mas01cr@509: extern "C" {
mas01cr@509: #include "audioDB_API.h"
mas01cr@509: }
mas01cr@509: #include "audioDB-internals.h"
mas01cr@589: #include "lshlib.h"
mas01cr@509: 
mas01cr@509: /*
mas01cr@509:  * Routines which are common to both indexed query and index creation:
mas01cr@509:  * we put them in their own file for build logistics.
mas01cr@509:  */
mas01cr@509: 
mas01cr@509: /* FIXME: there are several things wrong with this: the memory
mas01cr@509:  * discipline isn't ideal, the radius printing is a bit lame, the name
mas01cr@509:  * getting will succeed or fail depending on whether the path was
mas01cr@509:  * relative or absolute -- but most importantly encoding all that
mas01cr@509:  * information in a filename is going to lose: it's impossible to
mas01cr@509:  * maintain backwards-compatibility.  Instead we should probably store
mas01cr@509:  * the index metadata inside the audiodb instance. */
mas01cr@589: char *audiodb_index_get_name(const char *dbName, double radius, uint32_t sequenceLength) {
mas01cr@509:   char *indexName;
mas01cr@509:   if(strlen(dbName) > (ADB_MAXSTR - 32)) {
mas01cr@509:     return NULL;
mas01cr@509:   }
mas01cr@509:   indexName = new char[ADB_MAXSTR];  
mas01cr@509:   strncpy(indexName, dbName, ADB_MAXSTR);
mas01cr@509:   sprintf(indexName+strlen(dbName), ".lsh.%019.9f.%d", radius, sequenceLength);
mas01cr@509:   return indexName;
mas01cr@509: }
mas01cr@509: 
mas01cr@589: bool audiodb_index_exists(const char *dbName, double radius, uint32_t sequenceLength) {
mas01cr@509:   char *indexName = audiodb_index_get_name(dbName, radius, sequenceLength);
mas01cr@509:   if(!indexName) {
mas01cr@509:     return false;
mas01cr@509:   }
mas01cr@509:   struct stat st;
mas01cr@509:   if(stat(indexName, &st)) {
mas01cr@509:     delete [] indexName;
mas01cr@509:     return false;
mas01cr@509:   }
mas01cr@509:   /* FIXME: other stat checks here? */
mas01cr@509:   /* FIXME: is there any better way to check whether we can open a
mas01cr@509:    * file for reading than by opening a file for reading? */
mas01cr@509:   int fd = open(indexName, O_RDONLY);
mas01cr@509:   delete [] indexName;
mas01cr@509:   if(fd < 0) {
mas01cr@509:     return false;
mas01cr@509:   } else {
mas01cr@509:     close(fd);
mas01cr@509:     return true;
mas01cr@509:   }
mas01cr@509: }
mas01cr@509: 
mas01cr@509: /* FIXME: the indexName arg should be "const char *", but the LSH
mas01cr@509:  * library doesn't like that.
mas01cr@509:  */
mas01cr@509: LSH *audiodb_index_allocate(adb_t *adb, char *indexName, bool load_tables) {
mas01cr@509:   LSH *lsh;
mas01cr@509:   if(adb->cached_lsh) {
mas01cr@509:     if(!strncmp(adb->cached_lsh->get_indexName(), indexName, ADB_MAXSTR)) {
mas01cr@509:       return adb->cached_lsh;
mas01cr@509:     } else {
mas01cr@509:       delete adb->cached_lsh;
mas01cr@509:     }
mas01cr@509:   }
mas01cr@509:   lsh = new LSH(indexName, load_tables);
mas01cr@509:   if(load_tables) {
mas01cr@509:     adb->cached_lsh = lsh;
mas01cr@509:   } 
mas01cr@509:   return lsh;
mas01cr@509: }
mas01cr@509: 
mas01cr@589: vector<vector<float> > *audiodb_index_initialize_shingles(uint32_t sz, uint32_t dim, uint32_t seqLen) {
mas01cr@509:   std::vector<std::vector<float> > *vv = new vector<vector<float> >(sz);
mas01cr@589:   for(uint32_t i=0 ; i < sz ; i++) {
mas01cr@509:     (*vv)[i]=vector<float>(dim * seqLen);
mas01cr@509:   }
mas01cr@509:   return vv;
mas01cr@509: }
mas01cr@509: 
mas01cr@509: void audiodb_index_delete_shingles(vector<vector<float> > *vv) {
mas01cr@509:   delete vv;
mas01cr@509: }
mas01cr@509: 
mas01cr@589: void audiodb_index_make_shingle(vector<vector<float> >* vv, uint32_t idx, double* fvp, uint32_t dim, uint32_t seqLen){
mas01cr@589: 
mas01cr@509:   vector<float>::iterator ve = (*vv)[idx].end();
mas01cr@509:   vector<float>::iterator vi = (*vv)[idx].begin();
mas01cr@509:   // First feature vector in shingle
mas01cr@509:   if(idx == 0) {
mas01cr@509:     while(vi!=ve) {
mas01cr@509:       *vi++ = (float)(*fvp++);
mas01cr@509:     }
mas01cr@509:   } else {
mas01cr@509:     // Not first feature vector in shingle
mas01cr@509:     vector<float>::iterator ui=(*vv)[idx-1].begin() + dim;
mas01cr@509:     // Previous seqLen-1 dim-vectors
mas01cr@509:     while(vi!=ve-dim) {
mas01cr@509:       *vi++ = *ui++;
mas01cr@509:     }
mas01cr@509:     // Move data pointer to next feature vector
mas01cr@509:     fvp += ( seqLen + idx - 1 ) * dim ;
mas01cr@509:     // New d-vector
mas01cr@509:     while(vi!=ve) {
mas01cr@509:       *vi++ = (float)(*fvp++);
mas01cr@509:     }
mas01cr@509:   }
mas01cr@509: }
mas01cr@509: 
mas01cr@509: // in-place norming, no deletions.  If using power, return number of
mas01cr@509: // shingles above power threshold.
mas01cr@589: int audiodb_index_norm_shingles(vector<vector<float> >* vv, double* snp, double* spp, uint32_t dim, uint32_t seqLen, double radius, bool normed_vectors, bool use_pthreshold, float pthreshold) {
mas01cr@509:   int z = 0; // number of above-threshold shingles
mas01cr@509:   float l2norm;
mas01cr@509:   double power;
mas01cr@509:   float oneOverRadius = 1./(float)sqrt(radius); // Passed radius is really radius^2
mas01cr@509:   float oneOverSqrtl2NormDivRad = oneOverRadius;
mas01cr@589:   uint32_t shingleSize = seqLen * dim;
mas01cr@509: 
mas01cr@509:   if(!spp) {
mas01cr@509:     return -1;
mas01cr@509:   }
mas01cr@589:   for(uint32_t a=0; a<(*vv).size(); a++){
mas01cr@509:     l2norm = (float)(*snp++);
mas01cr@509:     if(normed_vectors)
mas01cr@509:       oneOverSqrtl2NormDivRad = (1./l2norm)*oneOverRadius;
mas01cr@509:     
mas01cr@589:     for(uint32_t b=0; b < shingleSize ; b++)
mas01cr@509:       (*vv)[a][b]*=oneOverSqrtl2NormDivRad;
mas01cr@509: 
mas01cr@509:     power = *spp++;
mas01cr@509:     if(use_pthreshold){
mas01cr@509:       if (power >= pthreshold)
mas01cr@509: 	z++;
mas01cr@509:     }
mas01cr@509:     else
mas01cr@509:       z++;	
mas01cr@509:   }
mas01cr@509:   return z;
mas01cr@509: }