/*M///////////////////////////////////////////////////////////////////////////////////////
//
// IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING.
//
// By downloading, copying, installing or using the software you agree to this license.
// If you do not agree to this license, do not download, install,
// copy or use the software.
//
//
// Intel License Agreement
//
// Copyright (C) 2000, Intel Corporation, all rights reserved.
// Third party copyrights are property of their respective owners.
//
// Redistribution and use in source and binary forms, with or without modification,
// are permitted provided that the following conditions are met:
//
// * Redistribution's of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
//
// * Redistribution's in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
//
// * The name of Intel Corporation may not be used to endorse or promote products
// derived from this software without specific prior written permission.
//
// This software is provided by the copyright holders and contributors "as is" and
// any express or implied warranties, including, but not limited to, the implied
// warranties of merchantability and fitness for a particular purpose are disclaimed.
// In no event shall the Intel Corporation or contributors be liable for any direct,
// indirect, incidental, special, exemplary, or consequential damages
// (including, but not limited to, procurement of substitute goods or services;
// loss of use, data, or profits; or business interruption) however caused
// and on any theory of liability, whether in contract, strict liability,
// or tort (including negligence or otherwise) arising in any way out of
// the use of this software, even if advised of the possibility of such damage.
//
//M*/
#include "_ml.h"
static const float ord_nan = FLT_MAX*0.5f;
static const int min_block_size = 1 << 16;
static const int block_size_delta = 1 << 10;
CvDTreeTrainData::CvDTreeTrainData()
{
var_idx = var_type = cat_count = cat_ofs = cat_map =
priors = priors_mult = counts = buf = direction = split_buf = 0;
tree_storage = temp_storage = 0;
clear();
}
CvDTreeTrainData::CvDTreeTrainData( const CvMat* _train_data, int _tflag,
const CvMat* _responses, const CvMat* _var_idx,
const CvMat* _sample_idx, const CvMat* _var_type,
const CvMat* _missing_mask, const CvDTreeParams& _params,
bool _shared, bool _add_labels )
{
var_idx = var_type = cat_count = cat_ofs = cat_map =
priors = priors_mult = counts = buf = direction = split_buf = 0;
tree_storage = temp_storage = 0;
set_data( _train_data, _tflag, _responses, _var_idx, _sample_idx,
_var_type, _missing_mask, _params, _shared, _add_labels );
}
CvDTreeTrainData::~CvDTreeTrainData()
{
clear();
}
bool CvDTreeTrainData::set_params( const CvDTreeParams& _params )
{
bool ok = false;
CV_FUNCNAME( "CvDTreeTrainData::set_params" );
__BEGIN__;
// set parameters
params = _params;
if( params.max_categories < 2 )
CV_ERROR( CV_StsOutOfRange, "params.max_categories should be >= 2" );
params.max_categories = MIN( params.max_categories, 15 );
if( params.max_depth < 0 )
CV_ERROR( CV_StsOutOfRange, "params.max_depth should be >= 0" );
params.max_depth = MIN( params.max_depth, 25 );
params.min_sample_count = MAX(params.min_sample_count,1);
if( params.cv_folds < 0 )
CV_ERROR( CV_StsOutOfRange,
"params.cv_folds should be =0 (the tree is not pruned) "
"or n>0 (tree is pruned using n-fold cross-validation)" );
if( params.cv_folds == 1 )
params.cv_folds = 0;
if( params.regression_accuracy < 0 )
CV_ERROR( CV_StsOutOfRange, "params.regression_accuracy should be >= 0" );
ok = true;
__END__;
return ok;
}
#define CV_CMP_NUM_PTR(a,b) (*(a) < *(b))
static CV_IMPLEMENT_QSORT_EX( icvSortIntPtr, int*, CV_CMP_NUM_PTR, int )
static CV_IMPLEMENT_QSORT_EX( icvSortDblPtr, double*, CV_CMP_NUM_PTR, int )
#define CV_CMP_PAIRS(a,b) ((a).val < (b).val)
static CV_IMPLEMENT_QSORT_EX( icvSortPairs, CvPair32s32f, CV_CMP_PAIRS, int )
void CvDTreeTrainData::set_data( const CvMat* _train_data, int _tflag,
const CvMat* _responses, const CvMat* _var_idx, const CvMat* _sample_idx,
const CvMat* _var_type, const CvMat* _missing_mask, const CvDTreeParams& _params,
bool _shared, bool _add_labels, bool _update_data )
{
CvMat* sample_idx = 0;
CvMat* var_type0 = 0;
CvMat* tmp_map = 0;
int** int_ptr = 0;
CvDTreeTrainData* data = 0;
CV_FUNCNAME( "CvDTreeTrainData::set_data" );
__BEGIN__;
int sample_all = 0, r_type = 0, cv_n;
int total_c_count = 0;
int tree_block_size, temp_block_size, max_split_size, nv_size, cv_size = 0;
int ds_step, dv_step, ms_step = 0, mv_step = 0; // {data|mask}{sample|var}_step
int vi, i;
char err[100];
const int *sidx = 0, *vidx = 0;
if( _update_data && data_root )
{
data = new CvDTreeTrainData( _train_data, _tflag, _responses, _var_idx,
_sample_idx, _var_type, _missing_mask, _params, _shared, _add_labels );
// compare new and old train data
if( !(data->var_count == var_count &&
cvNorm( data->var_type, var_type, CV_C ) < FLT_EPSILON &&
cvNorm( data->cat_count, cat_count, CV_C ) < FLT_EPSILON &&
cvNorm( data->cat_map, cat_map, CV_C ) < FLT_EPSILON) )
CV_ERROR( CV_StsBadArg,
"The new training data must have the same types and the input and output variables "
"and the same categories for categorical variables" );
cvReleaseMat( &priors );
cvReleaseMat( &priors_mult );
cvReleaseMat( &buf );
cvReleaseMat( &direction );
cvReleaseMat( &split_buf );
cvReleaseMemStorage( &temp_storage );
priors = data->priors; data->priors = 0;
priors_mult = data->priors_mult; data->priors_mult = 0;
buf = data->buf; data->buf = 0;
buf_count = data->buf_count; buf_size = data->buf_size;
sample_count = data->sample_count;
direction = data->direction; data->direction = 0;
split_buf = data->split_buf; data->split_buf = 0;
temp_storage = data->temp_storage; data->temp_storage = 0;
nv_heap = data->nv_heap; cv_heap = data->cv_heap;
data_root = new_node( 0, sample_count, 0, 0 );
EXIT;
}
clear();
var_all = 0;
rng = cvRNG(-1);
CV_CALL( set_params( _params ));
// check parameter types and sizes
CV_CALL( cvCheckTrainData( _train_data, _tflag, _missing_mask, &var_all, &sample_all ));
if( _tflag == CV_ROW_SAMPLE )
{
ds_step = _train_data->step/CV_ELEM_SIZE(_train_data->type);
dv_step = 1;
if( _missing_mask )
ms_step = _missing_mask->step, mv_step = 1;
}
else
{
dv_step = _train_data->step/CV_ELEM_SIZE(_train_data->type);
ds_step = 1;
if( _missing_mask )
mv_step = _missing_mask->step, ms_step = 1;
}
sample_count = sample_all;
var_count = var_all;
if( _sample_idx )
{
CV_CALL( sample_idx = cvPreprocessIndexArray( _sample_idx, sample_all ));
sidx = sample_idx->data.i;
sample_count = sample_idx->rows + sample_idx->cols - 1;
}
if( _var_idx )
{
CV_CALL( var_idx = cvPreprocessIndexArray( _var_idx, var_all ));
vidx = var_idx->data.i;
var_count = var_idx->rows + var_idx->cols - 1;
}
if( !CV_IS_MAT(_responses) ||
(CV_MAT_TYPE(_responses->type) != CV_32SC1 &&
CV_MAT_TYPE(_responses->type) != CV_32FC1) ||
_responses->rows != 1 && _responses->cols != 1 ||
_responses->rows + _responses->cols - 1 != sample_all )
CV_ERROR( CV_StsBadArg, "The array of _responses must be an integer or "
"floating-point vector containing as many elements as "
"the total number of samples in the training data matrix" );
CV_CALL( var_type0 = cvPreprocessVarType( _var_type, var_idx, var_all, &r_type ));
CV_CALL( var_type = cvCreateMat( 1, var_count+2, CV_32SC1 ));
cat_var_count = 0;
ord_var_count = -1;
is_classifier = r_type == CV_VAR_CATEGORICAL;
// step 0. calc the number of categorical vars
for( vi = 0; vi < var_count; vi++ )
{
var_type->data.i[vi] = var_type0->data.ptr[vi] == CV_VAR_CATEGORICAL ?
cat_var_count++ : ord_var_count--;
}
ord_var_count = ~ord_var_count;
cv_n = params.cv_folds;
// set the two last elements of var_type array to be able
// to locate responses and cross-validation labels using
// the corresponding get_* functions.
var_type->data.i[var_count] = cat_var_count;
var_type->data.i[var_count+1] = cat_var_count+1;
// in case of single ordered predictor we need dummy cv_labels
// for safe split_node_data() operation
have_labels = cv_n > 0 || ord_var_count == 1 && cat_var_count == 0 || _add_labels;
buf_size = (ord_var_count + get_work_var_count())*sample_count + 2;
shared = _shared;
buf_count = shared ? 3 : 2;
CV_CALL( buf = cvCreateMat( buf_count, buf_size, CV_32SC1 ));
CV_CALL( cat_count = cvCreateMat( 1, cat_var_count+1, CV_32SC1 ));
CV_CALL( cat_ofs = cvCreateMat( 1, cat_count->cols+1, CV_32SC1 ));
CV_CALL( cat_map = cvCreateMat( 1, cat_count->cols*10 + 128, CV_32SC1 ));
// now calculate the maximum size of split,
// create memory storage that will keep nodes and splits of the decision tree
// allocate root node and the buffer for the whole training data
max_split_size = cvAlign(sizeof(CvDTreeSplit) +
(MAX(0,sample_count - 33)/32)*sizeof(int),sizeof(void*));
tree_block_size = MAX((int)sizeof(CvDTreeNode)*8, max_split_size);
tree_block_size = MAX(tree_block_size + block_size_delta, min_block_size);
CV_CALL( tree_storage = cvCreateMemStorage( tree_block_size ));
CV_CALL( node_heap = cvCreateSet( 0, sizeof(*node_heap), sizeof(CvDTreeNode), tree_storage ));
nv_size = var_count*sizeof(int);
nv_size = MAX( nv_size, (int)sizeof(CvSetElem) );
temp_block_size = nv_size;
if( cv_n )
{
if( sample_count < cv_n*MAX(params.min_sample_count,10) )
CV_ERROR( CV_StsOutOfRange,
"The many folds in cross-validation for such a small dataset" );
cv_size = cvAlign( cv_n*(sizeof(int) + sizeof(double)*2), sizeof(double) );
temp_block_size = MAX(temp_block_size, cv_size);
}
temp_block_size = MAX( temp_block_size + block_size_delta, min_block_size );
CV_CALL( temp_storage = cvCreateMemStorage( temp_block_size ));
CV_CALL( nv_heap = cvCreateSet( 0, sizeof(*nv_heap), nv_size, temp_storage ));
if( cv_size )
CV_CALL( cv_heap = cvCreateSet( 0, sizeof(*cv_heap), cv_size, temp_storage ));
CV_CALL( data_root = new_node( 0, sample_count, 0, 0 ));
CV_CALL( int_ptr = (int**)cvAlloc( sample_count*sizeof(int_ptr[0]) ));
max_c_count = 1;
// transform the training data to convenient representation
for( vi = 0; vi <= var_count; vi++ )
{
int ci;
const uchar* mask = 0;
int m_step = 0, step;
const int* idata = 0;
const float* fdata = 0;
int num_valid = 0;
if( vi < var_count ) // analyze i-th input variable
{
int vi0 = vidx ? vidx[vi] : vi;
ci = get_var_type(vi);
step = ds_step; m_step = ms_step;
if( CV_MAT_TYPE(_train_data->type) == CV_32SC1 )
idata = _train_data->data.i + vi0*dv_step;
else
fdata = _train_data->data.fl + vi0*dv_step;
if( _missing_mask )
mask = _missing_mask->data.ptr + vi0*mv_step;
}
else // analyze _responses
{
ci = cat_var_count;
step = CV_IS_MAT_CONT(_responses->type) ?
1 : _responses->step / CV_ELEM_SIZE(_responses->type);
if( CV_MAT_TYPE(_responses->type) == CV_32SC1 )
idata = _responses->data.i;
else
fdata = _responses->data.fl;
}
if( vi < var_count && ci >= 0 ||
vi == var_count && is_classifier ) // process categorical variable or response
{
int c_count, prev_label;
int* c_map, *dst = get_cat_var_data( data_root, vi );
// copy data
for( i = 0; i < sample_count; i++ )
{
int val = INT_MAX, si = sidx ? sidx[i] : i;
if( !mask || !mask[si*m_step] )
{
if( idata )
val = idata[si*step];
else
{
float t = fdata[si*step];
val = cvRound(t);
if( val != t )
{
sprintf( err, "%d-th value of %d-th (categorical) "
"variable is not an integer", i, vi );
CV_ERROR( CV_StsBadArg, err );
}
}
if( val == INT_MAX )
{
sprintf( err, "%d-th value of %d-th (categorical) "
"variable is too large", i, vi );
CV_ERROR( CV_StsBadArg, err );
}
num_valid++;
}
dst[i] = val;
int_ptr[i] = dst + i;
}
// sort all the values, including the missing measurements
// that should all move to the end
icvSortIntPtr( int_ptr, sample_count, 0 );
//qsort( int_ptr, sample_count, sizeof(int_ptr[0]), icvCmpIntPtr );
c_count = num_valid > 0;
// count the categories
for( i = 1; i < num_valid; i++ )
c_count += *int_ptr[i] != *int_ptr[i-1];
if( vi > 0 )
max_c_count = MAX( max_c_count, c_count );
cat_count->data.i[ci] = c_count;
cat_ofs->data.i[ci] = total_c_count;
// resize cat_map, if need
if( cat_map->cols < total_c_count + c_count )
{
tmp_map = cat_map;
CV_CALL( cat_map = cvCreateMat( 1,
MAX(cat_map->cols*3/2,total_c_count+c_count), CV_32SC1 ));
for( i = 0; i < total_c_count; i++ )
cat_map->data.i[i] = tmp_map->data.i[i];
cvReleaseMat( &tmp_map );
}
c_map = cat_map->data.i + total_c_count;
total_c_count += c_count;
// compact the class indices and build the map
prev_label = ~*int_ptr[0];
c_count = -1;
for( i = 0; i < num_valid; i++ )
{
int cur_label = *int_ptr[i];
if( cur_label != prev_label )
c_map[++c_count] = prev_label = cur_label;
*int_ptr[i] = c_count;
}
// replace labels for missing values with -1
for( ; i < sample_count; i++ )
*int_ptr[i] = -1;
}
else if( ci < 0 ) // process ordered variable
{
CvPair32s32f* dst = get_ord_var_data( data_root, vi );
for( i = 0; i < sample_count; i++ )
{
float val = ord_nan;
int si = sidx ? sidx[i] : i;
if( !mask || !mask[si*m_step] )
{
if( idata )
val = (float)idata[si*step];
else
val = fdata[si*step];
if( fabs(val) >= ord_nan )
{
sprintf( err, "%d-th value of %d-th (ordered) "
"variable (=%g) is too large", i, vi, val );
CV_ERROR( CV_StsBadArg, err );
}
num_valid++;
}
dst[i].i = i;
dst[i].val = val;
}
icvSortPairs( dst, sample_count, 0 );
}
else // special case: process ordered response,
// it will be stored similarly to categorical vars (i.e. no pairs)
{
float* dst = get_ord_responses( data_root );
for( i = 0; i < sample_count; i++ )
{
float val = ord_nan;
int si = sidx ? sidx[i] : i;
if( idata )
val = (float)idata[si*step];
else
val = fdata[si*step];
if( fabs(val) >= ord_nan )
{
sprintf( err, "%d-th value of %d-th (ordered) "
"variable (=%g) is out of range", i, vi, val );
CV_ERROR( CV_StsBadArg, err );
}
dst[i] = val;
}
cat_count->data.i[cat_var_count] = 0;
cat_ofs->data.i[cat_var_count] = total_c_count;
num_valid = sample_count;
}
if( vi < var_count )
data_root->set_num_valid(vi, num_valid);
}
if( cv_n )
{
int* dst = get_labels(data_root);
CvRNG* r = &rng;
for( i = vi = 0; i < sample_count; i++ )
{
dst[i] = vi++;
vi &= vi < cv_n ? -1 : 0;
}
for( i = 0; i < sample_count; i++ )
{
int a = cvRandInt(r) % sample_count;
int b = cvRandInt(r) % sample_count;
CV_SWAP( dst[a], dst[b], vi );
}
}
cat_map->cols = MAX( total_c_count, 1 );
max_split_size = cvAlign(sizeof(CvDTreeSplit) +
(MAX(0,max_c_count - 33)/32)*sizeof(int),sizeof(void*));
CV_CALL( split_heap = cvCreateSet( 0, sizeof(*split_heap), max_split_size, tree_storage ));
have_priors = is_classifier && params.priors;
if( is_classifier )
{
int m = get_num_classes();
double sum = 0;
CV_CALL( priors = cvCreateMat( 1, m, CV_64F ));
for( i = 0; i < m; i++ )
{
double val = have_priors ? params.priors[i] : 1.;
if( val <= 0 )
CV_ERROR( CV_StsOutOfRange, "Every class weight should be positive" );
priors->data.db[i] = val;
sum += val;
}
// normalize weights
if( have_priors )
cvScale( priors, priors, 1./sum );
CV_CALL( priors_mult = cvCloneMat( priors ));
CV_CALL( counts = cvCreateMat( 1, m, CV_32SC1 ));
}
CV_CALL( direction = cvCreateMat( 1, sample_count, CV_8UC1 ));
CV_CALL( split_buf = cvCreateMat( 1, sample_count, CV_32SC1 ));
__END__;
if( data )
delete data;
cvFree( &int_ptr );
cvReleaseMat( &sample_idx );
cvReleaseMat( &var_type0 );
cvReleaseMat( &tmp_map );
}
CvDTreeNode* CvDTreeTrainData::subsample_data( const CvMat* _subsample_idx )
{
CvDTreeNode* root = 0;
CvMat* isubsample_idx = 0;
CvMat* subsample_co = 0;
CV_FUNCNAME( "CvDTreeTrainData::subsample_data" );
__BEGIN__;
if( !data_root )
CV_ERROR( CV_StsError, "No training data has been set" );
if( _subsample_idx )
CV_CALL( isubsample_idx = cvPreprocessIndexArray( _subsample_idx, sample_count ));
if( !isubsample_idx )
{
// make a copy of the root node
CvDTreeNode temp;
int i;
root = new_node( 0, 1, 0, 0 );
temp = *root;
*root = *data_root;
root->num_valid = temp.num_valid;
if( root->num_valid )
{
for( i = 0; i < var_count; i++ )
root->num_valid[i] = data_root->num_valid[i];
}
root->cv_Tn = temp.cv_Tn;
root->cv_node_risk = temp.cv_node_risk;
root->cv_node_error = temp.cv_node_error;
}
else
{
int* sidx = isubsample_idx->data.i;
// co - array of count/offset pairs (to handle duplicated values in _subsample_idx)
int* co, cur_ofs = 0;
int vi, i, total = data_root->sample_count;
int count = isubsample_idx->rows + isubsample_idx->cols - 1;
int work_var_count = get_work_var_count();
root = new_node( 0, count, 1, 0 );
CV_CALL( subsample_co = cvCreateMat( 1, total*2, CV_32SC1 ));
cvZero( subsample_co );
co = subsample_co->data.i;
for( i = 0; i < count; i++ )
co[sidx[i]*2]++;
for( i = 0; i < total; i++ )
{
if( co[i*2] )
{
co[i*2+1] = cur_ofs;
cur_ofs += co[i*2];
}
else
co[i*2+1] = -1;
}
for( vi = 0; vi < work_var_count; vi++ )
{
int ci = get_var_type(vi);
if( ci >= 0 || vi >= var_count )
{
const int* src = get_cat_var_data( data_root, vi );
int* dst = get_cat_var_data( root, vi );
int num_valid = 0;
for( i = 0; i < count; i++ )
{
int val = src[sidx[i]];
dst[i] = val;
num_valid += val >= 0;
}
if( vi < var_count )
root->set_num_valid(vi, num_valid);
}
else
{
const CvPair32s32f* src = get_ord_var_data( data_root, vi );
CvPair32s32f* dst = get_ord_var_data( root, vi );
int j = 0, idx, count_i;
int num_valid = data_root->get_num_valid(vi);
for( i = 0; i < num_valid; i++ )
{
idx = src[i].i;
count_i = co[idx*2];
if( count_i )
{
float val = src[i].val;
for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
{
dst[j].val = val;
dst[j].i = cur_ofs;
}
}
}
root->set_num_valid(vi, j);
for( ; i < total; i++ )
{
idx = src[i].i;
count_i = co[idx*2];
if( count_i )
{
float val = src[i].val;
for( cur_ofs = co[idx*2+1]; count_i > 0; count_i--, j++, cur_ofs++ )
{
dst[j].val = val;
dst[j].i = cur_ofs;
}
}
}
}
}
}
__END__;
cvReleaseMat( &isubsample_idx );
cvReleaseMat( &subsample_co );
return root;
}
void CvDTreeTrainData::get_vectors( const CvMat* _subsample_idx,
float* values, uchar* missing,
float* responses, bool get_class_idx )
{
CvMat* subsample_idx = 0;
CvMat* subsample_co = 0;
CV_FUNCNAME( "CvDTreeTrainData::get_vectors" );
__BEGIN__;
int i, vi, total = sample_count, count = total, cur_ofs = 0;
int* sidx = 0;
int* co = 0;
if( _subsample_idx )
{
CV_CALL( subsample_idx = cvPreprocessIndexArray( _subsample_idx, sample_count ));
sidx = subsample_idx->data.i;
CV_CALL( subsample_co = cvCreateMat( 1, sample_count*2, CV_32SC1 ));
co = subsample_co->data.i;
cvZero( subsample_co );
count = subsample_idx->cols + subsample_idx->rows - 1;
for( i = 0; i < count; i++ )
co[sidx[i]*2]++;
for( i = 0; i < total; i++ )
{
int count_i = co[i*2];
if( count_i )
{
co[i*2+1] = cur_ofs*var_count;
cur_ofs += count_i;
}
}
}
if( missing )
memset( missing, 1, count*var_count );
for( vi = 0; vi < var_count; vi++ )
{
int ci = get_var_type(vi);
if( ci >= 0 ) // categorical
{
float* dst = values + vi;
uchar* m = missing ? missing + vi : 0;
const int* src = get_cat_var_data(data_root, vi);
for( i = 0; i < count; i++, dst += var_count )
{
int idx = sidx ? sidx[i] : i;
int val = src[idx];
*dst = (float)val;
if( m )
{
*m = val < 0;
m += var_count;
}
}
}
else // ordered
{
float* dst = values + vi;
uchar* m = missing ? missing + vi : 0;
const CvPair32s32f* src = get_ord_var_data(data_root, vi);
int count1 = data_root->get_num_valid(vi);
for( i = 0; i < count1; i++ )
{
int idx = src[i].i;
int count_i = 1;
if( co )
{
count_i = co[idx*2];
cur_ofs = co[idx*2+1];
}
else
cur_ofs = idx*var_count;
if( count_i )
{
float val = src[i].val;
for( ; count_i > 0; count_i--, cur_ofs += var_count )
{
dst[cur_ofs] = val;
if( m )
m[cur_ofs] = 0;
}
}
}
}
}
// copy responses
if( responses )
{
if( is_classifier )
{
const int* src = get_class_labels(data_root);
for( i = 0; i < count; i++ )
{
int idx = sidx ? sidx[i] : i;
int val = get_class_idx ? src[idx] :
cat_map->data.i[cat_ofs->data.i[cat_var_count]+src[idx]];
responses[i] = (float)val;
}
}
else
{
const float* src = get_ord_responses(data_root);
for( i = 0; i < count; i++ )
{
int idx = sidx ? sidx[i] : i;
responses[i] = src[idx];
}
}
}
__END__;
cvReleaseMat( &subsample_idx );
cvReleaseMat( &subsample_co );
}
CvDTreeNode* CvDTreeTrainData::new_node( CvDTreeNode* parent, int count,
int storage_idx, int offset )
{
CvDTreeNode* node = (CvDTreeNode*)cvSetNew( node_heap );
node->sample_count = count;
node->depth = parent ? parent->depth + 1 : 0;
node->parent = parent;
node->left = node->right = 0;
node->split = 0;
node->value = 0;
node->class_idx = 0;
node->maxlr = 0.;
node->buf_idx = storage_idx;
node->offset = offset;
if( nv_heap )
node->num_valid = (int*)cvSetNew( nv_heap );
else
node->num_valid = 0;
node->alpha = node->node_risk = node->tree_risk = node->tree_error = 0.;
node->complexity = 0;
if( params.cv_folds > 0 && cv_heap )
{
int cv_n = params.cv_folds;
node->Tn = INT_MAX;
node->cv_Tn = (int*)cvSetNew( cv_heap );
node->cv_node_risk = (double*)cvAlignPtr(node->cv_Tn + cv_n, sizeof(double));
node->cv_node_error = node->cv_node_risk + cv_n;
}
else
{
node->Tn = 0;
node->cv_Tn = 0;
node->cv_node_risk = 0;
node->cv_node_error = 0;
}
return node;
}
CvDTreeSplit* CvDTreeTrainData::new_split_ord( int vi, float cmp_val,
int split_point, int inversed, float quality )
{
CvDTreeSplit* split = (CvDTreeSplit*)cvSetNew( split_heap );
split->var_idx = vi;
split->ord.c = cmp_val;
split->ord.split_point = split_point;
split->inversed = inversed;
split->quality = quality;
split->next = 0;
return split;
}
CvDTreeSplit* CvDTreeTrainData::new_split_cat( int vi, float quality )
{
CvDTreeSplit* split = (CvDTreeSplit*)cvSetNew( split_heap );
int i, n = (max_c_count + 31)/32;
split->var_idx = vi;
split->inversed = 0;
split->quality = quality;
for( i = 0; i < n; i++ )
split->subset[i] = 0;
split->next = 0;
return split;
}
void CvDTreeTrainData::free_node( CvDTreeNode* node )
{
CvDTreeSplit* split = node->split;
free_node_data( node );
while( split )
{
CvDTreeSplit* next = split->next;
cvSetRemoveByPtr( split_heap, split );
split = next;
}
node->split = 0;
cvSetRemoveByPtr( node_heap, node );
}
void CvDTreeTrainData::free_node_data( CvDTreeNode* node )
{
if( node->num_valid )
{
cvSetRemoveByPtr( nv_heap, node->num_valid );
node->num_valid = 0;
}
// do not free cv_* fields, as all the cross-validation related data is released at once.
}
void CvDTreeTrainData::free_train_data()
{
cvReleaseMat( &counts );
cvReleaseMat( &buf );
cvReleaseMat( &direction );
cvReleaseMat( &split_buf );
cvReleaseMemStorage( &temp_storage );
cv_heap = nv_heap = 0;
}
void CvDTreeTrainData::clear()
{
free_train_data();
cvReleaseMemStorage( &tree_storage );
cvReleaseMat( &var_idx );
cvReleaseMat( &var_type );
cvReleaseMat( &cat_count );
cvReleaseMat( &cat_ofs );
cvReleaseMat( &cat_map );
cvReleaseMat( &priors );
cvReleaseMat( &priors_mult );
node_heap = split_heap = 0;
sample_count = var_all = var_count = max_c_count = ord_var_count = cat_var_count = 0;
have_labels = have_priors = is_classifier = false;
buf_count = buf_size = 0;
shared = false;
data_root = 0;
rng = cvRNG(-1);
}
int CvDTreeTrainData::get_num_classes() const
{
return is_classifier ? cat_count->data.i[cat_var_count] : 0;
}
int CvDTreeTrainData::get_var_type(int vi) const
{
return var_type->data.i[vi];
}
int CvDTreeTrainData::get_work_var_count() const
{
return var_count + 1 + (have_labels ? 1 : 0);
}
CvPair32s32f* CvDTreeTrainData::get_ord_var_data( CvDTreeNode* n, int vi )
{
int oi = ~get_var_type(vi);
assert( 0 <= oi && oi < ord_var_count );
return (CvPair32s32f*)(buf->data.i + n->buf_idx*buf->cols +
n->offset + oi*n->sample_count*2);
}
int* CvDTreeTrainData::get_class_labels( CvDTreeNode* n )
{
return get_cat_var_data( n, var_count );
}
float* CvDTreeTrainData::get_ord_responses( CvDTreeNode* n )
{
return (float*)get_cat_var_data( n, var_count );
}
int* CvDTreeTrainData::get_labels( CvDTreeNode* n )
{
return have_labels ? get_cat_var_data( n, var_count + 1 ) : 0;
}
int* CvDTreeTrainData::get_cat_var_data( CvDTreeNode* n, int vi )
{
int ci = get_var_type(vi);
assert( 0 <= ci && ci <= cat_var_count + 1 );
return buf->data.i + n->buf_idx*buf->cols + n->offset +
(ord_var_count*2 + ci)*n->sample_count;
}
int CvDTreeTrainData::get_child_buf_idx( CvDTreeNode* n )
{
int idx = n->buf_idx + 1;
if( idx >= buf_count )
idx = shared ? 1 : 0;
return idx;
}
void CvDTreeTrainData::write_params( CvFileStorage* fs )
{
CV_FUNCNAME( "CvDTreeTrainData::write_params" );
__BEGIN__;
int vi, vcount = var_count;
cvWriteInt( fs, "is_classifier", is_classifier ? 1 : 0 );
cvWriteInt( fs, "var_all", var_all );
cvWriteInt( fs, "var_count", var_count );
cvWriteInt( fs, "ord_var_count", ord_var_count );
cvWriteInt( fs, "cat_var_count", cat_var_count );
cvStartWriteStruct( fs, "training_params", CV_NODE_MAP );
cvWriteInt( fs, "use_surrogates", params.use_surrogates ? 1 : 0 );
if( is_classifier )
{
cvWriteInt( fs, "max_categories", params.max_categories );
}
else
{
cvWriteReal( fs, "regression_accuracy", params.regression_accuracy );
}
cvWriteInt( fs, "max_depth", params.max_depth );
cvWriteInt( fs, "min_sample_count", params.min_sample_count );
cvWriteInt( fs, "cross_validation_folds", params.cv_folds );
if( params.cv_folds > 1 )
{
cvWriteInt( fs, "use_1se_rule", params.use_1se_rule ? 1 : 0 );
cvWriteInt( fs, "truncate_pruned_tree", params.truncate_pruned_tree ? 1 : 0 );
}
if( priors )
cvWrite( fs, "priors", priors );
cvEndWriteStruct( fs );
if( var_idx )
cvWrite( fs, "var_idx", var_idx );
cvStartWriteStruct( fs, "var_type", CV_NODE_SEQ+CV_NODE_FLOW );
for( vi = 0; vi < vcount; vi++ )
cvWriteInt( fs, 0, var_type->data.i[vi] >= 0 );
cvEndWriteStruct( fs );
if( cat_count && (cat_var_count > 0 || is_classifier) )
{
CV_ASSERT( cat_count != 0 );
cvWrite( fs, "cat_count", cat_count );
cvWrite( fs, "cat_map", cat_map );
}
__END__;
}
void CvDTreeTrainData::read_params( CvFileStorage* fs, CvFileNode* node )
{
CV_FUNCNAME( "CvDTreeTrainData::read_params" );
__BEGIN__;
CvFileNode *tparams_node, *vartype_node;
CvSeqReader reader;
int vi, max_split_size, tree_block_size;
is_classifier = (cvReadIntByName( fs, node, "is_classifier" ) != 0);
var_all = cvReadIntByName( fs, node, "var_all" );
var_count = cvReadIntByName( fs, node, "var_count", var_all );
cat_var_count = cvReadIntByName( fs, node, "cat_var_count" );
ord_var_count = cvReadIntByName( fs, node, "ord_var_count" );
tparams_node = cvGetFileNodeByName( fs, node, "training_params" );
if( tparams_node ) // training parameters are not necessary
{
params.use_surrogates = cvReadIntByName( fs, tparams_node, "use_surrogates", 1 ) != 0;
if( is_classifier )
{
params.max_categories = cvReadIntByName( fs, tparams_node, "max_categories" );
}
else
{
params.regression_accuracy =
(float)cvReadRealByName( fs, tparams_node, "regression_accuracy" );
}
params.max_depth = cvReadIntByName( fs, tparams_node, "max_depth" );
params.min_sample_count = cvReadIntByName( fs, tparams_node, "min_sample_count" );
params.cv_folds = cvReadIntByName( fs, tparams_node, "cross_validation_folds" );
if( params.cv_folds > 1 )
{
params.use_1se_rule = cvReadIntByName( fs, tparams_node, "use_1se_rule" ) != 0;
params.truncate_pruned_tree =
cvReadIntByName( fs, tparams_node, "truncate_pruned_tree" ) != 0;
}
priors = (CvMat*)cvReadByName( fs, tparams_node, "priors" );
if( priors )
{
if( !CV_IS_MAT(priors) )
CV_ERROR( CV_StsParseError, "priors must stored as a matrix" );
priors_mult = cvCloneMat( priors );
}
}
CV_CALL( var_idx = (CvMat*)cvReadByName( fs, node, "var_idx" ));
if( var_idx )
{
if( !CV_IS_MAT(var_idx) ||
var_idx->cols != 1 && var_idx->rows != 1 ||
var_idx->cols + var_idx->rows - 1 != var_count ||
CV_MAT_TYPE(var_idx->type) != CV_32SC1 )
CV_ERROR( CV_StsParseError,
"var_idx (if exist) must be valid 1d integer vector containing <var_count> elements" );
for( vi = 0; vi < var_count; vi++ )
if( (unsigned)var_idx->data.i[vi] >= (unsigned)var_all )
CV_ERROR( CV_StsOutOfRange, "some of var_idx elements are out of range" );
}
////// read var type
CV_CALL( var_type = cvCreateMat( 1, var_count + 2, CV_32SC1 ));
cat_var_count = 0;
ord_var_count = -1;
vartype_node = cvGetFileNodeByName( fs, node, "var_type" );
if( vartype_node && CV_NODE_TYPE(vartype_node->tag) == CV_NODE_INT && var_count == 1 )
var_type->data.i[0] = vartype_node->data.i ? cat_var_count++ : ord_var_count--;
else
{
if( !vartype_node || CV_NODE_TYPE(vartype_node->tag) != CV_NODE_SEQ ||
vartype_node->data.seq->total != var_count )
CV_ERROR( CV_StsParseError, "var_type must exist and be a sequence of 0's and 1's" );
cvStartReadSeq( vartype_node->data.seq, &reader );
for( vi = 0; vi < var_count; vi++ )
{
CvFileNode* n = (CvFileNode*)reader.ptr;
if( CV_NODE_TYPE(n->tag) != CV_NODE_INT || (n->data.i & ~1) )
CV_ERROR( CV_StsParseError, "var_type must exist and be a sequence of 0's and 1's" );
var_type->data.i[vi] = n->data.i ? cat_var_count++ : ord_var_count--;
CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
}
}
var_type->data.i[var_count] = cat_var_count;
ord_var_count = ~ord_var_count;
if( cat_var_count != cat_var_count || ord_var_count != ord_var_count )
CV_ERROR( CV_StsParseError, "var_type is inconsistent with cat_var_count and ord_var_count" );
//////
if( cat_var_count > 0 || is_classifier )
{
int ccount, total_c_count = 0;
CV_CALL( cat_count = (CvMat*)cvReadByName( fs, node, "cat_count" ));
CV_CALL( cat_map = (CvMat*)cvReadByName( fs, node, "cat_map" ));
if( !CV_IS_MAT(cat_count) || !CV_IS_MAT(cat_map) ||
cat_count->cols != 1 && cat_count->rows != 1 ||
CV_MAT_TYPE(cat_count->type) != CV_32SC1 ||
cat_count->cols + cat_count->rows - 1 != cat_var_count + is_classifier ||
cat_map->cols != 1 && cat_map->rows != 1 ||
CV_MAT_TYPE(cat_map->type) != CV_32SC1 )
CV_ERROR( CV_StsParseError,
"Both cat_count and cat_map must exist and be valid 1d integer vectors of an appropriate size" );
ccount = cat_var_count + is_classifier;
CV_CALL( cat_ofs = cvCreateMat( 1, ccount + 1, CV_32SC1 ));
cat_ofs->data.i[0] = 0;
max_c_count = 1;
for( vi = 0; vi < ccount; vi++ )
{
int val = cat_count->data.i[vi];
if( val <= 0 )
CV_ERROR( CV_StsOutOfRange, "some of cat_count elements are out of range" );
max_c_count = MAX( max_c_count, val );
cat_ofs->data.i[vi+1] = total_c_count += val;
}
if( cat_map->cols + cat_map->rows - 1 != total_c_count )
CV_ERROR( CV_StsBadSize,
"cat_map vector length is not equal to the total number of categories in all categorical vars" );
}
max_split_size = cvAlign(sizeof(CvDTreeSplit) +
(MAX(0,max_c_count - 33)/32)*sizeof(int),sizeof(void*));
tree_block_size = MAX((int)sizeof(CvDTreeNode)*8, max_split_size);
tree_block_size = MAX(tree_block_size + block_size_delta, min_block_size);
CV_CALL( tree_storage = cvCreateMemStorage( tree_block_size ));
CV_CALL( node_heap = cvCreateSet( 0, sizeof(node_heap[0]),
sizeof(CvDTreeNode), tree_storage ));
CV_CALL( split_heap = cvCreateSet( 0, sizeof(split_heap[0]),
max_split_size, tree_storage ));
__END__;
}
/////////////////////// Decision Tree /////////////////////////
CvDTree::CvDTree()
{
data = 0;
var_importance = 0;
default_model_name = "my_tree";
clear();
}
void CvDTree::clear()
{
cvReleaseMat( &var_importance );
if( data )
{
if( !data->shared )
delete data;
else
free_tree();
data = 0;
}
root = 0;
pruned_tree_idx = -1;
}
CvDTree::~CvDTree()
{
clear();
}
const CvDTreeNode* CvDTree::get_root() const
{
return root;
}
int CvDTree::get_pruned_tree_idx() const
{
return pruned_tree_idx;
}
CvDTreeTrainData* CvDTree::get_data()
{
return data;
}
bool CvDTree::train( const CvMat* _train_data, int _tflag,
const CvMat* _responses, const CvMat* _var_idx,
const CvMat* _sample_idx, const CvMat* _var_type,
const CvMat* _missing_mask, CvDTreeParams _params )
{
bool result = false;
CV_FUNCNAME( "CvDTree::train" );
__BEGIN__;
clear();
data = new CvDTreeTrainData( _train_data, _tflag, _responses,
_var_idx, _sample_idx, _var_type,
_missing_mask, _params, false );
CV_CALL( result = do_train(0));
__END__;
return result;
}
bool CvDTree::train( CvDTreeTrainData* _data, const CvMat* _subsample_idx )
{
bool result = false;
CV_FUNCNAME( "CvDTree::train" );
__BEGIN__;
clear();
data = _data;
data->shared = true;
CV_CALL( result = do_train(_subsample_idx));
__END__;
return result;
}
bool CvDTree::do_train( const CvMat* _subsample_idx )
{
bool result = false;
CV_FUNCNAME( "CvDTree::do_train" );
__BEGIN__;
root = data->subsample_data( _subsample_idx );
CV_CALL( try_split_node(root));
if( data->params.cv_folds > 0 )
CV_CALL( prune_cv());
if( !data->shared )
data->free_train_data();
result = true;
__END__;
return result;
}
void CvDTree::try_split_node( CvDTreeNode* node )
{
CvDTreeSplit* best_split = 0;
int i, n = node->sample_count, vi;
bool can_split = true;
double quality_scale;
calc_node_value( node );
if( node->sample_count <= data->params.min_sample_count ||
node->depth >= data->params.max_depth )
can_split = false;
if( can_split && data->is_classifier )
{
// check if we have a "pure" node,
// we assume that cls_count is filled by calc_node_value()
int* cls_count = data->counts->data.i;
int nz = 0, m = data->get_num_classes();
for( i = 0; i < m; i++ )
nz += cls_count[i] != 0;
if( nz == 1 ) // there is only one class
can_split = false;
}
else if( can_split )
{
if( sqrt(node->node_risk)/n < data->params.regression_accuracy )
can_split = false;
}
if( can_split )
{
best_split = find_best_split(node);
// TODO: check the split quality ...
node->split = best_split;
}
if( !can_split || !best_split )
{
data->free_node_data(node);
return;
}
quality_scale = calc_node_dir( node );
if( data->params.use_surrogates )
{
// find all the surrogate splits
// and sort them by their similarity to the primary one
for( vi = 0; vi < data->var_count; vi++ )
{
CvDTreeSplit* split;
int ci = data->get_var_type(vi);
if( vi == best_split->var_idx )
continue;
if( ci >= 0 )
split = find_surrogate_split_cat( node, vi );
else
split = find_surrogate_split_ord( node, vi );
if( split )
{
// insert the split
CvDTreeSplit* prev_split = node->split;
split->quality = (float)(split->quality*quality_scale);
while( prev_split->next &&
prev_split->next->quality > split->quality )
prev_split = prev_split->next;
split->next = prev_split->next;
prev_split->next = split;
}
}
}
split_node_data( node );
try_split_node( node->left );
try_split_node( node->right );
}
// calculate direction (left(-1),right(1),missing(0))
// for each sample using the best split
// the function returns scale coefficients for surrogate split quality factors.
// the scale is applied to normalize surrogate split quality relatively to the
// best (primary) split quality. That is, if a surrogate split is absolutely
// identical to the primary split, its quality will be set to the maximum value =
// quality of the primary split; otherwise, it will be lower.
// besides, the function compute node->maxlr,
// minimum possible quality (w/o considering the above mentioned scale)
// for a surrogate split. Surrogate splits with quality less than node->maxlr
// are not discarded.
double CvDTree::calc_node_dir( CvDTreeNode* node )
{
char* dir = (char*)data->direction->data.ptr;
int i, n = node->sample_count, vi = node->split->var_idx;
double L, R;
assert( !node->split->inversed );
if( data->get_var_type(vi) >= 0 ) // split on categorical var
{
const int* labels = data->get_cat_var_data(node,vi);
const int* subset = node->split->subset;
if( !data->have_priors )
{
int sum = 0, sum_abs = 0;
for( i = 0; i < n; i++ )
{
int idx = labels[i];
int d = idx >= 0 ? CV_DTREE_CAT_DIR(idx,subset) : 0;
sum += d; sum_abs += d & 1;
dir[i] = (char)d;
}
R = (sum_abs + sum) >> 1;
L = (sum_abs - sum) >> 1;
}
else
{
const int* responses = data->get_class_labels(node);
const double* priors = data->priors_mult->data.db;
double sum = 0, sum_abs = 0;
for( i = 0; i < n; i++ )
{
int idx = labels[i];
double w = priors[responses[i]];
int d = idx >= 0 ? CV_DTREE_CAT_DIR(idx,subset) : 0;
sum += d*w; sum_abs += (d & 1)*w;
dir[i] = (char)d;
}
R = (sum_abs + sum) * 0.5;
L = (sum_abs - sum) * 0.5;
}
}
else // split on ordered var
{
const CvPair32s32f* sorted = data->get_ord_var_data(node,vi);
int split_point = node->split->ord.split_point;
int n1 = node->get_num_valid(vi);
assert( 0 <= split_point && split_point < n1-1 );
if( !data->have_priors )
{
for( i = 0; i <= split_point; i++ )
dir[sorted[i].i] = (char)-1;
for( ; i < n1; i++ )
dir[sorted[i].i] = (char)1;
for( ; i < n; i++ )
dir[sorted[i].i] = (char)0;
L = split_point-1;
R = n1 - split_point + 1;
}
else
{
const int* responses = data->get_class_labels(node);
const double* priors = data->priors_mult->data.db;
L = R = 0;
for( i = 0; i <= split_point; i++ )
{
int idx = sorted[i].i;
double w = priors[responses[idx]];
dir[idx] = (char)-1;
L += w;
}
for( ; i < n1; i++ )
{
int idx = sorted[i].i;
double w = priors[responses[idx]];
dir[idx] = (char)1;
R += w;
}
for( ; i < n; i++ )
dir[sorted[i].i] = (char)0;
}
}
node->maxlr = MAX( L, R );
return node->split->quality/(L + R);
}
CvDTreeSplit* CvDTree::find_best_split( CvDTreeNode* node )
{
int vi;
CvDTreeSplit *best_split = 0, *split = 0, *t;
for( vi = 0; vi < data->var_count; vi++ )
{
int ci = data->get_var_type(vi);
if( node->get_num_valid(vi) <= 1 )
continue;
if( data->is_classifier )
{
if( ci >= 0 )
split = find_split_cat_class( node, vi );
else
split = find_split_ord_class( node, vi );
}
else
{
if( ci >= 0 )
split = find_split_cat_reg( node, vi );
else
split = find_split_ord_reg( node, vi );
}
if( split )
{
if( !best_split || best_split->quality < split->quality )
CV_SWAP( best_split, split, t );
if( split )
cvSetRemoveByPtr( data->split_heap, split );
}
}
return best_split;
}
CvDTreeSplit* CvDTree::find_split_ord_class( CvDTreeNode* node, int vi )
{
const float epsilon = FLT_EPSILON*2;
const CvPair32s32f* sorted = data->get_ord_var_data(node, vi);
const int* responses = data->get_class_labels(node);
int n = node->sample_count;
int n1 = node->get_num_valid(vi);
int m = data->get_num_classes();
const int* rc0 = data->counts->data.i;
int* lc = (int*)cvStackAlloc(m*sizeof(lc[0]));
int* rc = (int*)cvStackAlloc(m*sizeof(rc[0]));
int i, best_i = -1;
double lsum2 = 0, rsum2 = 0, best_val = 0;
const double* priors = data->have_priors ? data->priors_mult->data.db : 0;
// init arrays of class instance counters on both sides of the split
for( i = 0; i < m; i++ )
{
lc[i] = 0;
rc[i] = rc0[i];
}
// compensate for missing values
for( i = n1; i < n; i++ )
rc[responses[sorted[i].i]]--;
if( !priors )
{
int L = 0, R = n1;
for( i = 0; i < m; i++ )
rsum2 += (double)rc[i]*rc[i];
for( i = 0; i < n1 - 1; i++ )
{
int idx = responses[sorted[i].i];
int lv, rv;
L++; R--;
lv = lc[idx]; rv = rc[idx];
lsum2 += lv*2 + 1;
rsum2 -= rv*2 - 1;
lc[idx] = lv + 1; rc[idx] = rv - 1;
if( sorted[i].val + epsilon < sorted[i+1].val )
{
double val = (lsum2*R + rsum2*L)/((double)L*R);
if( best_val < val )
{
best_val = val;
best_i = i;
}
}
}
}
else
{
double L = 0, R = 0;
for( i = 0; i < m; i++ )
{
double wv = rc[i]*priors[i];
R += wv;
rsum2 += wv*wv;
}
for( i = 0; i < n1 - 1; i++ )
{
int idx = responses[sorted[i].i];
int lv, rv;
double p = priors[idx], p2 = p*p;
L += p; R -= p;
lv = lc[idx]; rv = rc[idx];
lsum2 += p2*(lv*2 + 1);
rsum2 -= p2*(rv*2 - 1);
lc[idx] = lv + 1; rc[idx] = rv - 1;
if( sorted[i].val + epsilon < sorted[i+1].val )
{
double val = (lsum2*R + rsum2*L)/((double)L*R);
if( best_val < val )
{
best_val = val;
best_i = i;
}
}
}
}
return best_i >= 0 ? data->new_split_ord( vi,
(sorted[best_i].val + sorted[best_i+1].val)*0.5f, best_i,
0, (float)best_val ) : 0;
}
void CvDTree::cluster_categories( const int* vectors, int n, int m,
int* csums, int k, int* labels )
{
// TODO: consider adding priors (class weights) and sample weights to the clustering algorithm
int iters = 0, max_iters = 100;
int i, j, idx;
double* buf = (double*)cvStackAlloc( (n + k)*sizeof(buf[0]) );
double *v_weights = buf, *c_weights = buf + k;
bool modified = true;
CvRNG* r = &data->rng;
// assign labels randomly
for( i = idx = 0; i < n; i++ )
{
int sum = 0;
const int* v = vectors + i*m;
labels[i] = idx++;
idx &= idx < k ? -1 : 0;
// compute weight of each vector
for( j = 0; j < m; j++ )
sum += v[j];
v_weights[i] = sum ? 1./sum : 0.;
}
for( i = 0; i < n; i++ )
{
int i1 = cvRandInt(r) % n;
int i2 = cvRandInt(r) % n;
CV_SWAP( labels[i1], labels[i2], j );
}
for( iters = 0; iters <= max_iters; iters++ )
{
// calculate csums
for( i = 0; i < k; i++ )
{
for( j = 0; j < m; j++ )
csums[i*m + j] = 0;
}
for( i = 0; i < n; i++ )
{
const int* v = vectors + i*m;
int* s = csums + labels[i]*m;
for( j = 0; j < m; j++ )
s[j] += v[j];
}
// exit the loop here, when we have up-to-date csums
if( iters == max_iters || !modified )
break;
modified = false;
// calculate weight of each cluster
for( i = 0; i < k; i++ )
{
const int* s = csums + i*m;
int sum = 0;
for( j = 0; j < m; j++ )
sum += s[j];
c_weights[i] = sum ? 1./sum : 0;
}
// now for each vector determine the closest cluster
for( i = 0; i < n; i++ )
{
const int* v = vectors + i*m;
double alpha = v_weights[i];
double min_dist2 = DBL_MAX;
int min_idx = -1;
for( idx = 0; idx < k; idx++ )
{
const int* s = csums + idx*m;
double dist2 = 0., beta = c_weights[idx];
for( j = 0; j < m; j++ )
{
double t = v[j]*alpha - s[j]*beta;
dist2 += t*t;
}
if( min_dist2 > dist2 )
{
min_dist2 = dist2;
min_idx = idx;
}
}
if( min_idx != labels[i] )
modified = true;
labels[i] = min_idx;
}
}
}
CvDTreeSplit* CvDTree::find_split_cat_class( CvDTreeNode* node, int vi )
{
CvDTreeSplit* split;
const int* labels = data->get_cat_var_data(node, vi);
const int* responses = data->get_class_labels(node);
int ci = data->get_var_type(vi);
int n = node->sample_count;
int m = data->get_num_classes();
int _mi = data->cat_count->data.i[ci], mi = _mi;
int* lc = (int*)cvStackAlloc(m*sizeof(lc[0]));
int* rc = (int*)cvStackAlloc(m*sizeof(rc[0]));
int* _cjk = (int*)cvStackAlloc(m*(mi+1)*sizeof(_cjk[0]))+m, *cjk = _cjk;
double* c_weights = (double*)cvStackAlloc( mi*sizeof(c_weights[0]) );
int* cluster_labels = 0;
int** int_ptr = 0;
int i, j, k, idx;
double L = 0, R = 0;
double best_val = 0;
int prevcode = 0, best_subset = -1, subset_i, subset_n, subtract = 0;
const double* priors = data->priors_mult->data.db;
// init array of counters:
// c_{jk} - number of samples that have vi-th input variable = j and response = k.
for( j = -1; j < mi; j++ )
for( k = 0; k < m; k++ )
cjk[j*m + k] = 0;
for( i = 0; i < n; i++ )
{
j = labels[i];
k = responses[i];
cjk[j*m + k]++;
}
if( m > 2 )
{
if( mi > data->params.max_categories )
{
mi = MIN(data->params.max_categories, n);
cjk += _mi*m;
cluster_labels = (int*)cvStackAlloc(mi*sizeof(cluster_labels[0]));
cluster_categories( _cjk, _mi, m, cjk, mi, cluster_labels );
}
subset_i = 1;
subset_n = 1 << mi;
}
else
{
assert( m == 2 );
int_ptr = (int**)cvStackAlloc( mi*sizeof(int_ptr[0]) );
for( j = 0; j < mi; j++ )
int_ptr[j] = cjk + j*2 + 1;
icvSortIntPtr( int_ptr, mi, 0 );
subset_i = 0;
subset_n = mi;
}
for( k = 0; k < m; k++ )
{
int sum = 0;
for( j = 0; j < mi; j++ )
sum += cjk[j*m + k];
rc[k] = sum;
lc[k] = 0;
}
for( j = 0; j < mi; j++ )
{
double sum = 0;
for( k = 0; k < m; k++ )
sum += cjk[j*m + k]*priors[k];
c_weights[j] = sum;
R += c_weights[j];
}
for( ; subset_i < subset_n; subset_i++ )
{
double weight;
int* crow;
double lsum2 = 0, rsum2 = 0;
if( m == 2 )
idx = (int)(int_ptr[subset_i] - cjk)/2;
else
{
int graycode = (subset_i>>1)^subset_i;
int diff = graycode ^ prevcode;
// determine index of the changed bit.
Cv32suf u;
idx = diff >= (1 << 16) ? 16 : 0;
u.f = (float)(((diff >> 16) | diff) & 65535);
idx += (u.i >> 23) - 127;
subtract = graycode < prevcode;
prevcode = graycode;
}
crow = cjk + idx*m;
weight = c_weights[idx];
if( weight < FLT_EPSILON )
continue;
if( !subtract )
{
for( k = 0; k < m; k++ )
{
int t = crow[k];
int lval = lc[k] + t;
int rval = rc[k] - t;
double p = priors[k], p2 = p*p;
lsum2 += p2*lval*lval;
rsum2 += p2*rval*rval;
lc[k] = lval; rc[k] = rval;
}
L += weight;
R -= weight;
}
else
{
for( k = 0; k < m; k++ )
{
int t = crow[k];
int lval = lc[k] - t;
int rval = rc[k] + t;
double p = priors[k], p2 = p*p;
lsum2 += p2*lval*lval;
rsum2 += p2*rval*rval;
lc[k] = lval; rc[k] = rval;
}
L -= weight;
R += weight;
}
if( L > FLT_EPSILON && R > FLT_EPSILON )
{
double val = (lsum2*R + rsum2*L)/((double)L*R);
if( best_val < val )
{
best_val = val;
best_subset = subset_i;
}
}
}
if( best_subset < 0 )
return 0;
split = data->new_split_cat( vi, (float)best_val );
if( m == 2 )
{
for( i = 0; i <= best_subset; i++ )
{
idx = (int)(int_ptr[i] - cjk) >> 1;
split->subset[idx >> 5] |= 1 << (idx & 31);
}
}
else
{
for( i = 0; i < _mi; i++ )
{
idx = cluster_labels ? cluster_labels[i] : i;
if( best_subset & (1 << idx) )
split->subset[i >> 5] |= 1 << (i & 31);
}
}
return split;
}
CvDTreeSplit* CvDTree::find_split_ord_reg( CvDTreeNode* node, int vi )
{
const float epsilon = FLT_EPSILON*2;
const CvPair32s32f* sorted = data->get_ord_var_data(node, vi);
const float* responses = data->get_ord_responses(node);
int n = node->sample_count;
int n1 = node->get_num_valid(vi);
int i, best_i = -1;
double best_val = 0, lsum = 0, rsum = node->value*n;
int L = 0, R = n1;
// compensate for missing values
for( i = n1; i < n; i++ )
rsum -= responses[sorted[i].i];
// find the optimal split
for( i = 0; i < n1 - 1; i++ )
{
float t = responses[sorted[i].i];
L++; R--;
lsum += t;
rsum -= t;
if( sorted[i].val + epsilon < sorted[i+1].val )
{
double val = (lsum*lsum*R + rsum*rsum*L)/((double)L*R);
if( best_val < val )
{
best_val = val;
best_i = i;
}
}
}
return best_i >= 0 ? data->new_split_ord( vi,
(sorted[best_i].val + sorted[best_i+1].val)*0.5f, best_i,
0, (float)best_val ) : 0;
}
CvDTreeSplit* CvDTree::find_split_cat_reg( CvDTreeNode* node, int vi )
{
CvDTreeSplit* split;
const int* labels = data->get_cat_var_data(node, vi);
const float* responses = data->get_ord_responses(node);
int ci = data->get_var_type(vi);
int n = node->sample_count;
int mi = data->cat_count->data.i[ci];
double* sum = (double*)cvStackAlloc( (mi+1)*sizeof(sum[0]) ) + 1;
int* counts = (int*)cvStackAlloc( (mi+1)*sizeof(counts[0]) ) + 1;
double** sum_ptr = 0;
int i, L = 0, R = 0;
double best_val = 0, lsum = 0, rsum = 0;
int best_subset = -1, subset_i;
for( i = -1; i < mi; i++ )
sum[i] = counts[i] = 0;
// calculate sum response and weight of each category of the input var
for( i = 0; i < n; i++ )
{
int idx = labels[i];
double s = sum[idx] + responses[i];
int nc = counts[idx] + 1;
sum[idx] = s;
counts[idx] = nc;
}
// calculate average response in each category
for( i = 0; i < mi; i++ )
{
R += counts[i];
rsum += sum[i];
sum[i] /= MAX(counts[i],1);
sum_ptr[i] = sum + i;
}
icvSortDblPtr( sum_ptr, mi, 0 );
// revert back to unnormalized sums
// (there should be a very little loss of accuracy)
for( i = 0; i < mi; i++ )
sum[i] *= counts[i];
for( subset_i = 0; subset_i < mi-1; subset_i++ )
{
int idx = (int)(sum_ptr[subset_i] - sum);
int ni = counts[idx];
if( ni )
{
double s = sum[idx];
lsum += s; L += ni;
rsum -= s; R -= ni;
if( L && R )
{
double val = (lsum*lsum*R + rsum*rsum*L)/((double)L*R);
if( best_val < val )
{
best_val = val;
best_subset = subset_i;
}
}
}
}
if( best_subset < 0 )
return 0;
split = data->new_split_cat( vi, (float)best_val );
for( i = 0; i <= best_subset; i++ )
{
int idx = (int)(sum_ptr[i] - sum);
split->subset[idx >> 5] |= 1 << (idx & 31);
}
return split;
}
CvDTreeSplit* CvDTree::find_surrogate_split_ord( CvDTreeNode* node, int vi )
{
const float epsilon = FLT_EPSILON*2;
const CvPair32s32f* sorted = data->get_ord_var_data(node, vi);
const char* dir = (char*)data->direction->data.ptr;
int n1 = node->get_num_valid(vi);
// LL - number of samples that both the primary and the surrogate splits send to the left
// LR - ... primary split sends to the left and the surrogate split sends to the right
// RL - ... primary split sends to the right and the surrogate split sends to the left
// RR - ... both send to the right
int i, best_i = -1, best_inversed = 0;
double best_val;
if( !data->have_priors )
{
int LL = 0, RL = 0, LR, RR;
int worst_val = cvFloor(node->maxlr), _best_val = worst_val;
int sum = 0, sum_abs = 0;
for( i = 0; i < n1; i++ )
{
int d = dir[sorted[i].i];
sum += d; sum_abs += d & 1;
}
// sum_abs = R + L; sum = R - L
RR = (sum_abs + sum) >> 1;
LR = (sum_abs - sum) >> 1;
// initially all the samples are sent to the right by the surrogate split,
// LR of them are sent to the left by primary split, and RR - to the right.
// now iteratively compute LL, LR, RL and RR for every possible surrogate split value.
for( i = 0; i < n1 - 1; i++ )
{
int d = dir[sorted[i].i];
if( d < 0 )
{
LL++; LR--;
if( LL + RR > _best_val && sorted[i].val + epsilon < sorted[i+1].val )
{
best_val = LL + RR;
best_i = i; best_inversed = 0;
}
}
else if( d > 0 )
{
RL++; RR--;
if( RL + LR > _best_val && sorted[i].val + epsilon < sorted[i+1].val )
{
best_val = RL + LR;
best_i = i; best_inversed = 1;
}
}
}
best_val = _best_val;
}
else
{
double LL = 0, RL = 0, LR, RR;
double worst_val = node->maxlr;
double sum = 0, sum_abs = 0;
const double* priors = data->priors_mult->data.db;
const int* responses = data->get_class_labels(node);
best_val = worst_val;
for( i = 0; i < n1; i++ )
{
int idx = sorted[i].i;
double w = priors[responses[idx]];
int d = dir[idx];
sum += d*w; sum_abs += (d & 1)*w;
}
// sum_abs = R + L; sum = R - L
RR = (sum_abs + sum)*0.5;
LR = (sum_abs - sum)*0.5;
// initially all the samples are sent to the right by the surrogate split,
// LR of them are sent to the left by primary split, and RR - to the right.
// now iteratively compute LL, LR, RL and RR for every possible surrogate split value.
for( i = 0; i < n1 - 1; i++ )
{
int idx = sorted[i].i;
double w = priors[responses[idx]];
int d = dir[idx];
if( d < 0 )
{
LL += w; LR -= w;
if( LL + RR > best_val && sorted[i].val + epsilon < sorted[i+1].val )
{
best_val = LL + RR;
best_i = i; best_inversed = 0;
}
}
else if( d > 0 )
{
RL += w; RR -= w;
if( RL + LR > best_val && sorted[i].val + epsilon < sorted[i+1].val )
{
best_val = RL + LR;
best_i = i; best_inversed = 1;
}
}
}
}
return best_i >= 0 && best_val > node->maxlr ? data->new_split_ord( vi,
(sorted[best_i].val + sorted[best_i+1].val)*0.5f, best_i,
best_inversed, (float)best_val ) : 0;
}
CvDTreeSplit* CvDTree::find_surrogate_split_cat( CvDTreeNode* node, int vi )
{
const int* labels = data->get_cat_var_data(node, vi);
const char* dir = (char*)data->direction->data.ptr;
int n = node->sample_count;
// LL - number of samples that both the primary and the surrogate splits send to the left
// LR - ... primary split sends to the left and the surrogate split sends to the right
// RL - ... primary split sends to the right and the surrogate split sends to the left
// RR - ... both send to the right
CvDTreeSplit* split = data->new_split_cat( vi, 0 );
int i, mi = data->cat_count->data.i[data->get_var_type(vi)], l_win = 0;
double best_val = 0;
double* lc = (double*)cvStackAlloc( (mi+1)*2*sizeof(lc[0]) ) + 1;
double* rc = lc + mi + 1;
for( i = -1; i < mi; i++ )
lc[i] = rc[i] = 0;
// for each category calculate the weight of samples
// sent to the left (lc) and to the right (rc) by the primary split
if( !data->have_priors )
{
int* _lc = (int*)cvStackAlloc((mi+2)*2*sizeof(_lc[0])) + 1;
int* _rc = _lc + mi + 1;
for( i = -1; i < mi; i++ )
_lc[i] = _rc[i] = 0;
for( i = 0; i < n; i++ )
{
int idx = labels[i];
int d = dir[i];
int sum = _lc[idx] + d;
int sum_abs = _rc[idx] + (d & 1);
_lc[idx] = sum; _rc[idx] = sum_abs;
}
for( i = 0; i < mi; i++ )
{
int sum = _lc[i];
int sum_abs = _rc[i];
lc[i] = (sum_abs - sum) >> 1;
rc[i] = (sum_abs + sum) >> 1;
}
}
else
{
const double* priors = data->priors_mult->data.db;
const int* responses = data->get_class_labels(node);
for( i = 0; i < n; i++ )
{
int idx = labels[i];
double w = priors[responses[i]];
int d = dir[i];
double sum = lc[idx] + d*w;
double sum_abs = rc[idx] + (d & 1)*w;
lc[idx] = sum; rc[idx] = sum_abs;
}
for( i = 0; i < mi; i++ )
{
double sum = lc[i];
double sum_abs = rc[i];
lc[i] = (sum_abs - sum) * 0.5;
rc[i] = (sum_abs + sum) * 0.5;
}
}
// 2. now form the split.
// in each category send all the samples to the same direction as majority
for( i = 0; i < mi; i++ )
{
double lval = lc[i], rval = rc[i];
if( lval > rval )
{
split->subset[i >> 5] |= 1 << (i & 31);
best_val += lval;
l_win++;
}
else
best_val += rval;
}
split->quality = (float)best_val;
if( split->quality <= node->maxlr || l_win == 0 || l_win == mi )
cvSetRemoveByPtr( data->split_heap, split ), split = 0;
return split;
}
void CvDTree::calc_node_value( CvDTreeNode* node )
{
int i, j, k, n = node->sample_count, cv_n = data->params.cv_folds;
const int* cv_labels = data->get_labels(node);
if( data->is_classifier )
{
// in case of classification tree:
// * node value is the label of the class that has the largest weight in the node.
// * node risk is the weighted number of misclassified samples,
// * j-th cross-validation fold value and risk are calculated as above,
// but using the samples with cv_labels(*)!=j.
// * j-th cross-validation fold error is calculated as the weighted number of
// misclassified samples with cv_labels(*)==j.
// compute the number of instances of each class
int* cls_count = data->counts->data.i;
const int* responses = data->get_class_labels(node);
int m = data->get_num_classes();
int* cv_cls_count = (int*)cvStackAlloc(m*cv_n*sizeof(cv_cls_count[0]));
double max_val = -1, total_weight = 0;
int max_k = -1;
double* priors = data->priors_mult->data.db;
for( k = 0; k < m; k++ )
cls_count[k] = 0;
if( cv_n == 0 )
{
for( i = 0; i < n; i++ )
cls_count[responses[i]]++;
}
else
{
for( j = 0; j < cv_n; j++ )
for( k = 0; k < m; k++ )
cv_cls_count[j*m + k] = 0;
for( i = 0; i < n; i++ )
{
j = cv_labels[i]; k = responses[i];
cv_cls_count[j*m + k]++;
}
for( j = 0; j < cv_n; j++ )
for( k = 0; k < m; k++ )
cls_count[k] += cv_cls_count[j*m + k];
}
if( data->have_priors && node->parent == 0 )
{
// compute priors_mult from priors, take the sample ratio into account.
double sum = 0;
for( k = 0; k < m; k++ )
{
int n_k = cls_count[k];
priors[k] = data->priors->data.db[k]*(n_k ? 1./n_k : 0.);
sum += priors[k];
}
sum = 1./sum;
for( k = 0; k < m; k++ )
priors[k] *= sum;
}
for( k = 0; k < m; k++ )
{
double val = cls_count[k]*priors[k];
total_weight += val;
if( max_val < val )
{
max_val = val;
max_k = k;
}
}
node->class_idx = max_k;
node->value = data->cat_map->data.i[
data->cat_ofs->data.i[data->cat_var_count] + max_k];
node->node_risk = total_weight - max_val;
for( j = 0; j < cv_n; j++ )
{
double sum_k = 0, sum = 0, max_val_k = 0;
max_val = -1; max_k = -1;
for( k = 0; k < m; k++ )
{
double w = priors[k];
double val_k = cv_cls_count[j*m + k]*w;
double val = cls_count[k]*w - val_k;
sum_k += val_k;
sum += val;
if( max_val < val )
{
max_val = val;
max_val_k = val_k;
max_k = k;
}
}
node->cv_Tn[j] = INT_MAX;
node->cv_node_risk[j] = sum - max_val;
node->cv_node_error[j] = sum_k - max_val_k;
}
}
else
{
// in case of regression tree:
// * node value is 1/n*sum_i(Y_i), where Y_i is i-th response,
// n is the number of samples in the node.
// * node risk is the sum of squared errors: sum_i((Y_i - <node_value>)^2)
// * j-th cross-validation fold value and risk are calculated as above,
// but using the samples with cv_labels(*)!=j.
// * j-th cross-validation fold error is calculated
// using samples with cv_labels(*)==j as the test subset:
// error_j = sum_(i,cv_labels(i)==j)((Y_i - <node_value_j>)^2),
// where node_value_j is the node value calculated
// as described in the previous bullet, and summation is done
// over the samples with cv_labels(*)==j.
double sum = 0, sum2 = 0;
const float* values = data->get_ord_responses(node);
double *cv_sum = 0, *cv_sum2 = 0;
int* cv_count = 0;
if( cv_n == 0 )
{
for( i = 0; i < n; i++ )
{
double t = values[i];
sum += t;
sum2 += t*t;
}
}
else
{
cv_sum = (double*)cvStackAlloc( cv_n*sizeof(cv_sum[0]) );
cv_sum2 = (double*)cvStackAlloc( cv_n*sizeof(cv_sum2[0]) );
cv_count = (int*)cvStackAlloc( cv_n*sizeof(cv_count[0]) );
for( j = 0; j < cv_n; j++ )
{
cv_sum[j] = cv_sum2[j] = 0.;
cv_count[j] = 0;
}
for( i = 0; i < n; i++ )
{
j = cv_labels[i];
double t = values[i];
double s = cv_sum[j] + t;
double s2 = cv_sum2[j] + t*t;
int nc = cv_count[j] + 1;
cv_sum[j] = s;
cv_sum2[j] = s2;
cv_count[j] = nc;
}
for( j = 0; j < cv_n; j++ )
{
sum += cv_sum[j];
sum2 += cv_sum2[j];
}
}
node->node_risk = sum2 - (sum/n)*sum;
node->value = sum/n;
for( j = 0; j < cv_n; j++ )
{
double s = cv_sum[j], si = sum - s;
double s2 = cv_sum2[j], s2i = sum2 - s2;
int c = cv_count[j], ci = n - c;
double r = si/MAX(ci,1);
node->cv_node_risk[j] = s2i - r*r*ci;
node->cv_node_error[j] = s2 - 2*r*s + c*r*r;
node->cv_Tn[j] = INT_MAX;
}
}
}
void CvDTree::complete_node_dir( CvDTreeNode* node )
{
int vi, i, n = node->sample_count, nl, nr, d0 = 0, d1 = -1;
int nz = n - node->get_num_valid(node->split->var_idx);
char* dir = (char*)data->direction->data.ptr;
// try to complete direction using surrogate splits
if( nz && data->params.use_surrogates )
{
CvDTreeSplit* split = node->split->next;
for( ; split != 0 && nz; split = split->next )
{
int inversed_mask = split->inversed ? -1 : 0;
vi = split->var_idx;
if( data->get_var_type(vi) >= 0 ) // split on categorical var
{
const int* labels = data->get_cat_var_data(node, vi);
const int* subset = split->subset;
for( i = 0; i < n; i++ )
{
int idx;
if( !dir[i] && (idx = labels[i]) >= 0 )
{
int d = CV_DTREE_CAT_DIR(idx,subset);
dir[i] = (char)((d ^ inversed_mask) - inversed_mask);
if( --nz )
break;
}
}
}
else // split on ordered var
{
const CvPair32s32f* sorted = data->get_ord_var_data(node, vi);
int split_point = split->ord.split_point;
int n1 = node->get_num_valid(vi);
assert( 0 <= split_point && split_point < n-1 );
for( i = 0; i < n1; i++ )
{
int idx = sorted[i].i;
if( !dir[idx] )
{
int d = i <= split_point ? -1 : 1;
dir[idx] = (char)((d ^ inversed_mask) - inversed_mask);
if( --nz )
break;
}
}
}
}
}
// find the default direction for the rest
if( nz )
{
for( i = nr = 0; i < n; i++ )
nr += dir[i] > 0;
nl = n - nr - nz;
d0 = nl > nr ? -1 : nr > nl;
}
// make sure that every sample is directed either to the left or to the right
for( i = 0; i < n; i++ )
{
int d = dir[i];
if( !d )
{
d = d0;
if( !d )
d = d1, d1 = -d1;
}
d = d > 0;
dir[i] = (char)d; // remap (-1,1) to (0,1)
}
}
void CvDTree::split_node_data( CvDTreeNode* node )
{
int vi, i, n = node->sample_count, nl, nr;
char* dir = (char*)data->direction->data.ptr;
CvDTreeNode *left = 0, *right = 0;
int* new_idx = data->split_buf->data.i;
int new_buf_idx = data->get_child_buf_idx( node );
int work_var_count = data->get_work_var_count();
// speedup things a little, especially for tree ensembles with a lots of small trees:
// do not physically split the input data between the left and right child nodes
// when we are not going to split them further,
// as calc_node_value() does not requires input features anyway.
bool split_input_data;
complete_node_dir(node);
for( i = nl = nr = 0; i < n; i++ )
{
int d = dir[i];
// initialize new indices for splitting ordered variables
new_idx[i] = (nl & (d-1)) | (nr & -d); // d ? ri : li
nr += d;
nl += d^1;
}
node->left = left = data->new_node( node, nl, new_buf_idx, node->offset );
node->right = right = data->new_node( node, nr, new_buf_idx, node->offset +
(data->ord_var_count + work_var_count)*nl );
split_input_data = node->depth + 1 < data->params.max_depth &&
(node->left->sample_count > data->params.min_sample_count ||
node->right->sample_count > data->params.min_sample_count);
// split ordered variables, keep both halves sorted.
for( vi = 0; vi < data->var_count; vi++ )
{
int ci = data->get_var_type(vi);
int n1 = node->get_num_valid(vi);
CvPair32s32f *src, *ldst0, *rdst0, *ldst, *rdst;
CvPair32s32f tl, tr;
if( ci >= 0 || !split_input_data )
continue;
src = data->get_ord_var_data(node, vi);
ldst0 = ldst = data->get_ord_var_data(left, vi);
rdst0 = rdst = data->get_ord_var_data(right, vi);
tl = ldst0[nl]; tr = rdst0[nr];
// split sorted
for( i = 0; i < n1; i++ )
{
int idx = src[i].i;
float val = src[i].val;
int d = dir[idx];
idx = new_idx[idx];
ldst->i = rdst->i = idx;
ldst->val = rdst->val = val;
ldst += d^1;
rdst += d;
}
left->set_num_valid(vi, (int)(ldst - ldst0));
right->set_num_valid(vi, (int)(rdst - rdst0));
// split missing
for( ; i < n; i++ )
{
int idx = src[i].i;
int d = dir[idx];
idx = new_idx[idx];
ldst->i = rdst->i = idx;
ldst->val = rdst->val = ord_nan;
ldst += d^1;
rdst += d;
}
ldst0[nl] = tl; rdst0[nr] = tr;
}
// split categorical vars, responses and cv_labels using new_idx relocation table
for( vi = 0; vi < work_var_count; vi++ )
{
int ci = data->get_var_type(vi);
int n1 = node->get_num_valid(vi), nr1 = 0;
int *src, *ldst0, *rdst0, *ldst, *rdst;
int tl, tr;
if( ci < 0 || (vi < data->var_count && !split_input_data) )
continue;
src = data->get_cat_var_data(node, vi);
ldst0 = ldst = data->get_cat_var_data(left, vi);
rdst0 = rdst = data->get_cat_var_data(right, vi);
tl = ldst0[nl]; tr = rdst0[nr];
for( i = 0; i < n; i++ )
{
int d = dir[i];
int val = src[i];
*ldst = *rdst = val;
ldst += d^1;
rdst += d;
nr1 += (val >= 0)&d;
}
if( vi < data->var_count )
{
left->set_num_valid(vi, n1 - nr1);
right->set_num_valid(vi, nr1);
}
ldst0[nl] = tl; rdst0[nr] = tr;
}
// deallocate the parent node data that is not needed anymore
data->free_node_data(node);
}
void CvDTree::prune_cv()
{
CvMat* ab = 0;
CvMat* temp = 0;
CvMat* err_jk = 0;
// 1. build tree sequence for each cv fold, calculate error_{Tj,beta_k}.
// 2. choose the best tree index (if need, apply 1SE rule).
// 3. store the best index and cut the branches.
CV_FUNCNAME( "CvDTree::prune_cv" );
__BEGIN__;
int ti, j, tree_count = 0, cv_n = data->params.cv_folds, n = root->sample_count;
// currently, 1SE for regression is not implemented
bool use_1se = data->params.use_1se_rule != 0 && data->is_classifier;
double* err;
double min_err = 0, min_err_se = 0;
int min_idx = -1;
CV_CALL( ab = cvCreateMat( 1, 256, CV_64F ));
// build the main tree sequence, calculate alpha's
for(;;tree_count++)
{
double min_alpha = update_tree_rnc(tree_count, -1);
if( cut_tree(tree_count, -1, min_alpha) )
break;
if( ab->cols <= tree_count )
{
CV_CALL( temp = cvCreateMat( 1, ab->cols*3/2, CV_64F ));
for( ti = 0; ti < ab->cols; ti++ )
temp->data.db[ti] = ab->data.db[ti];
cvReleaseMat( &ab );
ab = temp;
temp = 0;
}
ab->data.db[tree_count] = min_alpha;
}
ab->data.db[0] = 0.;
if( tree_count > 0 )
{
for( ti = 1; ti < tree_count-1; ti++ )
ab->data.db[ti] = sqrt(ab->data.db[ti]*ab->data.db[ti+1]);
ab->data.db[tree_count-1] = DBL_MAX*0.5;
CV_CALL( err_jk = cvCreateMat( cv_n, tree_count, CV_64F ));
err = err_jk->data.db;
for( j = 0; j < cv_n; j++ )
{
int tj = 0, tk = 0;
for( ; tk < tree_count; tj++ )
{
double min_alpha = update_tree_rnc(tj, j);
if( cut_tree(tj, j, min_alpha) )
min_alpha = DBL_MAX;
for( ; tk < tree_count; tk++ )
{
if( ab->data.db[tk] > min_alpha )
break;
err[j*tree_count + tk] = root->tree_error;
}
}
}
for( ti = 0; ti < tree_count; ti++ )
{
double sum_err = 0;
for( j = 0; j < cv_n; j++ )
sum_err += err[j*tree_count + ti];
if( ti == 0 || sum_err < min_err )
{
min_err = sum_err;
min_idx = ti;
if( use_1se )
min_err_se = sqrt( sum_err*(n - sum_err) );
}
else if( sum_err < min_err + min_err_se )
min_idx = ti;
}
}
pruned_tree_idx = min_idx;
free_prune_data(data->params.truncate_pruned_tree != 0);
__END__;
cvReleaseMat( &err_jk );
cvReleaseMat( &ab );
cvReleaseMat( &temp );
}
double CvDTree::update_tree_rnc( int T, int fold )
{
CvDTreeNode* node = root;
double min_alpha = DBL_MAX;
for(;;)
{
CvDTreeNode* parent;
for(;;)
{
int t = fold >= 0 ? node->cv_Tn[fold] : node->Tn;
if( t <= T || !node->left )
{
node->complexity = 1;
node->tree_risk = node->node_risk;
node->tree_error = 0.;
if( fold >= 0 )
{
node->tree_risk = node->cv_node_risk[fold];
node->tree_error = node->cv_node_error[fold];
}
break;
}
node = node->left;
}
for( parent = node->parent; parent && parent->right == node;
node = parent, parent = parent->parent )
{
parent->complexity += node->complexity;
parent->tree_risk += node->tree_risk;
parent->tree_error += node->tree_error;
parent->alpha = ((fold >= 0 ? parent->cv_node_risk[fold] : parent->node_risk)
- parent->tree_risk)/(parent->complexity - 1);
min_alpha = MIN( min_alpha, parent->alpha );
}
if( !parent )
break;
parent->complexity = node->complexity;
parent->tree_risk = node->tree_risk;
parent->tree_error = node->tree_error;
node = parent->right;
}
return min_alpha;
}
int CvDTree::cut_tree( int T, int fold, double min_alpha )
{
CvDTreeNode* node = root;
if( !node->left )
return 1;
for(;;)
{
CvDTreeNode* parent;
for(;;)
{
int t = fold >= 0 ? node->cv_Tn[fold] : node->Tn;
if( t <= T || !node->left )
break;
if( node->alpha <= min_alpha + FLT_EPSILON )
{
if( fold >= 0 )
node->cv_Tn[fold] = T;
else
node->Tn = T;
if( node == root )
return 1;
break;
}
node = node->left;
}
for( parent = node->parent; parent && parent->right == node;
node = parent, parent = parent->parent )
;
if( !parent )
break;
node = parent->right;
}
return 0;
}
void CvDTree::free_prune_data(bool cut_tree)
{
CvDTreeNode* node = root;
for(;;)
{
CvDTreeNode* parent;
for(;;)
{
// do not call cvSetRemoveByPtr( cv_heap, node->cv_Tn )
// as we will clear the whole cross-validation heap at the end
node->cv_Tn = 0;
node->cv_node_error = node->cv_node_risk = 0;
if( !node->left )
break;
node = node->left;
}
for( parent = node->parent; parent && parent->right == node;
node = parent, parent = parent->parent )
{
if( cut_tree && parent->Tn <= pruned_tree_idx )
{
data->free_node( parent->left );
data->free_node( parent->right );
parent->left = parent->right = 0;
}
}
if( !parent )
break;
node = parent->right;
}
if( data->cv_heap )
cvClearSet( data->cv_heap );
}
void CvDTree::free_tree()
{
if( root && data && data->shared )
{
pruned_tree_idx = INT_MIN;
free_prune_data(true);
data->free_node(root);
root = 0;
}
}
CvDTreeNode* CvDTree::predict( const CvMat* _sample,
const CvMat* _missing, bool preprocessed_input ) const
{
CvDTreeNode* result = 0;
int* catbuf = 0;
CV_FUNCNAME( "CvDTree::predict" );
__BEGIN__;
int i, step, mstep = 0;
const float* sample;
const uchar* m = 0;
CvDTreeNode* node = root;
const int* vtype;
const int* vidx;
const int* cmap;
const int* cofs;
if( !node )
CV_ERROR( CV_StsError, "The tree has not been trained yet" );
if( !CV_IS_MAT(_sample) || CV_MAT_TYPE(_sample->type) != CV_32FC1 ||
_sample->cols != 1 && _sample->rows != 1 ||
_sample->cols + _sample->rows - 1 != data->var_all && !preprocessed_input ||
_sample->cols + _sample->rows - 1 != data->var_count && preprocessed_input )
CV_ERROR( CV_StsBadArg,
"the input sample must be 1d floating-point vector with the same "
"number of elements as the total number of variables used for training" );
sample = _sample->data.fl;
step = CV_IS_MAT_CONT(_sample->type) ? 1 : _sample->step/sizeof(sample[0]);
if( data->cat_count && !preprocessed_input ) // cache for categorical variables
{
int n = data->cat_count->cols;
catbuf = (int*)cvStackAlloc(n*sizeof(catbuf[0]));
for( i = 0; i < n; i++ )
catbuf[i] = -1;
}
if( _missing )
{
if( !CV_IS_MAT(_missing) || !CV_IS_MASK_ARR(_missing) ||
!CV_ARE_SIZES_EQ(_missing, _sample) )
CV_ERROR( CV_StsBadArg,
"the missing data mask must be 8-bit vector of the same size as input sample" );
m = _missing->data.ptr;
mstep = CV_IS_MAT_CONT(_missing->type) ? 1 : _missing->step/sizeof(m[0]);
}
vtype = data->var_type->data.i;
vidx = data->var_idx && !preprocessed_input ? data->var_idx->data.i : 0;
cmap = data->cat_map ? data->cat_map->data.i : 0;
cofs = data->cat_ofs ? data->cat_ofs->data.i : 0;
while( node->Tn > pruned_tree_idx && node->left )
{
CvDTreeSplit* split = node->split;
int dir = 0;
for( ; !dir && split != 0; split = split->next )
{
int vi = split->var_idx;
int ci = vtype[vi];
i = vidx ? vidx[vi] : vi;
float val = sample[i*step];
if( m && m[i*mstep] )
continue;
if( ci < 0 ) // ordered
dir = val <= split->ord.c ? -1 : 1;
else // categorical
{
int c;
if( preprocessed_input )
c = cvRound(val);
else
{
c = catbuf[ci];
if( c < 0 )
{
int a = c = cofs[ci];
int b = cofs[ci+1];
int ival = cvRound(val);
if( ival != val )
CV_ERROR( CV_StsBadArg,
"one of input categorical variable is not an integer" );
while( a < b )
{
c = (a + b) >> 1;
if( ival < cmap[c] )
b = c;
else if( ival > cmap[c] )
a = c+1;
else
break;
}
if( c < 0 || ival != cmap[c] )
continue;
catbuf[ci] = c -= cofs[ci];
}
}
dir = CV_DTREE_CAT_DIR(c, split->subset);
}
if( split->inversed )
dir = -dir;
}
if( !dir )
{
double diff = node->right->sample_count - node->left->sample_count;
dir = diff < 0 ? -1 : 1;
}
node = dir < 0 ? node->left : node->right;
}
result = node;
__END__;
return result;
}
const CvMat* CvDTree::get_var_importance()
{
if( !var_importance )
{
CvDTreeNode* node = root;
double* importance;
if( !node )
return 0;
var_importance = cvCreateMat( 1, data->var_count, CV_64F );
cvZero( var_importance );
importance = var_importance->data.db;
for(;;)
{
CvDTreeNode* parent;
for( ;; node = node->left )
{
CvDTreeSplit* split = node->split;
if( !node->left || node->Tn <= pruned_tree_idx )
break;
for( ; split != 0; split = split->next )
importance[split->var_idx] += split->quality;
}
for( parent = node->parent; parent && parent->right == node;
node = parent, parent = parent->parent )
;
if( !parent )
break;
node = parent->right;
}
cvNormalize( var_importance, var_importance, 1., 0, CV_L1 );
}
return var_importance;
}
void CvDTree::write_split( CvFileStorage* fs, CvDTreeSplit* split )
{
int ci;
cvStartWriteStruct( fs, 0, CV_NODE_MAP + CV_NODE_FLOW );
cvWriteInt( fs, "var", split->var_idx );
cvWriteReal( fs, "quality", split->quality );
ci = data->get_var_type(split->var_idx);
if( ci >= 0 ) // split on a categorical var
{
int i, n = data->cat_count->data.i[ci], to_right = 0, default_dir;
for( i = 0; i < n; i++ )
to_right += CV_DTREE_CAT_DIR(i,split->subset) > 0;
// ad-hoc rule when to use inverse categorical split notation
// to achieve more compact and clear representation
default_dir = to_right <= 1 || to_right <= MIN(3, n/2) || to_right <= n/3 ? -1 : 1;
cvStartWriteStruct( fs, default_dir*(split->inversed ? -1 : 1) > 0 ?
"in" : "not_in", CV_NODE_SEQ+CV_NODE_FLOW );
for( i = 0; i < n; i++ )
{
int dir = CV_DTREE_CAT_DIR(i,split->subset);
if( dir*default_dir < 0 )
cvWriteInt( fs, 0, i );
}
cvEndWriteStruct( fs );
}
else
cvWriteReal( fs, !split->inversed ? "le" : "gt", split->ord.c );
cvEndWriteStruct( fs );
}
void CvDTree::write_node( CvFileStorage* fs, CvDTreeNode* node )
{
CvDTreeSplit* split;
cvStartWriteStruct( fs, 0, CV_NODE_MAP );
cvWriteInt( fs, "depth", node->depth );
cvWriteInt( fs, "sample_count", node->sample_count );
cvWriteReal( fs, "value", node->value );
if( data->is_classifier )
cvWriteInt( fs, "norm_class_idx", node->class_idx );
cvWriteInt( fs, "Tn", node->Tn );
cvWriteInt( fs, "complexity", node->complexity );
cvWriteReal( fs, "alpha", node->alpha );
cvWriteReal( fs, "node_risk", node->node_risk );
cvWriteReal( fs, "tree_risk", node->tree_risk );
cvWriteReal( fs, "tree_error", node->tree_error );
if( node->left )
{
cvStartWriteStruct( fs, "splits", CV_NODE_SEQ );
for( split = node->split; split != 0; split = split->next )
write_split( fs, split );
cvEndWriteStruct( fs );
}
cvEndWriteStruct( fs );
}
void CvDTree::write_tree_nodes( CvFileStorage* fs )
{
//CV_FUNCNAME( "CvDTree::write_tree_nodes" );
__BEGIN__;
CvDTreeNode* node = root;
// traverse the tree and save all the nodes in depth-first order
for(;;)
{
CvDTreeNode* parent;
for(;;)
{
write_node( fs, node );
if( !node->left )
break;
node = node->left;
}
for( parent = node->parent; parent && parent->right == node;
node = parent, parent = parent->parent )
;
if( !parent )
break;
node = parent->right;
}
__END__;
}
void CvDTree::write( CvFileStorage* fs, const char* name )
{
//CV_FUNCNAME( "CvDTree::write" );
__BEGIN__;
cvStartWriteStruct( fs, name, CV_NODE_MAP, CV_TYPE_NAME_ML_TREE );
get_var_importance();
data->write_params( fs );
if( var_importance )
cvWrite( fs, "var_importance", var_importance );
write( fs );
cvEndWriteStruct( fs );
__END__;
}
void CvDTree::write( CvFileStorage* fs )
{
//CV_FUNCNAME( "CvDTree::write" );
__BEGIN__;
cvWriteInt( fs, "best_tree_idx", pruned_tree_idx );
cvStartWriteStruct( fs, "nodes", CV_NODE_SEQ );
write_tree_nodes( fs );
cvEndWriteStruct( fs );
__END__;
}
CvDTreeSplit* CvDTree::read_split( CvFileStorage* fs, CvFileNode* fnode )
{
CvDTreeSplit* split = 0;
CV_FUNCNAME( "CvDTree::read_split" );
__BEGIN__;
int vi, ci;
if( !fnode || CV_NODE_TYPE(fnode->tag) != CV_NODE_MAP )
CV_ERROR( CV_StsParseError, "some of the splits are not stored properly" );
vi = cvReadIntByName( fs, fnode, "var", -1 );
if( (unsigned)vi >= (unsigned)data->var_count )
CV_ERROR( CV_StsOutOfRange, "Split variable index is out of range" );
ci = data->get_var_type(vi);
if( ci >= 0 ) // split on categorical var
{
int i, n = data->cat_count->data.i[ci], inversed = 0, val;
CvSeqReader reader;
CvFileNode* inseq;
split = data->new_split_cat( vi, 0 );
inseq = cvGetFileNodeByName( fs, fnode, "in" );
if( !inseq )
{
inseq = cvGetFileNodeByName( fs, fnode, "not_in" );
inversed = 1;
}
if( !inseq ||
(CV_NODE_TYPE(inseq->tag) != CV_NODE_SEQ && CV_NODE_TYPE(inseq->tag) != CV_NODE_INT))
CV_ERROR( CV_StsParseError,
"Either 'in' or 'not_in' tags should be inside a categorical split data" );
if( CV_NODE_TYPE(inseq->tag) == CV_NODE_INT )
{
val = inseq->data.i;
if( (unsigned)val >= (unsigned)n )
CV_ERROR( CV_StsOutOfRange, "some of in/not_in elements are out of range" );
split->subset[val >> 5] |= 1 << (val & 31);
}
else
{
cvStartReadSeq( inseq->data.seq, &reader );
for( i = 0; i < reader.seq->total; i++ )
{
CvFileNode* inode = (CvFileNode*)reader.ptr;
val = inode->data.i;
if( CV_NODE_TYPE(inode->tag) != CV_NODE_INT || (unsigned)val >= (unsigned)n )
CV_ERROR( CV_StsOutOfRange, "some of in/not_in elements are out of range" );
split->subset[val >> 5] |= 1 << (val & 31);
CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
}
}
// for categorical splits we do not use inversed splits,
// instead we inverse the variable set in the split
if( inversed )
for( i = 0; i < (n + 31) >> 5; i++ )
split->subset[i] ^= -1;
}
else
{
CvFileNode* cmp_node;
split = data->new_split_ord( vi, 0, 0, 0, 0 );
cmp_node = cvGetFileNodeByName( fs, fnode, "le" );
if( !cmp_node )
{
cmp_node = cvGetFileNodeByName( fs, fnode, "gt" );
split->inversed = 1;
}
split->ord.c = (float)cvReadReal( cmp_node );
}
split->quality = (float)cvReadRealByName( fs, fnode, "quality" );
__END__;
return split;
}
CvDTreeNode* CvDTree::read_node( CvFileStorage* fs, CvFileNode* fnode, CvDTreeNode* parent )
{
CvDTreeNode* node = 0;
CV_FUNCNAME( "CvDTree::read_node" );
__BEGIN__;
CvFileNode* splits;
int i, depth;
if( !fnode || CV_NODE_TYPE(fnode->tag) != CV_NODE_MAP )
CV_ERROR( CV_StsParseError, "some of the tree elements are not stored properly" );
CV_CALL( node = data->new_node( parent, 0, 0, 0 ));
depth = cvReadIntByName( fs, fnode, "depth", -1 );
if( depth != node->depth )
CV_ERROR( CV_StsParseError, "incorrect node depth" );
node->sample_count = cvReadIntByName( fs, fnode, "sample_count" );
node->value = cvReadRealByName( fs, fnode, "value" );
if( data->is_classifier )
node->class_idx = cvReadIntByName( fs, fnode, "norm_class_idx" );
node->Tn = cvReadIntByName( fs, fnode, "Tn" );
node->complexity = cvReadIntByName( fs, fnode, "complexity" );
node->alpha = cvReadRealByName( fs, fnode, "alpha" );
node->node_risk = cvReadRealByName( fs, fnode, "node_risk" );
node->tree_risk = cvReadRealByName( fs, fnode, "tree_risk" );
node->tree_error = cvReadRealByName( fs, fnode, "tree_error" );
splits = cvGetFileNodeByName( fs, fnode, "splits" );
if( splits )
{
CvSeqReader reader;
CvDTreeSplit* last_split = 0;
if( CV_NODE_TYPE(splits->tag) != CV_NODE_SEQ )
CV_ERROR( CV_StsParseError, "splits tag must stored as a sequence" );
cvStartReadSeq( splits->data.seq, &reader );
for( i = 0; i < reader.seq->total; i++ )
{
CvDTreeSplit* split;
CV_CALL( split = read_split( fs, (CvFileNode*)reader.ptr ));
if( !last_split )
node->split = last_split = split;
else
last_split = last_split->next = split;
CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
}
}
__END__;
return node;
}
void CvDTree::read_tree_nodes( CvFileStorage* fs, CvFileNode* fnode )
{
CV_FUNCNAME( "CvDTree::read_tree_nodes" );
__BEGIN__;
CvSeqReader reader;
CvDTreeNode _root;
CvDTreeNode* parent = &_root;
int i;
parent->left = parent->right = parent->parent = 0;
cvStartReadSeq( fnode->data.seq, &reader );
for( i = 0; i < reader.seq->total; i++ )
{
CvDTreeNode* node;
CV_CALL( node = read_node( fs, (CvFileNode*)reader.ptr, parent != &_root ? parent : 0 ));
if( !parent->left )
parent->left = node;
else
parent->right = node;
if( node->split )
parent = node;
else
{
while( parent && parent->right )
parent = parent->parent;
}
CV_NEXT_SEQ_ELEM( reader.seq->elem_size, reader );
}
root = _root.left;
__END__;
}
void CvDTree::read( CvFileStorage* fs, CvFileNode* fnode )
{
CvDTreeTrainData* _data = new CvDTreeTrainData();
_data->read_params( fs, fnode );
read( fs, fnode, _data );
get_var_importance();
}
// a special entry point for reading weak decision trees from the tree ensembles
void CvDTree::read( CvFileStorage* fs, CvFileNode* node, CvDTreeTrainData* _data )
{
CV_FUNCNAME( "CvDTree::read" );
__BEGIN__;
CvFileNode* tree_nodes;
clear();
data = _data;
tree_nodes = cvGetFileNodeByName( fs, node, "nodes" );
if( !tree_nodes || CV_NODE_TYPE(tree_nodes->tag) != CV_NODE_SEQ )
CV_ERROR( CV_StsParseError, "nodes tag is missing" );
pruned_tree_idx = cvReadIntByName( fs, node, "best_tree_idx", -1 );
read_tree_nodes( fs, tree_nodes );
__END__;
}
/* End of file. */