5f, best_i,
1713 0, (float)best_val ) : 0;
1717 CvDTreeSplit* CvDTree::find_split_cat_reg( CvDTreeNode* node, int vi )
1719 CvDTreeSplit* split;
1720 const int* labels = data->get_cat_var_data(node, vi);
1721 const float* responses = data->get_ord_responses(node);
1722 int ci = data->get_var_type(vi);
1723 int n = node->sample_count;
1724 int mi = data->cat_count->data.i[ci];
1725 double* sum = (double*)cvStackAlloc( (mi+1)*sizeof(sum[0]) ) + 1;
1726 int* counts = (int*)cvStackAlloc( (mi+1)*sizeof(counts[0]) ) + 1;
1727 double** sum_ptr = 0;
1728 int i, L = 0, R = 0;
1729 double best_val = 0, lsum = 0, rsum = 0;
1730 int best_subset = -1, subset_i;
1732 for( i = -1; i < mi; i++ )
1733 sum[i] = counts[i] = 0;
1735 // calculate sum response and weight of each category of the input var
1736 for( i = 0; i < n; i++ )
1738 int idx = labels[i];
1739 double s = sum[idx] + responses[i];
1740 int nc = counts[idx] + 1;
1745 // calculate average response in each category
1746 for( i = 0; i < mi; i++ )
1750 sum[i] /= MAX(counts[i],1);
1751 sum_ptr[i] = sum + i;
1754 icvSortDblPtr( sum_ptr, mi, 0 );
1756 // revert back to unnormalized sum
1757 // (there should be a very little loss of accuracy)
1758 for( i = 0; i < mi; i++ )
1759 sum[i] *= counts[i];
1761 for( subset_i = 0; subset_i < mi-1; subset_i++ )
1763 int idx = (int)(sum_ptr[subset_i] - sum);
1764 int ni = counts[idx];
1768 double s = sum[idx];
1774 double val = lsum*lsum/L + rsum*rsum/R;
1775 if( best_val < val )
1778 best_subset = subset_i;
1784 if( best_subset < 0 )
1787 split = data->new_split_cat( vi, (float)best_val );
1788 for( i = 0; i <= best_subset; i++ )
1790 int idx = (int)(sum_ptr[i] - sum);
1791 split->subset[idx >> 5] |= 1 << (idx & 31);
1798 CvDTreeSplit* CvDTree::find_surrogate_split_ord( CvDTreeNode* node, int vi )
1800 const float epsilon = FLT_EPSILON*2;
1801 const CvPair32s32f* sorted = data->get_ord_var_data(node, vi);
1802 const char* dir = (char*)data->direction->data.ptr;
1803 int n1 = node->get_num_valid(vi);
1804 // LL - number of samples that both the primary and the surrogate splits send to the left
1805 // LR - ... primary split sends to the left and the surrogate split sends to the right
1806 // RL - ... primary split sends to the right and the surrogate split sends to the left
1807 // RR - ... both send to the right
1808 int i, best_i = -1, best_inversed = 0;
1811 if( !data->have_priors )
1813 int LL = 0, RL = 0, LR, RR;
1814 int worst_val = cvFloor(node->maxlr), _best_val = worst_val;
1815 int sum = 0, sum_abs = 0;
1817 for( i = 0; i < n1; i++ )
1819 int d = dir[sorted[i].i];
1820 sum += d; sum_abs += d & 1;
1823 // sum_abs = R + L; sum = R - L
1824 RR = (sum_abs + sum) >> 1;
1825 LR = (sum_abs - sum) >> 1;
1827 // initially all the samples are sent to the right by the surrogate split,
1828 // LR of them are sent to the left by primary split, and RR - to the right.
1829 // now iteratively compute LL, LR, RL and RR for every possible surrogate split value.
1830 for( i = 0; i < n1 - 1; i++ )
1832 int d = dir[sorted[i].i];
1837 if( LL + RR > _best_val && sorted[i].val + epsilon < sorted[i+1].val )
1840 best_i = i; best_inversed = 0;
1846 if( RL + LR > _best_val && sorted[i].val + epsilon < sorted[i+1].val )
1849 best_i = i; best_inversed = 1;
1853 best_val = _best_val;
1857 double LL = 0, RL = 0, LR, RR;
1858 double worst_val = node->maxlr;
1859 double sum = 0, sum_abs = 0;
1860 const double* priors = data->priors->data.db;
1861 const int* responses = data->get_class_labels(node);
1862 best_val = worst_val;
1864 for( i = 0; i < n1; i++ )
1866 int idx = sorted[i].i;
1867 double w = priors[responses[idx]];
1869 sum += d*w; sum_abs += (d & 1)*w;
1872 // sum_abs = R + L; sum = R - L
1873 RR = (sum_abs + sum)*0.5;
1874 LR = (sum_abs - sum)*0.5;
1876 // initially all the samples are sent to the right by the surrogate split,
1877 // LR of them are sent to the left by primary split, and RR - to the right.
1878 // now iteratively compute LL, LR, RL and RR for every possible surrogate split value.
1879 for( i = 0; i < n1 - 1; i++ )
1881 int idx = sorted[i].i;
1882 double w = priors[responses[idx]];
1888 if( LL + RR > best_val && sorted[i].val + epsilon < sorted[i+1].val )
1891 best_i = i; best_inversed = 0;
1897 if( RL + LR > best_val && sorted[i].val + epsilon < sorted[i+1].val )
1900 best_i = i; best_inversed = 1;
1906 return best_i >= 0 ? data->new_split_ord( vi,
1907 (sorted[best_i].val + sorted[best_i+1].val)*0.5f, best_i,
1908 best_inversed, (float)best_val ) : 0;
1912 CvDTreeSplit* CvDTree::find_surrogate_split_cat( CvDTreeNode* node, int vi )
1914 const int* labels = data->get_cat_var_data(node, vi);
1915 const char* dir = (char*)data->direction->data.ptr;
1916 int n = node->sample_count;
1917 // LL - number of samples that both the primary and the surrogate splits send to the left
1918 // LR - ... primary split sends to the left and the surrogate split sends to the right
1919 // RL - ... primary split sends to the right and the surrogate split sends to the left
1920 // RR - ... both send to the right
1921 CvDTreeSplit* split = data->new_split_cat( vi, 0 );
1922 int i, mi = data->cat_count->data.i[data->get_var_type(vi)];
1923 double best_val = 0;
1924 double* lc = (double*)cvStackAlloc( (mi+1)*2*sizeof(lc[0]) ) + 1;
1925 double* rc = lc + mi + 1;
1927 for( i = -1; i < mi; i++ )
1930 // for each category calculate the weight of samples
1931 // sent to the left (lc) and to the right (rc) by the primary split
1932 if( !data->have_priors )
1934 int* _lc = data->counts->data.i + 1;
1935 int* _rc = _lc + mi + 1;
1937 for( i = -1; i < mi; i++ )
1938 _lc[i] = _rc[i] = 0;
1940 for( i = 0; i < n; i++ )
1942 int idx = labels[i];
1944 int sum = _lc[idx] + d;
1945 int sum_abs = _rc[idx] + (d & 1);
1946 _lc[idx] = sum; _rc[idx] = sum_abs;
1949 for( i = 0; i < mi; i++ )
1952 int sum_abs = _rc[i];
1953 lc[i] = (sum_abs - sum) >> 1;
1954 rc[i] = (sum_abs + sum) >> 1;
1959 const double* priors = data->priors->data.db;
1960 const int* responses = data->get_class_labels(node);
1962 for( i = 0; i < n; i++ )
1964 int idx = labels[i];
1965 double w = priors[responses[i]];
1967 double sum = lc[idx] + d*w;
1968 double sum_abs = rc[idx] + (d & 1)*w;
1969 lc[idx] = sum; rc[idx] = sum_abs;
1972 for( i = 0; i < mi; i++ )
1975 double sum_abs = rc[i];
1976 lc[i] = (sum_abs - sum) * 0.5;
1977 rc[i] = (sum_abs + sum) * 0.5;
1981 // 2. now form the split.
1982 // in each category send all the samples to the same direction as majority
1983 for( i = 0; i < mi; i++ )
1985 double lval = lc[i], rval = rc[i];
1988 split->subset[i >> 5] |= 1 << (i & 31);
1995 split->quality = (float)best_val;
1996 if( split->quality <= node->maxlr )
1997 cvSetRemoveByPtr( data->split_heap, split ), split = 0;
2003 void CvDTree::calc_node_value( CvDTreeNode* node )
2005 int i, j, k, n = node->sample_count, cv_n = data->params.cv_folds;
2006 const int* cv_labels = data->get_cv_labels(node);
2008 if( data->is_classifier )
2010 // in case of classification tree:
2011 // * node value is the label of the class that has the largest weight in the node.
2012 // * node risk is the weighted number of misclassified samples,
2013 // * j-th cross-validation fold value and risk are calculated as above,
2014 // but using the samples with cv_labels(*)!=j.
2015 // * j-th cross-validation fold error is calculated as the weighted number of
2016 // misclassified samples with cv_labels(*)==j.
2018 // compute the number of instances of each class
2019 int* cls_count = data->counts->data.i;
2020 const int* responses = data->get_class_labels(node);
2021 int m = data->get_num_classes();
2022 int* cv_cls_count = cls_count + m;
2023 double max_val = -1, total_weight = 0;
2025 double* priors = data->priors->data.db;
2027 for( k = 0; k < m; k++ )
2032 for( i = 0; i < n; i++ )
2033 cls_count[responses[i]]++;
2037 for( j = 0; j < cv_n; j++ )
2038 for( k = 0; k < m; k++ )
2039 cv_cls_count[j*m + k] = 0;
2041 for( i = 0; i < n; i++ )
2043 j = cv_labels[i]; k = responses[i];
2044 cv_cls_count[j*m + k]++;
2047 for( j = 0; j < cv_n; j++ )
2048 for( k = 0; k < m; k++ )
2049 cls_count[k] += cv_cls_count[j*m + k];
2052 for( k = 0; k < m; k++ )
2054 double val = cls_count[k]*priors[k];
2055 total_weight += val;
2063 node->class_idx = max_k;
2064 node->value = data->cat_map->data.i[
2065 data->cat_ofs->data.i[data->cat_var_count] + max_k];
2066 node->node_risk = total_weight - max_val;
2068 for( j = 0; j < cv_n; j++ )
2070 double sum_k = 0, sum = 0, max_val_k = 0;
2071 max_val = -1; max_k = -1;
2073 for( k = 0; k < m; k++ )
2075 double w = priors[k];
2076 double val_k = cv_cls_count[j*m + k]*w;
2077 double val = cls_count[k]*w - val_k;
2088 node->cv_Tn[j] = INT_MAX;
2089 node->cv_node_risk[j] = sum - max_val;
2090 node->cv_node_error[j] = sum_k - max_val_k;
2095 // in case of regression tree:
2096 // * node value is 1/n*sum_i(Y_i), where Y_i is i-th response,
2097 // n is the number of samples in the node.
2098 // * node risk is the sum of squared errors: sum_i((Y_i - <node_value>)^2)
2099 // * j-th cross-validation fold value and risk are calculated as above,
2100 // but using the samples with cv_labels(*)!=j.
2101 // * j-th cross-validation fold error is calculated
2102 // using samples with cv_labels(*)==j as the test subset:
2103 // error_j = sum_(i,cv_labels(i)==j)((Y_i - <node_value_j>)^2),
2104 // where node_value_j is the node value calculated
2105 // as described in the previous bullet, and summation is done
2106 // over the samples with cv_labels(*)==j.
2108 double sum = 0, sum2 = 0;
2109 const float* values = data->get_ord_responses(node);
2110 double *cv_sum = 0, *cv_sum2 = 0;
2115 // if cross-validation is not used, we even do not compute node_risk
2116 // (so the tree sequence T1>...>{root} may not be built).
2117 for( i = 0; i < n; i++ )
2122 cv_sum = (double*)cvStackAlloc( cv_n*sizeof(cv_sum[0]) );
2123 cv_sum2 = (double*)cvStackAlloc( cv_n*sizeof(cv_sum2[0]) );
2124 cv_count = (int*)cvStackAlloc( cv_n*sizeof(cv_count[0]) );
2126 for( j = 0; j < cv_n; j++ )
2128 cv_sum[j] = cv_sum2[j] = 0.;
2132 for( i = 0; i < n; i++ )
2135 double t = values[i];
2136 double s = cv_sum[j] + t;
2137 double s2 = cv_sum2[j] + t*t;
2138 int nc = cv_count[j] + 1;
2144 for( j = 0; j < cv_n; j++ )
2150 node->node_risk = sum2 - (sum/n)*sum;
2153 node->value = sum/n;
2155 for( j = 0; j < cv_n; j++ )
2157 double s = cv_sum[j], si = sum - s;
2158 double s2 = cv_sum2[j], s2i = sum2 - s2;
2159 int c = cv_count[j], ci = n - c;
2160 double r = si/MAX(ci,1);
2161 node->cv_node_risk[j] = s2i - r*r*ci;
2162 node->cv_node_error[j] = s2 - 2*r*s + c*r*r;
2163 node->cv_Tn[j] = INT_MAX;
2169 void CvDTree::complete_node_dir( CvDTreeNode* node )
2171 int vi, i, n = node->sample_count, nl, nr, d0 = 0, d1 = -1;
2172 int nz = n - node->get_num_valid(node->split->var_idx);
2173 char* dir = (char*)data->direction->data.ptr;
2175 // try to complete direction using surrogate splits
2176 if( nz && data->params.use_surrogates )
2178 CvDTreeSplit* split = node->split->next;
2179 for( ; split != 0 && nz; split = split->next )
2181 int inversed_mask = split->inversed ? -1 : 0;
2182 vi = split->var_idx;
2184 if( data->get_var_type(vi) >= 0 ) // split on categorical var
2186 const int* labels = data->get_cat_var_data(node, vi);
2187 const int* subset = split->subset;
2189 for( i = 0; i < n; i++ )
2192 if( !dir[i] && (idx = labels[i]) >= 0 )
2194 int d = DTREE_CAT_DIR(idx,subset);
2195 dir[i] = (char)((d ^ inversed_mask) - inversed_mask);
2201 else // split on ordered var
2203 const CvPair32s32f* sorted = data->get_ord_var_data(node, vi);
2204 int split_point = split->ord.split_point;
2205 int n1 = node->get_num_valid(vi);
2207 assert( 0 <= split_point && split_point < n-1 );
2209 for( i = 0; i < n1; i++ )
2211 int idx = sorted[i].i;
2214 int d = i <= split_point ? -1 : 1;
2215 dir[idx] = (char)((d ^ inversed_mask) - inversed_mask);
2224 // find the default direction for the rest
2227 for( i = nr = 0; i < n; i++ )
2230 d0 = nl > nr ? -1 : nr > nl;
2233 // make sure that every sample is directed either to the left or to the right
2234 for( i = 0; i < n; i++ )
2244 dir[i] = (char)d; // remap (-1,1) to (0,1)
2249 void CvDTree::split_node_data( CvDTreeNode* node )
2251 int vi, i, n = node->sample_count, nl, nr;
2252 char* dir = (char*)data->direction->data.ptr;
2253 CvDTreeNode *left = 0, *right = 0;
2254 int* new_idx = data->split_buf->data.i;
2255 int new_buf_idx = data->get_child_buf_idx( node );
2257 complete_node_dir(node);
2259 for( i = nl = nr = 0; i < n; i++ )
2262 // initialize new indices for splitting ordered variables
2263 new_idx[i] = (nl & (d-1)) | (nr & -d); // d ? ri : li
2268 node->left = left = data->new_node( node, nl, new_buf_idx, node->offset );
2269 node->right = right = data->new_node( node, nr, new_buf_idx, node->offset +
2270 (data->ord_var_count*2 + data->cat_var_count+1+data->have_cv_labels)*nl );
2272 // split ordered variables, keep both halves sorted.
2273 for( vi = 0; vi < data->var_count; vi++ )
2275 int ci = data->get_var_type(vi);
2276 int n1 = node->get_num_valid(vi);
2277 CvPair32s32f *src, *ldst0, *rdst0, *ldst, *rdst;
2278 CvPair32s32f tl, tr;
2283 src = data->get_ord_var_data(node, vi);
2284 ldst0 = ldst = data->get_ord_var_data(left, vi);
2285 rdst0 = rdst = data->get_ord_var_data(right, vi);
2286 tl = ldst0[nl]; tr = rdst0[nr];
2289 for( i = 0; i < n1; i++ )
2292 float val = src[i].val;
2295 ldst->i = rdst->i = idx;
2296 ldst->val = rdst->val = val;
2301 left->set_num_valid(vi, (int)(ldst - ldst0));
2302 right->set_num_valid(vi, (int)(rdst - rdst0));
2310 ldst->i = rdst->i = idx;
2311 ldst->val = rdst->val = ord_nan;
2316 ldst0[nl] = tl; rdst0[nr] = tr;
2319 // split categorical vars, responses and cv_labels using new_idx relocation table
2320 for( vi = 0; vi <= data->var_count + data->have_cv_labels + data->have_weights; vi++ )
2322 int ci = data->get_var_type(vi);
2323 int n1 = node->get_num_valid(vi), nr1 = 0;
2324 int *src, *ldst0, *rdst0, *ldst, *rdst;
2330 src = data->get_cat_var_data(node, vi);
2331 ldst0 = ldst = data->get_cat_var_data(left, vi);
2332 rdst0 = rdst = data->get_cat_var_data(right, vi);
2333 tl = ldst0[nl]; tr = rdst0[nr];
2335 for( i = 0; i < n; i++ )
2339 *ldst = *rdst = val;
2342 nr1 += (val >= 0)&d;
2345 if( vi < data->var_count )
2347 left->set_num_valid(vi, n1 - nr1);
2348 right->set_num_valid(vi, nr1);
2351 ldst0[nl] = tl; rdst0[nr] = tr;
2354 // deallocate the parent node data that is not needed anymore
2355 data->free_node_data(node);
2359 void CvDTree::prune_cv()
2365 // 1. build tree sequence for each cv fold, calculate error_{Tj,beta_k}.
2366 // 2. choose the best tree index (if need, apply 1SE rule).
2367 // 3. store the best index and cut the branches.
2369 CV_FUNCNAME( "CvDTree::prune_cv" );
2373 int ti, j, tree_count = 0, cv_n = data->params.cv_folds, n = root->sample_count;
2374 // currently, 1SE for regression is not implemented
2375 bool use_1se = data->params.use_1se_rule != 0 && data->is_classifier;
2377 double min_err = 0, min_err_se = 0;
2380 CV_CALL( ab = cvCreateMat( 1, 256, CV_64F ));
2382 // build the main tree sequence, calculate alpha's
2385 double min_alpha = update_tree_rnc(tree_count, -1);
2386 if( cut_tree(tree_count, -1, min_alpha) )
2389 if( ab->cols <= tree_count )
2391 CV_CALL( temp = cvCreateMat( 1, ab->cols*3/2, CV_64F ));
2392 for( ti = 0; ti < ab->cols; ti++ )
2393 temp->data.db[ti] = ab->data.db[ti];
2394 cvReleaseMat( &ab );
2399 ab->data.db[tree_count] = min_alpha;
2402 ab->data.db[0] = 0.;
2403 for( ti = 1; ti < tree_count-1; ti++ )
2404 ab->data.db[ti] = sqrt(ab->data.db[ti]*ab->data.db[ti+1]);
2405 ab->data.db[tree_count-1] = DBL_MAX*0.5;
2407 CV_CALL( err_jk = cvCreateMat( cv_n, tree_count, CV_64F ));
2408 err = err_jk->data.db;
2410 for( j = 0; j < cv_n; j++ )
2413 for( ; tk < tree_count; tj++ )
2415 double min_alpha = update_tree_rnc(tj, j);
2416 if( cut_tree(tj, j, min_alpha) )
2417 min_alpha = DBL_MAX;
2419 for( ; tk < tree_count; tk++ )
2421 if( ab->data.db[tk] > min_alpha )
2423 err[j*tree_count + tk] = root->tree_error;
2428 for( ti = 0; ti < tree_count; ti++ )
2431 for( j = 0; j < cv_n; j++ )
2432 sum_err += err[j*tree_count + ti];
2433 if( ti == 0 || sum_err < min_err )
2438 min_err_se = sqrt( sum_err*(n - sum_err) );
2440 else if( sum_err < min_err + min_err_se )
2444 pruned_tree_idx = min_idx;
2445 free_prune_data(data->params.truncate_pruned_tree != 0);
2449 cvReleaseMat( &err_jk );
2450 cvReleaseMat( &ab );
2451 cvReleaseMat( &temp );
2455 double CvDTree::update_tree_rnc( int T, int fold )
2457 CvDTreeNode* node = root;
2458 double min_alpha = DBL_MAX;
2462 CvDTreeNode* parent;
2465 int t = fold >= 0 ? node->cv_Tn[fold] : node->Tn;
2466 if( t <= T || !node->left )
2468 node->complexity = 1;
2469 node->tree_risk = node->node_risk;
2470 node->tree_error = 0.;
2473 node->tree_risk = node->cv_node_risk[fold];
2474 node->tree_error = node->cv_node_error[fold];
2481 for( parent = node->parent; parent && parent->right == node;
2482 node = parent, parent = parent->parent )
2484 parent->complexity += node->complexity;
2485 parent->tree_risk += node->tree_risk;
2486 parent->tree_error += node->tree_error;
2488 parent->alpha = ((fold >= 0 ? parent->cv_node_risk[fold] : parent->node_risk)
2489 - parent->tree_risk)/(parent->complexity - 1);
2490 min_alpha = MIN( min_alpha, parent->alpha );
2496 parent->complexity = node->complexity;
2497 parent->tree_risk = node->tree_risk;
2498 parent->tree_error = node->tree_error;
2499 node = parent->right;
2506 int CvDTree::cut_tree( int T, int fold, double min_alpha )
2508 CvDTreeNode* node = root;
2514 CvDTreeNode* parent;
2517 int t = fold >= 0 ? node->cv_Tn[fold] : node->Tn;
2518 if( t <= T || !node->left )
2520 if( node->alpha <= min_alpha + FLT_EPSILON )
2523 node->cv_Tn[fold] = T;
2533 for( parent = node->parent; parent && parent->right == node;
2534 node = parent, parent = parent->parent )
2540 node = parent->right;
2547 void CvDTree::free_prune_data(bool cut_tree)
2549 CvDTreeNode* node = root;
2553 CvDTreeNode* parent;
2556 // do not call cvSetRemoveByPtr( cv_heap, node->cv_Tn )
2557 // as we will clear the whole cross-validation heap at the end
2559 node->cv_node_error = node->cv_node_risk = 0;
2565 for( parent = node->parent; parent && parent->right == node;
2566 node = parent, parent = parent->parent )
2568 if( cut_tree && parent->Tn <= pruned_tree_idx )
2570 data->free_node( parent->left );
2571 data->free_node( parent->right );
2572 parent->left = parent->right = 0;
2579 node = parent->right;
2583 cvClearSet( data->cv_heap );
2587 void CvDTree::free_tree()
2589 if( root && data && data->shared )
2591 pruned_tree_idx = INT_MIN;
2592 free_prune_data(true);
2593 data->free_node(root);
2599 CvDTreeNode* CvDTree::predict( const CvMat* _sample,
2600 const CvMat* _missing, bool preprocessed_input ) const
2602 CvDTreeNode* result = 0;
2605 CV_FUNCNAME( "CvDTree::predict" );
2609 int i, step, mstep = 0;
2610 const float* sample;
2612 CvDTreeNode* node = root;
2619 CV_ERROR( CV_StsError, "The tree has not been trained yet" );
2621 if( !CV_IS_MAT(_sample) || CV_MAT_TYPE(_sample->type) != CV_32FC1 ||
2622 _sample->cols != 1 && _sample->rows != 1 ||
2623 _sample->cols + _sample->rows - 1 != data->var_all && !preprocessed_input ||
2624 _sample->cols + _sample->rows - 1 != data->var_count && preprocessed_input )
2625 CV_ERROR( CV_StsBadArg,
2626 "the input sample must be 1d floating-point vector with the same "
2627 "number of elements as the total number of variables used for training" );
2629 sample = _sample->data.fl;
2630 step = CV_IS_MAT_CONT(_sample->type) ? 1 : _sample->step/sizeof(data[0]);
2632 if( data->cat_count && !preprocessed_input ) // cache for categorical variables
2634 int n = data->cat_count->cols;
2635 catbuf = (int*)cvStackAlloc(n*sizeof(catbuf[0]));
2636 for( i = 0; i < n; i++ )
2642 if( !CV_IS_MAT(_missing) || !CV_IS_MASK_ARR(_missing) ||
2643 !CV_ARE_SIZES_EQ(_missing, _sample) )
2644 CV_ERROR( CV_StsBadArg,
2645 "the missing data mask must be 8-bit vector of the same size as input sample" );
2646 m = _missing->data.ptr;
2647 mstep = CV_IS_MAT_CONT(_missing->type) ? 1 : _missing->step/sizeof(m[0]);
2650 vtype = data->var_type->data.i;
2651 vidx = data->var_idx && !preprocessed_input ? data->var_idx->data.i : 0;
2652 cmap = data->cat_map->data.i;
2653 cofs = data->cat_ofs->data.i;
2655 while( node->Tn > pruned_tree_idx && node->left )
2657 CvDTreeSplit* split = node->split;
2659 for( ; !dir && split != 0; split = split->next )
2661 int vi = split->var_idx;
2663 i = vidx ? vidx[vi] : vi;
2664 float val = sample[i*step];
2665 if( m && m[i*mstep] )
2667 if( ci < 0 ) // ordered
2668 dir = val <= split->ord.c ? -1 : 1;
2672 if( preprocessed_input )
2679 int a = c = cofs[ci];
2681 int ival = cvRound(val);
2683 CV_ERROR( CV_StsBadArg,
2684 "one of input categorical variable is not an integer" );
2689 if( ival < cmap[c] )
2691 else if( ival > cmap[c] )
2697 if( c < 0 || ival != cmap[c] )
2700 catbuf[ci] = c -= cofs[ci];
2703 dir = DTREE_CAT_DIR(c, split->subset);
2706 if( split->inversed )
2712 double diff = node->right->sample_count - node->left->sample_count;
2713 dir = diff < 0 ? -1 : 1;
2715 node = dir < 0 ? node->left : node->right;
2726 const CvMat* CvDTree::get_var_importance()
2728 if( !var_importance )
2730 CvDTreeNode* node = root;
2734 var_importance = cvCreateMat( 1, data->var_count, CV_64F );
2735 cvZero( var_importance );
2736 importance = var_importance->data.db;
2740 CvDTreeNode* parent;
2741 for( ;; node = node->left )
2743 CvDTreeSplit* split = node->split;
2745 if( !node->left || node->Tn <= pruned_tree_idx )
2748 for( ; split != 0; split = split->next )
2749 importance[split->var_idx] += split->quality;
2752 for( parent = node->parent; parent && parent->right == node;
2753 node = parent, parent = parent->parent )
2759 node = parent->right;
2762 cvNormalize( var_importance, var_importance, 1., 0, CV_L1 );
2765 return var_importance;
2769 void CvDTree::write_train_data_params( CvFileStorage* fs )
2771 CV_FUNCNAME( "CvDTree::write_train_data_params" );
2775 int vi, vcount = data->var_count;
2777 cvWriteInt( fs, "is_classifier", data->is_classifier ? 1 : 0 );
2778 cvWriteInt( fs, "var_all", data->var_all );
2779 cvWriteInt( fs, "var_count", data->var_count );
2780 cvWriteInt( fs, "ord_var_count", data->ord_var_count );
2781 cvWriteInt( fs, "cat_var_count", data->cat_var_count );
2783 cvStartWriteStruct( fs, "training_params", CV_NODE_MAP );
2784 cvWriteInt( fs, "use_surrogates", data->params.use_surrogates ? 1 : 0 );
2786 if( data->is_classifier )
2788 cvWriteInt( fs, "max_categories", data->params.max_categories );
2792 cvWriteReal( fs, "regression_accuracy", data->params.regression_accuracy );
2795 cvWriteInt( fs, "max_depth", data->params.max_depth );
2796 cvWriteInt( fs, "min_sample_count", data->params.min_sample_count );
2797 cvWriteInt( fs, "cross_validation_folds", data->params.cv_folds );
2799 if( data->params.cv_folds > 1 )
2801 cvWriteInt( fs, "use_1se_rule", data->params.use_1se_rule ? 1 : 0 );
2802 cvWriteInt( fs, "truncate_pruned_tree", data->params.truncate_pruned_tree ? 1 : 0 );
2806 cvWrite( fs, "priors", data->priors );
2808 cvEndWriteStruct( fs );
2811 cvWrite( fs, "var_idx", data->var_idx );
2813 cvStartWriteStruct( fs, "var_type", CV_NODE_SEQ+CV_NODE_FLOW );
2815 for( vi = 0; vi < vcount; vi++ )
2816 cvWriteInt( fs, 0, data->var_type->data.i[vi] >= 0 );
2818 cvEndWriteStruct( fs );
2820 if( data->cat_count && (data->cat_var_count > 0 || data->is_classifier) )
2822 CV_ASSERT( data->cat_count != 0 );
2823 cvWrite( fs, "cat_count", data->cat_count );
2824 cvWrite( fs, "cat_map", data->cat_map );
2831 void CvDTree::write_split( CvFileStorage* fs, CvDTreeSplit* split )
2835 cvStartWriteStruct( fs, 0, CV_NODE_MAP + CV_NODE_FLOW );
2836 cvWriteInt( fs, "var", split->var_idx );
2837 cvWriteReal( fs, "quality", split->quality );
2839 ci = data->get_var_type(split->var_idx);
2840 if( ci >= 0 ) // split on a categorical var
2842 int i, n = data->cat_count->data.i[ci], to_right = 0, default_dir;
2843 for( i = 0; i < n; i++ )
2844 to_right += DTREE_CAT_DIR(i,split->subset) > 0;
2846 // ad-hoc rule when to use inverse categorical split notation
2847 // to achieve more compact and clear representation
2848 default_dir = to_right <= 1 || to_right <= MIN(3, n/2) || to_right <= n/3 ? -1 : 1;
2850 cvStartWriteStruct( fs, default_dir*(split->inversed ? -1 : 1) > 0 ?
2851 "in" : "not_in", CV_NODE_SEQ+CV_NODE_FLOW );
2852 for( i = 0; i < n; i++ )
2854 int dir = DTREE_CAT_DIR(i,split->subset);
2855 if( dir*default_dir < 0 )
2856 cvWriteInt( fs, 0, i );
2858 cvEndWriteStruct( fs );
2861 cvWriteReal( fs, !split->inversed ? "le" : "gt", split->ord.c );
2863 cvEndWriteStruct( fs );
2867 void CvDTree::write_node( CvFileStorage* fs, CvDTreeNode* node )
2869 CvDTreeSplit* split;
2871 cvStartWriteStruct( fs, 0, CV_NODE_MAP );
2873 cvWriteInt( fs, "depth", node->depth );
2874 cvWriteInt( fs, "sample_count", node->sample_count );
2875 cvWriteReal( fs, "value", node->value );
2877 if( data->is_classifier )
2878 cvWriteInt( fs, "norm_class_idx", node->class_idx );
2880 cvWriteInt( fs, "Tn", node->Tn );
2881 cvWriteInt( fs, "complexity", node->complexity );
2882 cvWriteReal( fs, "alpha", node->alpha );
2883 cvWriteReal( fs, "node_risk", node->node_risk );
2884 cvWriteReal( fs, "tree_risk", node->tree_risk );
2885 cvWriteReal( fs, "tree_error", node->tree_error );
2889 cvStartWriteStruct( fs, "splits", CV_NODE_SEQ );
2891 for( split = node->split; split != 0; split = split->next )
2892 write_split( fs, split );
2894 cvEndWriteStruct( fs );
2897 cvEndWriteStruct( fs );
2901 void CvDTree::write_tree_nodes( CvFileStorage* fs )
2903 //CV_FUNCNAME( "CvDTree::write_tree_nodes" );
2907 CvDTreeNode* node = root;
2909 // traverse the tree and save all the nodes in depth-first order
2912 CvDTreeNode* parent;
2915 write_node( fs, node );
2921 for( parent = node->parent; parent && parent->right == node;
2922 node = parent, parent = parent->parent )
2928 node = parent->right;
2935 void CvDTree::write( CvFileStorage* fs, const char* name )
2937 //CV_FUNCNAME( "CvDTree::write" );
2941 cvStartWriteStruct( fs, name, CV_NODE_MAP, CV_TYPE_NAME_ML_TREE );
2943 write_train_data_params( fs );
2945 cvWriteInt( fs, "best_tree_idx", pruned_tree_idx );
2946 get_var_importance();
2947 cvWrite( fs, "var_importance", var_importance );
2949 cvStartWriteStruct( fs, "nodes", CV_NODE_SEQ );
2950 write_tree_nodes( fs );
2951 cvEndWriteStruct( fs );
2953 cvEndWriteStruct( fs );
2959 void CvDTree::read_train_data_params( CvFileStorage* fs, CvFileNode* node )
2961 CV_FUNCNAME( "CvDTree::read_train_data_params" );
2965 CvDTreeParams params;
2966 CvFileNode *tparams_node, *vartype_node;
2968 int is_classifier, vi, cat_var_count, ord_var_count;
2969 int max_split_size, tree_block_size;
2971 data = new CvDTreeTrainData;
2973 is_classifier = (cvReadIntByName( fs, node, "is_classifier" ) != 0);
2974 data->is_classifier = (is_classifier != 0);
2975 data->var_all = cvReadIntByName( fs, node, "var_all" );
2976 data->var_count = cvReadIntByName( fs, node, "var_count", data->var_all );
2977 data->cat_var_count = cvReadIntByName( fs, node, "cat_var_count" );
2978 data->ord_var_count = cvReadIntByName( fs, node, "ord_var_count" );
2980 tparams_node = cvGetFileNodeByName( fs, node, "training_params" );
2982 if( tparams_node ) // training parameters are not necessary
2984 data->params.use_surrogates = cvReadIntByName( fs, tparams_node, "use_surrogates", 1 ) != 0;
2988 data->params.max_categories = cvReadIntByName( fs, tparams_node, "max_categories" );
2992 data->params.regression_accuracy =
2993 (float)cvReadRealByName( fs, tparams_node, "regression_accuracy" );
2996 data->params.max_depth = cvReadIntByName( fs, tparams_node, "max_depth" );
2997 data->params.min_sample_count = cvReadIntByName( fs, tparams_node, "min_sample_count" );
2998 data->params.cv_folds = cvReadIntByName( fs, tparams_node, "cross_validation_folds" );
3000 if( data->params.cv_folds > 1 )
3002 data->params.use_1se_rule = cvReadIntByName( fs, tparams_node, "use_1se_rule" ) != 0;
3003 data->params.truncate_pruned_tree =
3004 cvReadIntByName( fs, tparams_node, "truncate_pruned_tree" ) != 0;
3007 data->priors = (CvMat*)cvReadByName( fs, tparams_node, "priors" );
3008 if( data->priors && !CV_IS_MAT(data->priors) )
3009 CV_ERROR( CV_StsParseError, "priors must stored as a matrix" );
3012 CV_CALL( data->var_idx = (CvMat*)cvReadByName( fs, node, "var_idx" ));
3015 if( !CV_IS_MAT(data->var_idx) ||
3016 data->var_idx->cols != 1 && data->var_idx->rows != 1 ||
3017 data->var_idx->cols + data->var_idx->rows - 1 != data->var_count ||
3018 CV_MAT_TYPE(data->var_idx->type) != CV_32SC1 )
3019 CV_ERROR( CV_StsParseError,
3020 "var_idx (if exist) must be valid 1d integer vector containing <var_count> elements" );
3022 for( vi = 0; vi < data->var_count; vi++ )
3023 if( (unsigned)data->var_idx->data.i[vi] >= (unsigned)data->var_all )
3024 CV_ERROR( CV_StsOutOfRange, "some of var_idx elements are out of range" );
3027 ////// read var type
3028 CV_CALL( data->var_type = cvCreateMat( 1, data->var_count + 2, CV_32SC1 ));
3030 vartype_node = cvGetFileNodeByName( fs, node, "var_type" );
3031 if( !vartype_node || CV_NODE_TYPE(vartype_node->tag) != CV_NODE_SEQ ||
3032 vartype_node->data.seq->total != data->var_count )
3033 CV_ERROR( CV_StsParseError, "var_type must exist and be a sequence of 0's and 1's" );
3035 cvStartReadSeq( vartype_node->data.seq, &reader );
3039 for( vi = 0; vi < data->var_count; vi++ )
3041 CvFileNode* n = (CvFileNode*)reader.ptr;
3042 if( CV_NODE_TYPE(n->tag) != CV_NODE_INT || (n->data.i & ~1) )
3043 CV_ERROR( CV_StsParseError, "var_type must exist and be a sequence of 0's and 1's" );
3044 data->var_type->data.i[vi] = n->data.i ? cat_var_count++ : ord_var_count--;
3045 CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
3048 ord_var_count = ~ord_var_count;
3049 if( cat_var_count != data->cat_var_count || ord_var_count != data->ord_var_count )
3050 CV_ERROR( CV_StsParseError, "var_type is inconsistent with cat_var_count and ord_var_count" );
3053 if( data->cat_var_count > 0 || is_classifier )
3055 int ccount, max_c_count = 0, total_c_count = 0;
3056 CV_CALL( data->cat_count = (CvMat*)cvReadByName( fs, node, "cat_count" ));
3057 CV_CALL( data->cat_map = (CvMat*)cvReadByName( fs, node, "cat_map" ));
3059 if( !CV_IS_MAT(data->cat_count) || !CV_IS_MAT(data->cat_map) ||
3060 data->cat_count->cols != 1 && data->cat_count->rows != 1 ||
3061 CV_MAT_TYPE(data->cat_count->type) != CV_32SC1 ||
3062 data->cat_count->cols + data->cat_count->rows - 1 != cat_var_count + is_classifier ||
3063 data->cat_map->cols != 1 && data->cat_map->rows != 1 ||
3064 CV_MAT_TYPE(data->cat_map->type) != CV_32SC1 )
3065 CV_ERROR( CV_StsParseError,
3066 "Both cat_count and cat_map must exist and be valid 1d integer vectors of an appropriate size" );
3068 ccount = cat_var_count + is_classifier;
3070 CV_CALL( data->cat_ofs = cvCreateMat( 1, ccount + 1, CV_32SC1 ));
3071 data->cat_ofs->data.i[0] = 0;
3073 for( vi = 0; vi < ccount; vi++ )
3075 int val = data->cat_count->data.i[vi];
3077 CV_ERROR( CV_StsOutOfRange, "some of cat_count elements are out of range" );
3078 max_c_count = MAX( max_c_count, val );
3079 data->cat_ofs->data.i[vi+1] = total_c_count += val;
3082 if( data->cat_map->cols + data->cat_map->rows - 1 != total_c_count )
3083 CV_ERROR( CV_StsBadSize,
3084 "cat_map vector length is not equal to the total number of categories in all categorical vars" );
3086 data->max_c_count = max_c_count;
3089 max_split_size = cvAlign(sizeof(CvDTreeSplit) +
3090 (MAX(0,data->max_c_count - 33)/32)*sizeof(int),sizeof(void*));
3092 tree_block_size = MAX((int)sizeof(CvDTreeNode)*8, max_split_size);
3093 tree_block_size = MAX(tree_block_size + block_size_delta, min_block_size);
3094 CV_CALL( data->tree_storage = cvCreateMemStorage( tree_block_size ));
3095 CV_CALL( data->node_heap = cvCreateSet( 0, sizeof(data->node_heap[0]),
3096 sizeof(CvDTreeNode), data->tree_storage ));
3097 CV_CALL( data->split_heap = cvCreateSet( 0, sizeof(data->split_heap[0]),
3098 max_split_size, data->tree_storage ));
3104 CvDTreeSplit* CvDTree::read_split( CvFileStorage* fs, CvFileNode* fnode )
3106 CvDTreeSplit* split = 0;
3108 CV_FUNCNAME( "CvDTree::read_split" );
3114 if( !fnode || CV_NODE_TYPE(fnode->tag) != CV_NODE_MAP )
3115 CV_ERROR( CV_StsParseError, "some of the splits are not stored properly" );
3117 vi = cvReadIntByName( fs, fnode, "var", -1 );
3118 if( (unsigned)vi >= (unsigned)data->var_count )
3119 CV_ERROR( CV_StsOutOfRange, "Split variable index is out of range" );
3121 ci = data->get_var_type(vi);
3122 if( ci >= 0 ) // split on categorical var
3124 int i, n = data->cat_count->data.i[ci], inversed = 0;
3127 split = data->new_split_cat( vi, 0 );
3128 inseq = cvGetFileNodeByName( fs, fnode, "in" );
3131 inseq = cvGetFileNodeByName( fs, fnode, "not_in" );
3134 if( !inseq || CV_NODE_TYPE(inseq->tag) != CV_NODE_SEQ )
3135 CV_ERROR( CV_StsParseError,
3136 "Either 'in' or 'not_in' tags should be inside a categorical split data" );
3138 cvStartReadSeq( inseq->data.seq, &reader );
3140 for( i = 0; i < reader.seq->total; i++ )
3142 CvFileNode* inode = (CvFileNode*)reader.ptr;
3143 int val = inode->data.i;
3144 if( CV_NODE_TYPE(inode->tag) != CV_NODE_INT || (unsigned)val >= (unsigned)n )
3145 CV_ERROR( CV_StsOutOfRange, "some of in/not_in elements are out of range" );
3147 split->subset[val >> 5] |= 1 << (val & 31);
3148 CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
3151 // for categorical splits we do not use inversed splits,
3152 // instead we inverse the variable set in the split
3154 for( i = 0; i < (n + 31) >> 5; i++ )
3155 split->subset[i] ^= -1;
3159 CvFileNode* cmp_node;
3160 split = data->new_split_ord( vi, 0, 0, 0, 0 );
3162 cmp_node = cvGetFileNodeByName( fs, fnode, "le" );
3165 cmp_node = cvGetFileNodeByName( fs, fnode, "gt" );
3166 split->inversed = 1;
3169 split->ord.c = (float)cvReadReal( cmp_node );
3172 split->quality = (float)cvReadRealByName( fs, fnode, "quality" );
3180 CvDTreeNode* CvDTree::read_node( CvFileStorage* fs, CvFileNode* fnode, CvDTreeNode* parent )
3182 CvDTreeNode* node = 0;
3184 CV_FUNCNAME( "CvDTree::read_node" );
3191 if( !fnode || CV_NODE_TYPE(fnode->tag) != CV_NODE_MAP )
3192 CV_ERROR( CV_StsParseError, "some of the tree elements are not stored properly" );
3194 CV_CALL( node = data->new_node( parent, 0, 0, 0 ));
3195 depth = cvReadIntByName( fs, fnode, "depth", -1 );
3196 if( depth != node->depth )
3197 CV_ERROR( CV_StsParseError, "incorrect node depth" );
3199 node->sample_count = cvReadIntByName( fs, fnode, "sample_count" );
3200 node->value = cvReadRealByName( fs, fnode, "value" );
3201 if( data->is_classifier )
3202 node->class_idx = cvReadIntByName( fs, fnode, "norm_class_idx" );
3204 node->Tn = cvReadIntByName( fs, fnode, "Tn" );
3205 node->complexity = cvReadIntByName( fs, fnode, "complexity" );
3206 node->alpha = cvReadRealByName( fs, fnode, "alpha" );
3207 node->node_risk = cvReadRealByName( fs, fnode, "node_risk" );
3208 node->tree_risk = cvReadRealByName( fs, fnode, "tree_risk" );
3209 node->tree_error = cvReadRealByName( fs, fnode, "tree_error" );
3211 splits = cvGetFileNodeByName( fs, fnode, "splits" );
3215 CvDTreeSplit* last_split = 0;
3217 if( CV_NODE_TYPE(splits->tag) != CV_NODE_SEQ )
3218 CV_ERROR( CV_StsParseError, "splits tag must stored as a sequence" );
3220 cvStartReadSeq( splits->data.seq, &reader );
3221 for( i = 0; i < reader.seq->total; i++ )
3223 CvDTreeSplit* split;
3224 CV_CALL( split = read_split( fs, (CvFileNode*)reader.ptr ));
3226 node->split = last_split = split;
3228 last_split = last_split->next = split;
3230 CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
3240 void CvDTree::read_tree_nodes( CvFileStorage* fs, CvFileNode* fnode )
3242 CV_FUNCNAME( "CvDTree::read_tree_nodes" );
3248 CvDTreeNode* parent = &_root;
3250 parent->left = parent->right = parent->parent = 0;
3252 cvStartReadSeq( fnode->data.seq, &reader );
3254 for( i = 0; i < reader.seq->total; i++ )
3258 CV_CALL( node = read_node( fs, (CvFileNode*)reader.ptr, parent != &_root ? parent : 0 ));
3260 parent->left = node;
3262 parent->right = node;
3267 while( parent && parent->right )
3268 parent = parent->parent;
3271 CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
3280 void CvDTree::read( CvFileStorage* fs, CvFileNode* fnode )
3282 CV_FUNCNAME( "CvDTree::read" );
3286 CvFileNode* tree_nodes;
3289 read_train_data_params( fs, fnode );
3291 tree_nodes = cvGetFileNodeByName( fs, fnode, "nodes" );
3292 if( !tree_nodes || CV_NODE_TYPE(tree_nodes->tag) != CV_NODE_SEQ )
3293 CV_ERROR( CV_StsParseError, "nodes tag is missing" );
3295 pruned_tree_idx = cvReadIntByName( fs, fnode, "best_tree_idx", -1 );
3297 read_tree_nodes( fs, tree_nodes );
3298 get_var_importance(); // recompute variable importance