annotate index.cpp @ 320:a995e5ad999a large_adb

working LARGE_ADB support. Activiate at creation time with -N --ntracks 20001 or greater, or with --datasize 1356 or greater, or both. LARGE_ADB blocks non-indexed QUERY.
author mas01mc
date Wed, 20 Aug 2008 13:50:58 +0000
parents b9eff6896943
children da2272e029b3
rev   line source
mas01mc@292 1 // LSH indexing
mas01mc@292 2 //
mas01mc@292 3 // Construct a persistent LSH table structure
mas01mc@292 4 // Store at the same location as dbName
mas01mc@292 5 // Naming convention:
mas01mc@292 6 // dbName.lsh.${radius}.${sequenceLength}
mas01mc@292 7 //
mas01mc@292 8 //
mas01mc@292 9 // Author: Michael Casey
mas01mc@292 10 // Date: 23 June 2008
mas01mc@319 11 //
mas01mc@319 12 // 19th August 2008 - added O2_FLAG_LARGE_ADB support
mas01mc@292 13
mas01mc@292 14 #include "audioDB.h"
mas01mc@292 15 #include "ReporterBase.h"
mas01mc@292 16
mas01mc@292 17
mas01mc@292 18 /************************* LSH point index to audioDB conversion *****************/
mas01mc@319 19 Uns32T audioDB::index_to_trackID(Uns32T lshID, Uns32T nPntBits){
mas01mc@319 20 assert(nPntBits);
mas01mc@319 21 return lshID>>nPntBits;
mas01mc@292 22 }
mas01mc@292 23
mas01mc@319 24 Uns32T audioDB::index_to_trackPos(Uns32T lshID, Uns32T nPntBits){
mas01mc@319 25 assert(nPntBits);
mas01mc@319 26 return lshID&((1<<nPntBits)-1);
mas01mc@292 27 }
mas01mc@292 28
mas01mc@319 29 Uns32T audioDB::index_from_trackInfo(Uns32T trackID, Uns32T spos, Uns32T nPntBits){
mas01mc@319 30 assert(nPntBits);
mas01mc@319 31 return (trackID << nPntBits) | spos;
mas01mc@292 32 }
mas01mc@292 33
mas01mc@292 34 /************************* LSH indexing and query initialization *****************/
mas01mc@292 35
mas01mc@292 36 char* audioDB::index_get_name(const char*dbName, double radius, Uns32T sequenceLength){
mas01mc@292 37 char* indexName = new char[MAXSTR];
mas01mc@292 38 // Attempt to make new file
mas01mc@292 39 if(strlen(dbName) > (MAXSTR - 32))
mas01mc@292 40 error("dbName is too long for LSH index filename appendages");
mas01mc@292 41 strncpy(indexName, dbName, MAXSTR);
mas01mc@292 42 sprintf(indexName+strlen(dbName), ".lsh.%019.9f.%d", radius, sequenceLength);
mas01mc@292 43 return indexName;
mas01mc@292 44 }
mas01mc@292 45
mas01mc@292 46 // return true if index exists else return false
mas01mc@292 47 int audioDB::index_exists(const char* dbName, double radius, Uns32T sequenceLength){
mas01mc@292 48 // Test to see if file exists
mas01mc@292 49 char* indexName = index_get_name(dbName, radius, sequenceLength);
mas01mc@292 50 lshfid = open (indexName, O_RDONLY);
mas01mc@292 51 delete[] indexName;
mas01mc@292 52 close(lshfid);
mas01mc@292 53
mas01mc@292 54 if(lshfid<0)
mas01mc@292 55 return false;
mas01mc@292 56 else
mas01mc@292 57 return true;
mas01mc@292 58 }
mas01mc@292 59
mas01mc@308 60 LSH* audioDB::index_allocate(char* indexName, bool load_hashTables){
mas01mc@308 61 LSH* gIndx=SERVER_LSH_INDEX_SINGLETON;
mas01mc@308 62 if(isServer && gIndx && (strncmp(gIndx->get_indexName(), indexName, MAXSTR)==0) )
mas01mc@308 63 audioDB::lsh = gIndx; // Use the global SERVER resident index
mas01mc@308 64 else{
mas01mc@308 65 if(audioDB::lsh)
mas01mc@308 66 delete audioDB::lsh;
mas01mc@308 67 audioDB::lsh = new LSH(indexName, load_hashTables);
mas01mc@308 68 }
mas01mc@308 69 assert(audioDB::lsh);
mas01mc@308 70 return audioDB::lsh;
mas01mc@308 71 }
mas01mc@308 72
mas01mc@292 73 vector<vector<float> >* audioDB::index_initialize_shingles(Uns32T sz){
mas01mc@292 74 if(vv)
mas01mc@292 75 delete vv;
mas01mc@292 76 vv = new vector<vector<float> >(sz);
mas01mc@292 77 for(Uns32T i=0 ; i < sz ; i++)
mas01mc@292 78 (*vv)[i]=vector<float>(dbH->dim*sequenceLength); // allocate shingle storage
mas01mc@292 79 return vv;
mas01mc@292 80 }
mas01mc@292 81
mas01mc@292 82 /******************** LSH indexing audioDB database access forall s \in {S} ***********************/
mas01mc@292 83
mas01mc@292 84 // Prepare the AudioDB database for read access and allocate auxillary memory
mas01mc@292 85 void audioDB::index_initialize(double **snp, double **vsnp, double **spp, double **vspp, Uns32T *dvp) {
mas01mc@319 86 if (!(dbH->flags & O2_FLAG_POWER)) {
mas01mc@319 87 error("INDEXed database must be power-enabled", dbName);
mas01mc@319 88 }
mas01mc@319 89
mas01mc@319 90 double *snpp = *snp, *sppp = 0;
mas01mc@319 91
mas01mc@292 92 *dvp = dbH->length / (dbH->dim * sizeof(double)); // number of database vectors
mas01mc@292 93 *snp = new double[*dvp]; // songs norm pointer: L2 norm table for each vector
mas01mc@292 94 *spp = new double[*dvp]; // song powertable pointer
mas01mc@292 95 sppp = *spp;
mas01mc@319 96 memcpy(*snp, l2normTable, *dvp * sizeof(double));
mas01mc@292 97 memcpy(*spp, powerTable, *dvp * sizeof(double));
mas01mc@319 98
mas01mc@319 99
mas01mc@292 100 for(Uns32T i = 0; i < dbH->numFiles; i++){
mas01mc@292 101 if(trackTable[i] >= sequenceLength) {
mas01mc@292 102 sequence_sum(snpp, trackTable[i], sequenceLength);
mas01mc@292 103 sequence_sqrt(snpp, trackTable[i], sequenceLength);
mas01mc@292 104
mas01mc@292 105 sequence_sum(sppp, trackTable[i], sequenceLength);
mas01mc@292 106 sequence_average(sppp, trackTable[i], sequenceLength);
mas01mc@292 107 }
mas01mc@292 108 snpp += trackTable[i];
mas01mc@292 109 sppp += trackTable[i];
mas01mc@292 110 }
mas01mc@319 111
mas01mc@292 112 *vsnp = *snp;
mas01mc@292 113 *vspp = *spp;
mas01mc@319 114
mas01mc@292 115 // Move the feature vector read pointer to start of fetures in database
mas01mc@292 116 lseek(dbfid, dbH->dataOffset, SEEK_SET);
mas01mc@292 117 }
mas01mc@292 118
mas01mc@292 119
mas01mc@292 120 /************************ LSH indexing ***********************************/
mas01mc@292 121 void audioDB::index_index_db(const char* dbName){
mas01mc@292 122 char* newIndexName;
mas01mc@292 123 double *fvp = 0, *sNorm = 0, *snPtr = 0, *sPower = 0, *spPtr = 0;
mas01mc@292 124 Uns32T dbVectors = 0;
mas01mc@292 125
mas01mc@320 126
mas01mc@292 127 printf("INDEX: initializing header\n");
mas01mc@292 128 // Check if audioDB exists, initialize header and open database for read
mas01mc@292 129 forWrite = false;
mas01mc@292 130 initDBHeader(dbName);
mas01mc@292 131
mas01mc@320 132 if(dbH->flags & O2_FLAG_POWER)
mas01mc@320 133 usingPower = true;
mas01mc@320 134
mas01mc@320 135 if(dbH->flags & O2_FLAG_TIMES)
mas01mc@320 136 usingTimes = true;
mas01mc@320 137
mas01mc@292 138 newIndexName = index_get_name(dbName, radius, sequenceLength);
mas01mc@292 139
mas01mc@292 140 // Set unit norming flag override
mas01mc@292 141 audioDB::normalizedDistance = !audioDB::no_unit_norming;
mas01mc@292 142
mas01mc@320 143 printf("INDEX: dim %d\n", (int)dbH->dim);
mas01mc@292 144 printf("INDEX: R %f\n", radius);
mas01mc@292 145 printf("INDEX: seqlen %d\n", sequenceLength);
mas01mc@292 146 printf("INDEX: lsh_w %f\n", lsh_param_w);
mas01mc@292 147 printf("INDEX: lsh_k %d\n", lsh_param_k);
mas01mc@292 148 printf("INDEX: lsh_m %d\n", lsh_param_m);
mas01mc@292 149 printf("INDEX: lsh_N %d\n", lsh_param_N);
mas01mc@296 150 printf("INDEX: lsh_C %d\n", lsh_param_ncols);
mas01mc@292 151 printf("INDEX: lsh_b %d\n", lsh_param_b);
mas01mc@292 152 printf("INDEX: normalized? %s\n", normalizedDistance?"true":"false");
mas01mc@292 153 fflush(stdout);
mas01mc@292 154
mas01mc@292 155
mas01mc@292 156 if((lshfid = open(newIndexName,O_RDONLY))<0){
mas01mc@292 157 printf("INDEX: constructing new LSH index\n");
mas01mc@292 158 printf("INDEX: making index file %s\n", newIndexName);
mas01mc@292 159 fflush(stdout);
mas01mc@292 160 // Construct new LSH index
mas01mc@292 161 lsh = new LSH((float)lsh_param_w, lsh_param_k,
mas01mc@292 162 lsh_param_m,
mas01mc@292 163 (Uns32T)(sequenceLength*dbH->dim),
mas01mc@292 164 lsh_param_N,
mas01mc@292 165 lsh_param_ncols,
mas01mc@292 166 (float)radius);
mas01mc@292 167 assert(lsh);
mas01mc@292 168
mas01mc@292 169 Uns32T endTrack = lsh_param_b;
mas01mc@292 170 if( endTrack > dbH->numFiles)
mas01mc@292 171 endTrack = dbH->numFiles;
mas01mc@292 172 // Insert up to lsh_param_b tracks
mas01mc@320 173 if( ! (dbH->flags & O2_FLAG_LARGE_ADB) ){
mas01mc@320 174 index_initialize(&sNorm, &snPtr, &sPower, &spPtr, &dbVectors);
mas01mc@319 175 }
mas01mc@320 176 index_insert_tracks(0, endTrack, &fvp, &sNorm, &snPtr, &sPower, &spPtr);
mas01mc@292 177 lsh->serialize(newIndexName, lsh_in_core?O2_SERIAL_FILEFORMAT2:O2_SERIAL_FILEFORMAT1);
mas01mc@292 178
mas01mc@292 179 // Clean up
mas01mc@292 180 delete lsh;
mas01mc@308 181 lsh = 0;
mas01mc@292 182 close(lshfid);
mas01mc@292 183 }
mas01mc@292 184
mas01mc@292 185 // Attempt to open LSH file
mas01mc@292 186 if((lshfid = open(newIndexName,O_RDONLY))>0){
mas01mc@292 187 printf("INDEX: merging with existing LSH index\n");
mas01mc@292 188 fflush(stdout);
mas01mc@292 189
mas01mc@292 190 // Get the lsh header info and find how many tracks are inserted already
mas01mc@292 191 lsh = new LSH(newIndexName, false); // lshInCore=false to avoid loading hashTables here
mas01mc@292 192 assert(lsh);
mas01mc@319 193 Uns32T maxs = index_to_trackID(lsh->get_maxp(), lsh_n_point_bits)+1;
mas01mc@292 194 delete lsh;
mas01mc@308 195 lsh = 0;
mas01mc@292 196
mas01mc@292 197 // This allows for updating index after more tracks are inserted into audioDB
mas01mc@292 198 for(Uns32T startTrack = maxs; startTrack < dbH->numFiles; startTrack+=lsh_param_b){
mas01mc@292 199
mas01mc@292 200 Uns32T endTrack = startTrack + lsh_param_b;
mas01mc@292 201 if( endTrack > dbH->numFiles)
mas01mc@292 202 endTrack = dbH->numFiles;
mas01mc@292 203 printf("Indexing track range: %d - %d\n", startTrack, endTrack);
mas01mc@292 204 fflush(stdout);
mas01mc@292 205 lsh = new LSH(newIndexName, lsh_in_core); // Initialize core memory for LSH tables
mas01mc@292 206 assert(lsh);
mas01mc@292 207
mas01mc@292 208 // Insert up to lsh_param_b database tracks
mas01mc@292 209 index_insert_tracks(startTrack, endTrack, &fvp, &sNorm, &snPtr, &sPower, &spPtr);
mas01mc@292 210
mas01mc@292 211 // Serialize to file
mas01mc@292 212 lsh->serialize(newIndexName, lsh_in_core?O2_SERIAL_FILEFORMAT2:O2_SERIAL_FILEFORMAT1); // Serialize core LSH heap to disk
mas01mc@292 213 delete lsh;
mas01mc@308 214 lsh = 0;
mas01mc@292 215 }
mas01mc@292 216
mas01mc@292 217 close(lshfid);
mas01mc@292 218 printf("INDEX: done constructing LSH index.\n");
mas01mc@292 219 fflush(stdout);
mas01mc@292 220
mas01mc@292 221 }
mas01mc@292 222 else{
mas01mc@292 223 error("Something's wrong with LSH index file");
mas01mc@292 224 exit(1);
mas01mc@292 225 }
mas01mc@292 226
mas01mc@292 227 delete[] newIndexName;
mas01mc@320 228 delete[] sNorm;
mas01mc@320 229 delete[] sPower;
mas01mc@292 230 }
mas01mc@292 231
mas01mc@319 232
mas01mc@319 233 // initialize auxillary track data from filesystem
mas01mc@319 234 // pre-conditions:
mas01mc@319 235 // dbH->flags & O2_FLAG_LARGE_ADB
mas01mc@319 236 // feature data allocated and copied (fvp)
mas01mc@319 237 //
mas01mc@319 238 // post-conditions:
mas01mc@319 239 // allocated power data
mas01mc@319 240 // allocated l2norm data
mas01mc@319 241 //
mas01mc@319 242 void audioDB::init_track_aux_data(Uns32T trackID, double* fvp, double** sNormpp,double** snPtrp, double** sPowerp, double** spPtrp){
mas01mc@319 243 if( !(dbH->flags & O2_FLAG_LARGE_ADB) )
mas01mc@319 244 error("error: init_track_large_adb required O2_FLAG_LARGE_ADB");
mas01mc@319 245
mas01mc@319 246 // Allocate and read the power sequence
mas01mc@319 247 if(trackTable[trackID]>=sequenceLength){
mas01mc@319 248
mas01mc@319 249 // Open and check dimensions of power file
mas01mc@319 250 powerfd = open(powerFileNameTable+trackID*O2_FILETABLE_ENTRY_SIZE, O_RDONLY);
mas01mc@319 251 if (powerfd < 0) {
mas01mc@319 252 error("failed to open power file", powerFileNameTable+trackID*O2_FILETABLE_ENTRY_SIZE);
mas01mc@319 253 }
mas01mc@319 254 if (fstat(powerfd, &statbuf) < 0) {
mas01mc@319 255 error("fstat error finding size of power file", powerFileNameTable+trackID*O2_FILETABLE_ENTRY_SIZE, "fstat");
mas01mc@319 256 }
mas01mc@319 257
mas01mc@319 258 if( (statbuf.st_size - sizeof(int)) / (sizeof(double)) != trackTable[trackID] )
mas01mc@319 259 error("Dimension mismatch: numPowers != numVectors", powerFileNameTable+trackID*O2_FILETABLE_ENTRY_SIZE);
mas01mc@320 260
mas01mc@319 261 *sPowerp = new double[trackTable[trackID]]; // Allocate memory for power values
mas01mc@319 262 assert(*sPowerp);
mas01mc@319 263 *spPtrp = *sPowerp;
mas01mc@319 264 insertPowerData(trackTable[trackID], powerfd, *sPowerp);
mas01mc@319 265 if (0 < powerfd) {
mas01mc@319 266 close(powerfd);
mas01mc@319 267 }
mas01mc@319 268
mas01mc@319 269 sequence_sum(*sPowerp, trackTable[trackID], sequenceLength);
mas01mc@319 270 sequence_average(*sPowerp, trackTable[trackID], sequenceLength);
mas01mc@319 271 powerTable = 0;
mas01mc@319 272
mas01mc@319 273 // Allocate and calculate the l2norm sequence
mas01mc@319 274 *sNormpp = new double[trackTable[trackID]];
mas01mc@319 275 assert(*sNormpp);
mas01mc@319 276 *snPtrp = *sNormpp;
mas01mc@319 277 unitNorm(fvp, dbH->dim, trackTable[trackID], *sNormpp);
mas01mc@319 278 sequence_sum(*sNormpp, trackTable[trackID], sequenceLength);
mas01mc@319 279 sequence_sqrt(*sNormpp, trackTable[trackID], sequenceLength);
mas01mc@319 280 }
mas01mc@319 281 }
mas01mc@319 282
mas01mc@292 283 void audioDB::index_insert_tracks(Uns32T start_track, Uns32T end_track,
mas01mc@292 284 double** fvpp, double** sNormpp,double** snPtrp,
mas01mc@292 285 double** sPowerp, double** spPtrp){
mas01mc@292 286 size_t nfv = 0;
mas01mc@292 287 double* fvp = 0; // Keep pointer for memory allocation and free() for track data
mas01mc@292 288 Uns32T trackID = 0;
mas01mc@292 289
mas01mc@292 290 VERB_LOG(1, "indexing tracks...");
mas01mc@292 291
mas01mc@319 292 int trackfd = dbfid;
mas01mc@292 293 for(trackID = start_track ; trackID < end_track ; trackID++ ){
mas01mc@319 294 if( dbH->flags & O2_FLAG_LARGE_ADB ){
mas01mc@319 295 // Open and check dimensions of feature file
mas01mc@319 296 initInputFile(featureFileNameTable+trackID*O2_FILETABLE_ENTRY_SIZE, false); // nommap, file pointer at correct position
mas01mc@319 297 trackfd = infid;
mas01mc@319 298 }
mas01mc@319 299 read_data(trackfd, trackID, &fvp, &nfv); // over-writes fvp and nfv
mas01mc@292 300 *fvpp = fvp; // Protect memory allocation and free() for track data
mas01mc@319 301
mas01mc@319 302 if( dbH->flags & O2_FLAG_LARGE_ADB )
mas01mc@319 303 // Load power and calculate power and l2norm sequence sums
mas01mc@319 304 init_track_aux_data(trackID, fvp, sNormpp, snPtrp, sPowerp, spPtrp);
mas01mc@319 305
mas01mc@292 306 if(!index_insert_track(trackID, fvpp, snPtrp, spPtrp))
mas01mc@292 307 break;
mas01mc@319 308 if ( dbH->flags & O2_FLAG_LARGE_ADB ){
mas01mc@319 309 close(infid);
mas01mc@320 310 delete[] *sNormpp;
mas01mc@320 311 delete[] *sPowerp;
mas01mc@319 312 *sNormpp = *sPowerp = *snPtrp = *snPtrp = 0;
mas01mc@319 313 }
mas01mc@319 314 } // end for(trackID = start_track ; ... )
mas01mc@292 315 std::cout << "finished inserting." << endl;
mas01mc@292 316 }
mas01mc@292 317
mas01mc@292 318 int audioDB::index_insert_track(Uns32T trackID, double** fvpp, double** snpp, double** sppp){
mas01mc@292 319 // Loop over the current input track's vectors
mas01cr@305 320 Uns32T numVecs = 0;
mas01cr@305 321 if (trackTable[trackID] > O2_MAXTRACKLEN) {
mas01cr@305 322 if (O2_MAXTRACKLEN < sequenceLength - 1) {
mas01cr@305 323 numVecs = 0;
mas01cr@305 324 } else {
mas01cr@305 325 numVecs = O2_MAXTRACKLEN - sequenceLength + 1;
mas01cr@305 326 }
mas01cr@305 327 } else {
mas01cr@305 328 if (trackTable[trackID] < sequenceLength - 1) {
mas01cr@305 329 numVecs = 0;
mas01cr@305 330 } else {
mas01cr@305 331 numVecs = trackTable[trackID] - sequenceLength + 1;
mas01cr@305 332 }
mas01cr@305 333 }
mas01mc@292 334
mas01mc@319 335 Uns32T numVecsAboveThreshold = 0, collisionCount = 0;
mas01mc@319 336 if(numVecs){
mas01mc@319 337 vv = index_initialize_shingles(numVecs);
mas01mc@319 338
mas01mc@319 339 for( Uns32T pointID = 0 ; pointID < numVecs; pointID++ )
mas01mc@319 340 index_make_shingle(vv, pointID, *fvpp, dbH->dim, sequenceLength);
mas01mc@319 341
mas01mc@319 342 numVecsAboveThreshold = index_norm_shingles(vv, *snpp, *sppp);
mas01mc@319 343 collisionCount = index_insert_shingles(vv, trackID, *sppp);
mas01mc@319 344 }
mas01mc@292 345 float meanCollisionCount = numVecsAboveThreshold?(float)collisionCount/numVecsAboveThreshold:0;
mas01mc@292 346
mas01mc@292 347 /* index_norm_shingles() only goes as far as the end of the
mas01mc@292 348 sequence, which is right, but the space allocated is for the
mas01mc@292 349 whole track. */
mas01mc@292 350
mas01mc@292 351 /* But numVecs will be <trackTable[track] if trackTable[track]>O2_MAXTRACKLEN
mas01mc@292 352 * So let's be certain the pointers are in the correct place
mas01mc@292 353 */
mas01mc@292 354
mas01mc@319 355 if( !(dbH->flags & O2_FLAG_LARGE_ADB) ){
mas01mc@319 356 *snpp += trackTable[trackID];
mas01mc@319 357 *sppp += trackTable[trackID];
mas01mc@319 358 *fvpp += trackTable[trackID] * dbH->dim;
mas01mc@319 359 }
mas01mc@292 360
mas01mc@292 361 std::cout << " n=" << trackTable[trackID] << " n'=" << numVecsAboveThreshold << " E[#c]=" << lsh->get_mean_collision_rate() << " E[#p]=" << meanCollisionCount << endl;
mas01mc@292 362 std::cout.flush();
mas01mc@292 363 return true;
mas01mc@292 364 }
mas01mc@292 365
mas01mc@292 366 Uns32T audioDB::index_insert_shingles(vector<vector<float> >* vv, Uns32T trackID, double* spp){
mas01mc@292 367 Uns32T collisionCount = 0;
mas01mc@292 368 cout << "[" << trackID << "]" << fileTable+trackID*O2_FILETABLE_ENTRY_SIZE;
mas01mc@311 369 for( Uns32T pointID=0 ; pointID < (*vv).size(); pointID+=sequenceHop)
mas01mc@311 370 if(!use_absolute_threshold || (use_absolute_threshold && (*spp >= absolute_threshold))){
mas01mc@319 371 collisionCount += lsh->insert_point((*vv)[pointID], index_from_trackInfo(trackID, pointID, lsh_n_point_bits));
mas01mc@311 372 spp+=sequenceHop;
mas01mc@311 373 }
mas01mc@292 374 return collisionCount;
mas01mc@292 375 }
mas01mc@292 376
mas01mc@292 377 /********************* LSH shingle construction ***************************/
mas01mc@292 378
mas01mc@292 379 // Construct shingles out of a feature matrix
mas01mc@292 380 // inputs:
mas01mc@292 381 // idx is vector index in feature matrix
mas01mc@292 382 // fvp is base feature matrix pointer double* [numVecs x dbH->dim]
mas01mc@292 383 //
mas01mc@292 384 // pre-conditions:
mas01mc@292 385 // dbH->dim
mas01mc@292 386 // sequenceLength
mas01mc@292 387 // idx < numVectors - sequenceLength + 1
mas01mc@292 388 //
mas01mc@292 389 // post-conditions:
mas01mc@292 390 // (*vv)[idx] contains a shingle with dbH->dim*sequenceLength float values
mas01mc@292 391
mas01mc@292 392 void audioDB::index_make_shingle(vector<vector<float> >* vv, Uns32T idx, double* fvp, Uns32T dim, Uns32T seqLen){
mas01mc@292 393 assert(idx<(*vv).size());
mas01mc@292 394 vector<float>::iterator ve = (*vv)[idx].end();
mas01mc@292 395 vi=(*vv)[idx].begin(); // shingle iterator
mas01mc@292 396 // First feature vector in shingle
mas01mc@292 397 if(idx==0){
mas01mc@292 398 while(vi!=ve)
mas01mc@292 399 *vi++ = (float)(*fvp++);
mas01mc@292 400 }
mas01mc@292 401 // Not first feature vector in shingle
mas01mc@292 402 else{
mas01mc@292 403 vector<float>::iterator ui=(*vv)[idx-1].begin() + dim; // previous shingle iterator
mas01mc@292 404 // Previous seqLen-1 dim-vectors
mas01mc@292 405 while(vi!=ve-dim)
mas01mc@292 406 *vi++=*ui++;
mas01mc@292 407 // Move data pointer to next feature vector
mas01mc@292 408 fvp += ( seqLen + idx - 1 ) * dim ;
mas01mc@292 409 // New d-vector
mas01mc@292 410 while(vi!=ve)
mas01mc@292 411 *vi++ = (float)(*fvp++);
mas01mc@292 412 }
mas01mc@292 413 }
mas01mc@292 414
mas01mc@292 415 // norm shingles
mas01mc@292 416 // in-place norming, no deletions
mas01mc@292 417 // If using power, return number of shingles above power threshold
mas01mc@292 418 int audioDB::index_norm_shingles(vector<vector<float> >* vv, double* snp, double* spp){
mas01mc@292 419 int z = 0; // number of above-threshold shingles
mas01mc@292 420 float l2norm;
mas01mc@292 421 double power;
mas01mc@292 422 float oneOverRadius = 1./(float)sqrt(radius); // Passed radius is really radius^2
mas01mc@292 423 float oneOverSqrtl2NormDivRad = oneOverRadius;
mas01mc@292 424 if(!spp)
mas01mc@292 425 error("LSH indexing and query requires a power feature using -w or -W");
mas01mc@292 426 Uns32T shingleSize = sequenceLength*dbH->dim;
mas01mc@292 427 for(Uns32T a=0; a<(*vv).size(); a++){
mas01mc@292 428 l2norm = (float)(*snp++);
mas01mc@292 429 if(audioDB::normalizedDistance)
mas01mc@292 430 oneOverSqrtl2NormDivRad = (1./l2norm)*oneOverRadius;
mas01mc@292 431
mas01mc@292 432 for(Uns32T b=0; b < shingleSize ; b++)
mas01mc@292 433 (*vv)[a][b]*=oneOverSqrtl2NormDivRad;
mas01mc@292 434
mas01mc@292 435 power = *spp++;
mas01mc@292 436 if(use_absolute_threshold){
mas01mc@292 437 if ( power >= absolute_threshold )
mas01mc@292 438 z++;
mas01mc@292 439 }
mas01mc@292 440 else
mas01mc@292 441 z++;
mas01mc@292 442 }
mas01mc@292 443 return z;
mas01mc@292 444 }
mas01mc@292 445
mas01mc@292 446
mas01mc@292 447 /*********************** LSH retrieval ****************************/
mas01mc@292 448
mas01mc@292 449
mas01mc@292 450 // return true if indexed query performed else return false
mas01mc@292 451 int audioDB::index_init_query(const char* dbName){
mas01mc@292 452
mas01mc@292 453 if(!(index_exists(dbName, radius, sequenceLength)))
mas01mc@292 454 return false;
mas01mc@292 455
mas01mc@292 456 char* indexName = index_get_name(dbName, radius, sequenceLength);
mas01mc@292 457
mas01mc@292 458 // Test to see if file exists
mas01mc@292 459 if((lshfid = open (indexName, O_RDONLY)) < 0){
mas01mc@292 460 delete[] indexName;
mas01mc@292 461 return false;
mas01mc@292 462 }
mas01mc@292 463
mas01mc@308 464 lsh = index_allocate(indexName, false); // Get the header only here
mas01mc@292 465 sequenceLength = lsh->get_lshHeader()->dataDim / dbH->dim; // shingleDim / vectorDim
mas01mc@292 466
mas01mc@311 467 if(lsh!=SERVER_LSH_INDEX_SINGLETON){
mas01mc@308 468 if( fabs(radius - lsh->get_radius())>fabs(O2_DISTANCE_TOLERANCE))
mas01mc@308 469 printf("*** Warning: adb_radius (%f) != lsh_radius (%f) ***\n", radius, lsh->get_radius());
mas01mc@320 470 printf("INDEX: dim %d\n", (int)dbH->dim);
mas01mc@308 471 printf("INDEX: R %f\n", lsh->get_radius());
mas01mc@308 472 printf("INDEX: seqlen %d\n", sequenceLength);
mas01mc@308 473 printf("INDEX: w %f\n", lsh->get_lshHeader()->get_binWidth());
mas01mc@308 474 printf("INDEX: k %d\n", lsh->get_lshHeader()->get_numFuns());
mas01mc@308 475 printf("INDEX: L (m*(m-1))/2 %d\n", lsh->get_lshHeader()->get_numTables());
mas01mc@308 476 printf("INDEX: N %d\n", lsh->get_lshHeader()->get_numRows());
mas01mc@319 477 printf("INDEX: s %d\n", index_to_trackID(lsh->get_maxp(), lsh_n_point_bits));
mas01mc@308 478 printf("INDEX: Opened LSH index file %s\n", indexName);
mas01mc@308 479 fflush(stdout);
mas01mc@308 480 }
mas01mc@292 481
mas01mc@292 482 // Check to see if we are loading hash tables into core, and do so if true
mas01mc@292 483 if((lsh->get_lshHeader()->flags&O2_SERIAL_FILEFORMAT2) || lsh_in_core){
mas01mc@308 484 if(SERVER_LSH_INDEX_SINGLETON)
mas01mc@308 485 fprintf(stderr,"INDEX: using persistent hash tables: %s\n", lsh->get_indexName());
mas01mc@308 486 else
mas01mc@308 487 printf("INDEX: loading hash tables into core %s\n", (lsh->get_lshHeader()->flags&O2_SERIAL_FILEFORMAT2)?"FORMAT2":"FORMAT1");
mas01mc@308 488 lsh = index_allocate(indexName, true);
mas01mc@292 489 }
mas01mc@292 490
mas01mc@292 491 delete[] indexName;
mas01mc@292 492 return true;
mas01mc@292 493 }
mas01mc@292 494
mas01mc@292 495 // *Static* approximate NN point reporter callback method for lshlib
mas01mc@292 496 void audioDB::index_add_point_approximate(void* instancePtr, Uns32T pointID, Uns32T qpos, float dist){
mas01mc@292 497 assert(instancePtr); // We need an instance for this callback
mas01mc@292 498 audioDB* myself = (audioDB*) instancePtr; // Use explicit cast to recover "this" instance
mas01mc@319 499 Uns32T trackID = index_to_trackID(pointID, myself->lsh_n_point_bits);
mas01mc@319 500 Uns32T spos = index_to_trackPos(pointID, myself->lsh_n_point_bits);
mas01mc@292 501 // Skip identity in query_from_key
mas01mc@292 502 if( !myself->query_from_key || (myself->query_from_key && ( trackID != myself->query_from_key_index )) )
mas01mc@292 503 myself->reporter->add_point(trackID, qpos, spos, dist);
mas01mc@292 504 }
mas01mc@292 505
mas01mc@292 506 // *Static* exact NN point reporter callback method for lshlib
mas01mc@292 507 // Maintain a queue of points to pass to query_points() for exact evaluation
mas01mc@292 508 void audioDB::index_add_point_exact(void* instancePtr, Uns32T pointID, Uns32T qpos, float dist){
mas01mc@292 509 assert(instancePtr); // We need an instance for this callback
mas01mc@292 510 audioDB* myself = (audioDB*) instancePtr; // Use explicit cast to recover "this" instance
mas01mc@319 511 Uns32T trackID = index_to_trackID(pointID, myself->lsh_n_point_bits);
mas01mc@319 512 Uns32T spos = index_to_trackPos(pointID, myself->lsh_n_point_bits);
mas01mc@292 513 // Skip identity in query_from_key
mas01mc@292 514 if( !myself->query_from_key || (myself->query_from_key && ( trackID != myself->query_from_key_index )) )
mas01mc@292 515 myself->index_insert_exact_evaluation_queue(trackID, qpos, spos);
mas01mc@292 516 }
mas01mc@292 517
mas01mc@292 518 void audioDB::initialize_exact_evalutation_queue(){
mas01mc@292 519 if(exact_evaluation_queue)
mas01mc@292 520 delete exact_evaluation_queue;
mas01mc@292 521 exact_evaluation_queue = new priority_queue<PointPair, std::vector<PointPair>, std::less<PointPair> >;
mas01mc@292 522 }
mas01mc@292 523
mas01mc@292 524 void audioDB::index_insert_exact_evaluation_queue(Uns32T trackID, Uns32T qpos, Uns32T spos){
mas01mc@292 525 PointPair p(trackID, qpos, spos);
mas01mc@292 526 exact_evaluation_queue->push(p);
mas01mc@292 527 }
mas01mc@292 528
mas01mc@292 529 // return 0: if index does not exist
mas01mc@292 530 // return nqv: if index exists
mas01mc@292 531 int audioDB::index_query_loop(const char* dbName, Uns32T queryIndex) {
mas01mc@292 532
mas01mc@320 533 unsigned int numVectors = 0;
mas01mc@320 534 double *query = 0, *query_data = 0;
mas01mc@320 535 double *qNorm = 0, *qnPtr = 0, *qPower = 0, *qpPtr = 0;
mas01mc@320 536 double meanQdur = 0;
mas01mc@292 537 void (*add_point_func)(void*,Uns32T,Uns32T,float);
mas01mc@292 538
mas01mc@292 539 // Set the point-reporter callback based on the value of lsh_exact
mas01mc@292 540 if(lsh_exact){
mas01mc@292 541 initialize_exact_evalutation_queue();
mas01mc@292 542 add_point_func = &index_add_point_exact;
mas01mc@292 543 }
mas01mc@292 544 else
mas01mc@292 545 add_point_func = &index_add_point_approximate;
mas01mc@292 546
mas01mc@292 547 if(!index_init_query(dbName)) // sets-up LSH index structures for querying
mas01mc@292 548 return 0;
mas01mc@292 549
mas01mc@292 550 char* database = index_get_name(dbName, radius, sequenceLength);
mas01mc@292 551
mas01mc@292 552 if(query_from_key)
mas01mc@292 553 set_up_query_from_key(&query_data, &query, &qNorm, &qnPtr, &qPower, &qpPtr, &meanQdur, &numVectors, queryIndex);
mas01mc@292 554 else
mas01mc@292 555 set_up_query(&query_data, &query, &qNorm, &qnPtr, &qPower, &qpPtr, &meanQdur, &numVectors); // get query vectors
mas01mc@292 556
mas01mc@292 557 VERB_LOG(1, "retrieving tracks...");
mas01mc@292 558
mas01mc@292 559 assert(pointNN>0 && pointNN<=O2_MAXNN);
mas01mc@292 560 assert(trackNN>0 && trackNN<=O2_MAXNN);
mas01mc@292 561
mas01mc@292 562 gettimeofday(&tv1, NULL);
mas01mc@292 563 // query vector index
mas01mc@292 564 Uns32T Nq = (numVectors>O2_MAXTRACKLEN?O2_MAXTRACKLEN:numVectors) - sequenceLength + 1;
mas01mc@292 565 vv = index_initialize_shingles(Nq); // allocate memory to copy query vectors to shingles
mas01mc@292 566 cout << "Nq=" << Nq; cout.flush();
mas01mc@292 567 // Construct shingles from query features
mas01mc@292 568 for( Uns32T pointID = 0 ; pointID < Nq ; pointID++ )
mas01mc@292 569 index_make_shingle(vv, pointID, query, dbH->dim, sequenceLength);
mas01mc@292 570
mas01mc@292 571 // Normalize query vectors
mas01mc@292 572 Uns32T numVecsAboveThreshold = index_norm_shingles( vv, qnPtr, qpPtr );
mas01mc@292 573 cout << " Nq'=" << numVecsAboveThreshold << endl; cout.flush();
mas01mc@292 574
mas01mc@292 575 // Nq contains number of inspected points in query file,
mas01mc@292 576 // numVecsAboveThreshold is number of points with power >= absolute_threshold
mas01mc@292 577 double* qpp = qpPtr; // Keep original qpPtr for possible exact evaluation
mas01mc@292 578 if(usingQueryPoint && numVecsAboveThreshold){
mas01mc@292 579 if((lsh->get_lshHeader()->flags&O2_SERIAL_FILEFORMAT2) || lsh_in_core)
mas01mc@292 580 lsh->retrieve_point((*vv)[0], queryPoint, add_point_func, (void*)this);
mas01mc@292 581 else
mas01mc@292 582 lsh->serial_retrieve_point(database, (*vv)[0], queryPoint, add_point_func, (void*)this);
mas01mc@292 583 }
mas01mc@292 584 else if(numVecsAboveThreshold)
mas01mc@292 585 for( Uns32T pointID = 0 ; pointID < Nq; pointID++ )
mas01mc@292 586 if(!use_absolute_threshold || (use_absolute_threshold && (*qpp++ >= absolute_threshold)))
mas01mc@292 587 if((lsh->get_lshHeader()->flags&O2_SERIAL_FILEFORMAT2) || lsh_in_core)
mas01mc@292 588 lsh->retrieve_point((*vv)[pointID], pointID, add_point_func, (void*)this);
mas01mc@292 589 else
mas01mc@292 590 lsh->serial_retrieve_point(database, (*vv)[pointID], pointID, add_point_func, (void*)this);
mas01mc@292 591
mas01mc@292 592 if(lsh_exact)
mas01mc@292 593 // Perform exact distance computation on point pairs in exact_evaluation_queue
mas01mc@292 594 query_loop_points(query, qnPtr, qpPtr, meanQdur, numVectors);
mas01mc@292 595
mas01mc@292 596 gettimeofday(&tv2,NULL);
mas01mc@292 597 VERB_LOG(1,"elapsed time: %ld msec\n",
mas01mc@292 598 (tv2.tv_sec*1000 + tv2.tv_usec/1000) -
mas01mc@292 599 (tv1.tv_sec*1000 + tv1.tv_usec/1000))
mas01mc@292 600
mas01mc@292 601 // Close the index file
mas01mc@292 602 close(lshfid);
mas01mc@292 603
mas01mc@292 604 // Clean up
mas01mc@292 605 if(query_data)
mas01mc@292 606 delete[] query_data;
mas01mc@292 607 if(qNorm)
mas01mc@292 608 delete[] qNorm;
mas01mc@292 609 if(qPower)
mas01mc@292 610 delete[] qPower;
mas01mc@292 611 if(database)
mas01mc@292 612 delete[] database;
mas01mc@292 613
mas01mc@292 614 return Nq;
mas01mc@292 615 }
mas01mc@292 616