基於OpenCV ML的Affinity Propagation原型實現

春節實現的APCluster，會在最近有空的時候和RBM一起嘗試提交到OpenCV ml中。這是一個原型代碼，缺少很多錯誤控制的流程，但已支持CvSparseMat。

mlapcluster.h

#ifndef GUARD_mlapcluster_h #define GUARD_mlapcluster_h #include <ml.h> #define CV_TYPE_NAME_ML_APCLUSTER "opencv-ml-affinity-propagation-cluster" struct CV_EXPORTS CvAPCParams { int maxiteration; int stopcriterion; double lambda; CvAPCParams() : maxiteration(2000), stopcriterion(200), lambda(0.9) {} CvAPCParams( int _maxiteration, int _stopcriterion, double _lambda ) : maxiteration(_maxiteration), stopcriterion(_stopcriterion), lambda(_lambda) {} }; class CV_EXPORTS CvAPCluster : public CvStatModel { private: CvAPCParams* params; public: CvAPCluster( CvAPCParams* _params ) : params(_params) {} virtual ~CvAPCluster() { clear(); } virtual bool train( const CvMat* _train_data, const CvMat* _response ); virtual bool train( const CvSparseMat* _train_data, const CvMat* _response ); virtual void clear(); virtual void write( CvFileStorage* fs, const char* name ); virtual void read( CvFileStorage* fs, CvFileNode* root_node ); }; #endif

mlapcluster.cpp

#include "mlapcluster.h" #include <limits> bool CvAPCluster::train( const CvMat* similarity, const CvMat* labels ) { int stop = 0; CV_FUNCNAME( "CvAPCluster::train" ); __BEGIN__; CvMat* rsp; CvMat* avl; int* rspmaxidx; double* rspmaxval; double* rspmaxval2; double* diagrsp; double& lam = params->lambda; double _lam = 1.-lam; int num; if ( !CV_IS_MAT(similarity) || similarity->rows != similarity->cols || CV_MAT_TYPE( similarity->type ) != CV_64FC1 ) CV_ERROR( CV_StsBadArg, "similarity matrix is a double-point matrix with equal rows and cols" ); num = similarity->rows; if ( !CV_IS_MAT(labels) || labels->cols != num || CV_MAT_TYPE( labels->type ) != CV_32SC1 ) CV_ERROR( CV_StsBadArg, "labels array (when passed) must be a valid 1d integer vector of <sample_count> elements" ); rsp = cvCreateMat( num, num, CV_64FC1 ); avl = cvCreateMat( num, num, CV_64FC1 ); cvZero( rsp ); cvZero( avl ); rspmaxidx = (int*)cvAlloc( num*sizeof(rspmaxidx[0]) ); rspmaxval = (double*)cvAlloc( num*sizeof(rspmaxval[0]) ); rspmaxval2 = (double*)cvAlloc( num*sizeof(rspmaxval2[0]) ); diagrsp = (double*)cvAlloc( num*sizeof(diagrsp[0]) ); for ( int k = 0; k < params->maxiteration; k++ ) { double* rsp_vec = rsp->data.db; double* avl_vec = avl->data.db; double* sim_vec = similarity->data.db; double* sim_vec2 = sim_vec; for ( int i = 0; i < num; i++ ) { int yk = -1; double y = -std::numeric_limits<double>::max(), y2 = -std::numeric_limits<double>::max(); for ( int j = 0; j < num; j++ ) { double t = *sim_vec2 + *avl_vec; if ( t > y ) { y = t; yk = j; } else if ( t > y2 ) y2 = t; sim_vec2++; avl_vec++; } for ( int j = 0; j < num; j++ ) { if ( j != yk ) *rsp_vec = *rsp_vec*lam+(*sim_vec-y)*_lam; else *rsp_vec = *rsp_vec*lam+(*sim_vec-y2)*_lam; rsp_vec++; sim_vec++; } } int* rspmaxidx_vec = rspmaxidx; double* rspmaxval_vec = rspmaxval; double* rspmaxval2_vec = rspmaxval2; for ( int i = 0; i < num; i++, rspmaxidx_vec++, rspmaxval_vec++, rspmaxval2_vec++ ) { *rspmaxidx_vec = -1; *rspmaxval_vec = 0; *rspmaxval2_vec = 0; } rsp_vec = rsp->data.db; double* diagrsp_vec = diagrsp; for ( int i = 0; i < num; i++, diagrsp_vec++ ) { rspmaxidx_vec = rspmaxidx; rspmaxval_vec = rspmaxval; rspmaxval2_vec = rspmaxval2; for ( int j = 0; j < num; j++, rspmaxidx_vec++, rspmaxval_vec++, rspmaxval2_vec++ ) { if ( i != j ) { if ( *rsp_vec > *rspmaxval_vec ) { *rspmaxidx_vec = i; *rspmaxval_vec = *rsp_vec; } else if ( *rsp_vec > *rspmaxval2_vec ) *rspmaxval2_vec = *rsp_vec; } else *diagrsp_vec = *rsp_vec; rsp_vec++; } } avl_vec = avl->data.db; diagrsp_vec = diagrsp; for ( int i = 0; i < num; i++, diagrsp_vec++ ) { rspmaxidx_vec = rspmaxidx; rspmaxval_vec = rspmaxval; rspmaxval2_vec = rspmaxval2; for ( int j = 0; j < num; j++, rspmaxidx_vec++, rspmaxval_vec++, rspmaxval2_vec++ ) { double tmp; if ( i != j ) { if ( i != *rspmaxidx_vec ) tmp = *diagrsp_vec + *rspmaxval_vec; else tmp = *diagrsp_vec + *rspmaxval2_vec; if ( tmp > 0 ) tmp = 0; } else tmp = *rspmaxval_vec; *avl_vec = *avl_vec*lam+tmp*_lam; avl_vec++; } } stop++; int* cls_vec = labels->data.i; rsp_vec = rsp->data.db; avl_vec = avl->data.db; for ( int i = 0; i < num; i++, cls_vec++ ) { int maxidx = i; double maxval = -std::numeric_limits<double>::max(); for ( int j = 0; j < num; j++ ) { double t = *rsp_vec + *avl_vec; if ( t > maxval ) { maxval = t; maxidx = j; } rsp_vec++; avl_vec++; } if ( *cls_vec != maxidx ) stop = 0; *cls_vec = maxidx; } if ( stop > params->stopcriterion ) break; } cvFree( &diagrsp ); cvFree( &rspmaxval2 ); cvFree( &rspmaxval ); cvFree( &rspmaxidx ); cvReleaseMat( &avl ); cvReleaseMat( &rsp ); __END__; return stop > params->stopcriterion; } struct CvSparseNode2D { int i; int k; double val; }; bool CvAPCluster::train( const CvSparseMat* similarity, const CvMat* labels ) { int stop = 0; CV_FUNCNAME( "CvAPCluster::train" ); __BEGIN__; int* psize; int* psize_vec; int num; double* rsp; double* avl; double* rsp_vec; double* avl_vec; CvSparseNode2D* nodes; CvSparseNode2D** node_entries; int* rspmaxidx; double* rspmaxval; double* rspmaxval2; double* diagrsp; int total = 0; double& lam = params->lambda; double _lam = 1.-lam; if ( !CV_IS_SPARSE_MAT(similarity) || similarity->dims != 2 || similarity->size[0] != similarity->size[1] || CV_MAT_TYPE( similarity->type ) != CV_64FC1 ) CV_ERROR( CV_StsBadArg, "similarity matrix is a double-point sparse matrix with equal rows and cols" ); num = similarity->size[0]; if ( !CV_IS_MAT(labels) || labels->cols != num || CV_MAT_TYPE( labels->type ) != CV_32SC1 ) CV_ERROR( CV_StsBadArg, "labels array (when passed) must be a valid 1d integer vector of <sample_count> elements" ); psize = (int*)cvAlloc( num*sizeof(psize[0]) ); psize_vec = psize; for ( int i = 0; i < num; i++, psize_vec++ ) *psize_vec = 0; CvSparseMatIterator mat_iterator; CvSparseNode* node; node = cvInitSparseMatIterator( similarity, &mat_iterator ); for ( ; node != 0; node = cvGetNextSparseNode( &mat_iterator ) ) { const int* idx = CV_NODE_IDX( similarity, node ); psize[idx[0]]++; total++; } rsp = (double*)cvAlloc( total*sizeof(rsp[0]) ); avl = (double*)cvAlloc( total*sizeof(avl[0]) ); rsp_vec = rsp; avl_vec = avl; nodes = (CvSparseNode2D*)cvAlloc( total*sizeof(nodes[0]) ); node_entries = (CvSparseNode2D**)cvAlloc( num*sizeof(node_entries[0]) ); psize_vec = psize; node_entries[0] = nodes; for ( int i = 1; i < num; i++, psize_vec++ ) node_entries[i] = node_entries[i-1] + *psize_vec; node = cvInitSparseMatIterator( similarity, &mat_iterator ); for ( ; node != 0; node = cvGetNextSparseNode( &mat_iterator ), rsp_vec++, avl_vec++ ) { *rsp_vec = *avl_vec = 0; const int* idx = CV_NODE_IDX( similarity, node ); CvSparseNode2D*& node_entry = node_entries[idx[0]]; node_entry->i = idx[0]; node_entry->k = idx[1]; node_entry->val = *(double*)CV_NODE_VAL( similarity, node ); node_entry++; } cvFree( &node_entries ); rspmaxidx = (int*)cvAlloc( num*sizeof(rspmaxidx[0]) ); rspmaxval = (double*)cvAlloc( num*sizeof(rspmaxval[0]) ); rspmaxval2 = (double*)cvAlloc( num*sizeof(rspmaxval2[0]) ); diagrsp = (double*)cvAlloc( num*sizeof(diagrsp[0]) ); for ( int k = 0; k < params->maxiteration; k++ ) { CvSparseNode2D* nodes_vec = nodes; psize_vec = psize; CvSparseNode2D* nodes_vec2 = nodes; avl_vec = avl; rsp_vec = rsp; for ( int i = 0; i < num; i++, psize_vec++ ) { int yk = -1; double y = -std::numeric_limits<double>::max(), y2 = -std::numeric_limits<double>::max(); for ( int j = 0; j < *psize_vec; j++ ) { double t = nodes_vec2->val + *avl_vec; if ( t > y ) { y = t; yk = nodes_vec2->k; } else if ( t > y2 ) y2 = t; nodes_vec2++; avl_vec++; } for ( int j = 0; j < *psize_vec; j++ ) { if ( nodes_vec->k != yk ) *rsp_vec = *rsp_vec*lam+(nodes_vec->val-y)*_lam; else *rsp_vec = *rsp_vec*lam+(nodes_vec->val-y2)*_lam; nodes_vec++; rsp_vec++; } } int* rspmaxidx_vec = rspmaxidx; double* rspmaxval_vec = rspmaxval; double* rspmaxval2_vec = rspmaxval2; for ( int i = 0; i < num; i++, rspmaxidx_vec++, rspmaxval_vec++, rspmaxval2_vec++ ) { *rspmaxidx_vec = -1; *rspmaxval_vec = 0; *rspmaxval2_vec = 0; } rsp_vec = rsp; psize_vec = psize; nodes_vec = nodes; double* diagrsp_vec = diagrsp; for ( int i = 0; i < num; i++, psize_vec++, diagrsp_vec++ ) for ( int j = 0; j < *psize_vec; j++ ) { if ( i != nodes_vec->k ) { if ( *rsp_vec > rspmaxval[nodes_vec->k] ) { rspmaxidx[nodes_vec->k] = i; rspmaxval[nodes_vec->k] = *rsp_vec; } else if ( *rsp_vec > rspmaxval2[nodes_vec->k] ) rspmaxval2[nodes_vec->k] = *rsp_vec; } else *diagrsp_vec = *rsp_vec; nodes_vec++; rsp_vec++; } avl_vec = avl; diagrsp_vec = diagrsp; psize_vec = psize; nodes_vec = nodes; for ( int i = 0; i < num; i++, psize_vec++, diagrsp_vec++ ) for ( int j = 0; j < *psize_vec; j++ ) { double tmp; if ( i != nodes_vec->k ) { if ( i != rspmaxidx[nodes_vec->k] ) tmp = *diagrsp_vec + rspmaxval[nodes_vec->k]; else tmp = *diagrsp_vec + rspmaxval2[nodes_vec->k]; if ( tmp > 0 ) tmp = 0; } else tmp = rspmaxval[nodes_vec->k]; *avl_vec = *avl_vec*lam+tmp*_lam; nodes_vec++; avl_vec++; } stop++; int* cls_vec = labels->data.i; rsp_vec = rsp; avl_vec = avl; psize_vec = psize; nodes_vec = nodes; for ( int i = 0; i < num; i++, psize_vec++, cls_vec++ ) { int maxidx = i; double maxval = -std::numeric_limits<double>::max(); for ( int j = 0; j < *psize_vec; j++ ) { double t = *rsp_vec + *avl_vec; if ( t > maxval ) { maxval = t; maxidx = nodes_vec->k; } nodes_vec++; rsp_vec++; avl_vec++; } if ( *cls_vec != maxidx ) stop = 0; *cls_vec = maxidx; } if ( stop > params->stopcriterion ) break; } cvFree( &diagrsp ); cvFree( &rspmaxval2 ); cvFree( &rspmaxval ); cvFree( &rspmaxidx ); cvFree( &rsp ); cvFree( &avl ); cvFree( &psize ); cvFree( &nodes ); __END__; return stop > params->stopcriterion; } void CvAPCluster::clear() { } void CvAPCluster::write( CvFileStorage* fs, const char* name ) { } void CvAPCluster::read( CvFileStorage* fs, CvFileNode* root_node ) { }

測試例子

#include "mlapcluster.h" #include <iostream> int main() { CvAPCParams* params = new CvAPCParams(2000, 200, 0.5); CvAPCluster* apcluster = new CvAPCluster(params); //CvMat* dist = cvCreateMat( 25, 25, CV_64FC1 ); int sizes[] = {25, 25}; CvSparseMat* dist = cvCreateSparseMat( 2, sizes, CV_64FC1 ); cvZero( dist ); freopen( "ToyProblemSimilarities.txt", "r", stdin ); double total = 0; for ( int k = 0; k < 25*24; k++ ) { int i, j; double d; std::cin>>i>>j>>d; cvSetReal2D( dist, i-1, j-1, d ); total += d; } total = -15.561256; //total = total / (25*24); printf("preference: %f/n", total); for ( int k = 0; k < 25; k++ ) cvSetReal2D( dist, k, k, total ); CvMat* response = cvCreateMat( 1, 25, CV_32SC1 ); long long t; t = cvGetTickCount(); apcluster->train( dist, response ); t = cvGetTickCount() - t; printf( "apcluster trained in %lld ms./n", t/1000000 ); for ( int i = 0; i < 25; i++ ) printf("%d ", response->data.i[i]); printf("/n"); return 0; }