annotate index.cpp @ 325:57fde215d913

Fixed indexing error in non-large ADB introduced with large ADB.
author mas01mc
date Fri, 29 Aug 2008 16:39:53 +0000
parents c93be2f3a674
children cb53c45718e6
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@324 11 //
mas01mc@324 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@324 19 Uns32T audioDB::index_to_trackID(Uns32T lshID, Uns32T nPntBits){
mas01mc@324 20 assert(nPntBits);
mas01mc@324 21 return lshID>>nPntBits;
mas01mc@292 22 }
mas01mc@292 23
mas01mc@324 24 Uns32T audioDB::index_to_trackPos(Uns32T lshID, Uns32T nPntBits){
mas01mc@324 25 assert(nPntBits);
mas01mc@324 26 return lshID&((1<<nPntBits)-1);
mas01mc@292 27 }
mas01mc@292 28
mas01mc@324 29 Uns32T audioDB::index_from_trackInfo(Uns32T trackID, Uns32T spos, Uns32T nPntBits){
mas01mc@324 30 assert(nPntBits);
mas01mc@324 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@324 60 // If we are a server and have a memory-resident index, check the indexName against the resident index (using get_indexName())
mas01mc@324 61 // If they match, i.e. path+dbName_resident == path+dbName_requested, use
mas01mc@324 62 // the memory-resident index.
mas01mc@324 63 // Else allocate a new LSH instance and load the index from disk
mas01mc@308 64 LSH* audioDB::index_allocate(char* indexName, bool load_hashTables){
mas01mc@308 65 LSH* gIndx=SERVER_LSH_INDEX_SINGLETON;
mas01mc@308 66 if(isServer && gIndx && (strncmp(gIndx->get_indexName(), indexName, MAXSTR)==0) )
mas01mc@308 67 audioDB::lsh = gIndx; // Use the global SERVER resident index
mas01mc@308 68 else{
mas01mc@308 69 if(audioDB::lsh)
mas01mc@308 70 delete audioDB::lsh;
mas01mc@308 71 audioDB::lsh = new LSH(indexName, load_hashTables);
mas01mc@308 72 }
mas01mc@308 73 assert(audioDB::lsh);
mas01mc@308 74 return audioDB::lsh;
mas01mc@308 75 }
mas01mc@308 76
mas01mc@292 77 vector<vector<float> >* audioDB::index_initialize_shingles(Uns32T sz){
mas01mc@292 78 if(vv)
mas01mc@292 79 delete vv;
mas01mc@292 80 vv = new vector<vector<float> >(sz);
mas01mc@292 81 for(Uns32T i=0 ; i < sz ; i++)
mas01mc@292 82 (*vv)[i]=vector<float>(dbH->dim*sequenceLength); // allocate shingle storage
mas01mc@292 83 return vv;
mas01mc@292 84 }
mas01mc@292 85
mas01mc@292 86 /******************** LSH indexing audioDB database access forall s \in {S} ***********************/
mas01mc@292 87
mas01mc@292 88 // Prepare the AudioDB database for read access and allocate auxillary memory
mas01mc@292 89 void audioDB::index_initialize(double **snp, double **vsnp, double **spp, double **vspp, Uns32T *dvp) {
mas01mc@324 90 if (!(dbH->flags & O2_FLAG_POWER)) {
mas01mc@324 91 error("INDEXed database must be power-enabled", dbName);
mas01mc@324 92 }
mas01mc@324 93
mas01mc@325 94 double *snpp = 0, *sppp = 0;
mas01mc@324 95
mas01mc@292 96 *dvp = dbH->length / (dbH->dim * sizeof(double)); // number of database vectors
mas01mc@292 97 *snp = new double[*dvp]; // songs norm pointer: L2 norm table for each vector
mas01mc@325 98 snpp = *snp;
mas01mc@292 99 *spp = new double[*dvp]; // song powertable pointer
mas01mc@292 100 sppp = *spp;
mas01mc@325 101
mas01mc@324 102 memcpy(*snp, l2normTable, *dvp * sizeof(double));
mas01mc@292 103 memcpy(*spp, powerTable, *dvp * sizeof(double));
mas01mc@324 104
mas01mc@324 105
mas01mc@292 106 for(Uns32T i = 0; i < dbH->numFiles; i++){
mas01mc@292 107 if(trackTable[i] >= sequenceLength) {
mas01mc@292 108 sequence_sum(snpp, trackTable[i], sequenceLength);
mas01mc@292 109 sequence_sqrt(snpp, trackTable[i], sequenceLength);
mas01mc@292 110
mas01mc@292 111 sequence_sum(sppp, trackTable[i], sequenceLength);
mas01mc@292 112 sequence_average(sppp, trackTable[i], sequenceLength);
mas01mc@292 113 }
mas01mc@292 114 snpp += trackTable[i];
mas01mc@292 115 sppp += trackTable[i];
mas01mc@292 116 }
mas01mc@324 117
mas01mc@292 118 *vsnp = *snp;
mas01mc@292 119 *vspp = *spp;
mas01mc@324 120
mas01mc@292 121 // Move the feature vector read pointer to start of fetures in database
mas01mc@292 122 lseek(dbfid, dbH->dataOffset, SEEK_SET);
mas01mc@292 123 }
mas01mc@292 124
mas01mc@292 125
mas01mc@292 126 /************************ LSH indexing ***********************************/
mas01mc@292 127 void audioDB::index_index_db(const char* dbName){
mas01mc@292 128 char* newIndexName;
mas01mc@292 129 double *fvp = 0, *sNorm = 0, *snPtr = 0, *sPower = 0, *spPtr = 0;
mas01mc@292 130 Uns32T dbVectors = 0;
mas01mc@292 131
mas01mc@324 132
mas01mc@292 133 printf("INDEX: initializing header\n");
mas01mc@292 134 // Check if audioDB exists, initialize header and open database for read
mas01mc@292 135 forWrite = false;
mas01mc@292 136 initDBHeader(dbName);
mas01mc@292 137
mas01mc@324 138 if(dbH->flags & O2_FLAG_POWER)
mas01mc@324 139 usingPower = true;
mas01mc@324 140
mas01mc@324 141 if(dbH->flags & O2_FLAG_TIMES)
mas01mc@324 142 usingTimes = true;
mas01mc@324 143
mas01mc@292 144 newIndexName = index_get_name(dbName, radius, sequenceLength);
mas01mc@292 145
mas01mc@292 146 // Set unit norming flag override
mas01mc@292 147 audioDB::normalizedDistance = !audioDB::no_unit_norming;
mas01mc@292 148
mas01mc@324 149 printf("INDEX: dim %d\n", (int)dbH->dim);
mas01mc@292 150 printf("INDEX: R %f\n", radius);
mas01mc@292 151 printf("INDEX: seqlen %d\n", sequenceLength);
mas01mc@292 152 printf("INDEX: lsh_w %f\n", lsh_param_w);
mas01mc@292 153 printf("INDEX: lsh_k %d\n", lsh_param_k);
mas01mc@292 154 printf("INDEX: lsh_m %d\n", lsh_param_m);
mas01mc@292 155 printf("INDEX: lsh_N %d\n", lsh_param_N);
mas01mc@296 156 printf("INDEX: lsh_C %d\n", lsh_param_ncols);
mas01mc@292 157 printf("INDEX: lsh_b %d\n", lsh_param_b);
mas01mc@292 158 printf("INDEX: normalized? %s\n", normalizedDistance?"true":"false");
mas01mc@292 159 fflush(stdout);
mas01mc@292 160
mas01mc@292 161
mas01mc@292 162 if((lshfid = open(newIndexName,O_RDONLY))<0){
mas01mc@292 163 printf("INDEX: constructing new LSH index\n");
mas01mc@292 164 printf("INDEX: making index file %s\n", newIndexName);
mas01mc@292 165 fflush(stdout);
mas01mc@292 166 // Construct new LSH index
mas01mc@292 167 lsh = new LSH((float)lsh_param_w, lsh_param_k,
mas01mc@292 168 lsh_param_m,
mas01mc@292 169 (Uns32T)(sequenceLength*dbH->dim),
mas01mc@292 170 lsh_param_N,
mas01mc@292 171 lsh_param_ncols,
mas01mc@292 172 (float)radius);
mas01mc@292 173 assert(lsh);
mas01mc@292 174
mas01mc@292 175 Uns32T endTrack = lsh_param_b;
mas01mc@292 176 if( endTrack > dbH->numFiles)
mas01mc@292 177 endTrack = dbH->numFiles;
mas01mc@292 178 // Insert up to lsh_param_b tracks
mas01mc@324 179 if( ! (dbH->flags & O2_FLAG_LARGE_ADB) ){
mas01mc@324 180 index_initialize(&sNorm, &snPtr, &sPower, &spPtr, &dbVectors);
mas01mc@324 181 }
mas01mc@324 182 index_insert_tracks(0, endTrack, &fvp, &sNorm, &snPtr, &sPower, &spPtr);
mas01mc@292 183 lsh->serialize(newIndexName, lsh_in_core?O2_SERIAL_FILEFORMAT2:O2_SERIAL_FILEFORMAT1);
mas01mc@292 184
mas01mc@292 185 // Clean up
mas01mc@292 186 delete lsh;
mas01mc@308 187 lsh = 0;
mas01mc@292 188 close(lshfid);
mas01mc@292 189 }
mas01mc@292 190
mas01mc@292 191 // Attempt to open LSH file
mas01mc@292 192 if((lshfid = open(newIndexName,O_RDONLY))>0){
mas01mc@292 193 printf("INDEX: merging with existing LSH index\n");
mas01mc@292 194 fflush(stdout);
mas01mc@292 195
mas01mc@292 196 // Get the lsh header info and find how many tracks are inserted already
mas01mc@292 197 lsh = new LSH(newIndexName, false); // lshInCore=false to avoid loading hashTables here
mas01mc@292 198 assert(lsh);
mas01mc@324 199 Uns32T maxs = index_to_trackID(lsh->get_maxp(), lsh_n_point_bits)+1;
mas01mc@292 200 delete lsh;
mas01mc@308 201 lsh = 0;
mas01mc@292 202
mas01mc@292 203 // This allows for updating index after more tracks are inserted into audioDB
mas01mc@292 204 for(Uns32T startTrack = maxs; startTrack < dbH->numFiles; startTrack+=lsh_param_b){
mas01mc@292 205
mas01mc@292 206 Uns32T endTrack = startTrack + lsh_param_b;
mas01mc@292 207 if( endTrack > dbH->numFiles)
mas01mc@292 208 endTrack = dbH->numFiles;
mas01mc@292 209 printf("Indexing track range: %d - %d\n", startTrack, endTrack);
mas01mc@292 210 fflush(stdout);
mas01mc@292 211 lsh = new LSH(newIndexName, lsh_in_core); // Initialize core memory for LSH tables
mas01mc@292 212 assert(lsh);
mas01mc@292 213
mas01mc@292 214 // Insert up to lsh_param_b database tracks
mas01mc@292 215 index_insert_tracks(startTrack, endTrack, &fvp, &sNorm, &snPtr, &sPower, &spPtr);
mas01mc@292 216
mas01mc@292 217 // Serialize to file
mas01mc@292 218 lsh->serialize(newIndexName, lsh_in_core?O2_SERIAL_FILEFORMAT2:O2_SERIAL_FILEFORMAT1); // Serialize core LSH heap to disk
mas01mc@292 219 delete lsh;
mas01mc@308 220 lsh = 0;
mas01mc@292 221 }
mas01mc@292 222
mas01mc@292 223 close(lshfid);
mas01mc@292 224 printf("INDEX: done constructing LSH index.\n");
mas01mc@292 225 fflush(stdout);
mas01mc@292 226
mas01mc@292 227 }
mas01mc@292 228 else{
mas01mc@292 229 error("Something's wrong with LSH index file");
mas01mc@292 230 exit(1);
mas01mc@292 231 }
mas01mc@292 232
mas01mc@324 233 delete[] newIndexName;
mas01mc@324 234 delete[] sNorm;
mas01mc@324 235 delete[] sPower;
mas01mc@324 236 }
mas01mc@292 237
mas01mc@292 238
mas01mc@324 239 // initialize auxillary track data from filesystem
mas01mc@324 240 // pre-conditions:
mas01mc@324 241 // dbH->flags & O2_FLAG_LARGE_ADB
mas01mc@324 242 // feature data allocated and copied (fvp)
mas01mc@324 243 //
mas01mc@324 244 // post-conditions:
mas01mc@324 245 // allocated power data
mas01mc@324 246 // allocated l2norm data
mas01mc@324 247 //
mas01mc@324 248 void audioDB::init_track_aux_data(Uns32T trackID, double* fvp, double** sNormpp,double** snPtrp, double** sPowerp, double** spPtrp){
mas01mc@324 249 if( !(dbH->flags & O2_FLAG_LARGE_ADB) )
mas01mc@324 250 error("error: init_track_large_adb required O2_FLAG_LARGE_ADB");
mas01mc@292 251
mas01mc@324 252 // Allocate and read the power sequence
mas01mc@324 253 if(trackTable[trackID]>=sequenceLength){
mas01mc@324 254
mas01mc@324 255 char* prefixedString = new char[O2_MAXFILESTR];
mas01mc@324 256 char* tmpStr = prefixedString;
mas01mc@324 257 // Open and check dimensions of power file
mas01mc@324 258 strncpy(prefixedString, powerFileNameTable+trackID*O2_FILETABLE_ENTRY_SIZE, O2_MAXFILESTR);
mas01mc@324 259 prefix_name((char ** const)&prefixedString, adb_feature_root);
mas01mc@324 260 if(prefixedString!=tmpStr)
mas01mc@324 261 delete[] tmpStr;
mas01mc@324 262 powerfd = open(prefixedString, O_RDONLY);
mas01mc@324 263 if (powerfd < 0) {
mas01mc@324 264 error("failed to open power file", prefixedString);
mas01mc@324 265 }
mas01mc@324 266 if (fstat(powerfd, &statbuf) < 0) {
mas01mc@324 267 error("fstat error finding size of power file", prefixedString, "fstat");
mas01mc@324 268 }
mas01mc@324 269
mas01mc@324 270 if( (statbuf.st_size - sizeof(int)) / (sizeof(double)) != trackTable[trackID] )
mas01mc@324 271 error("Dimension mismatch: numPowers != numVectors", prefixedString);
mas01mc@324 272
mas01mc@324 273 *sPowerp = new double[trackTable[trackID]]; // Allocate memory for power values
mas01mc@324 274 assert(*sPowerp);
mas01mc@324 275 *spPtrp = *sPowerp;
mas01mc@324 276 insertPowerData(trackTable[trackID], powerfd, *sPowerp);
mas01mc@324 277 if (0 < powerfd) {
mas01mc@324 278 close(powerfd);
mas01mc@324 279 }
mas01mc@324 280
mas01mc@324 281 sequence_sum(*sPowerp, trackTable[trackID], sequenceLength);
mas01mc@324 282 sequence_average(*sPowerp, trackTable[trackID], sequenceLength);
mas01mc@324 283 powerTable = 0;
mas01mc@324 284
mas01mc@324 285 // Allocate and calculate the l2norm sequence
mas01mc@324 286 *sNormpp = new double[trackTable[trackID]];
mas01mc@324 287 assert(*sNormpp);
mas01mc@324 288 *snPtrp = *sNormpp;
mas01mc@324 289 unitNorm(fvp, dbH->dim, trackTable[trackID], *sNormpp);
mas01mc@324 290 sequence_sum(*sNormpp, trackTable[trackID], sequenceLength);
mas01mc@324 291 sequence_sqrt(*sNormpp, trackTable[trackID], sequenceLength);
mas01mc@324 292 }
mas01mc@292 293 }
mas01mc@292 294
mas01mc@292 295 void audioDB::index_insert_tracks(Uns32T start_track, Uns32T end_track,
mas01mc@292 296 double** fvpp, double** sNormpp,double** snPtrp,
mas01mc@292 297 double** sPowerp, double** spPtrp){
mas01mc@292 298 size_t nfv = 0;
mas01mc@292 299 double* fvp = 0; // Keep pointer for memory allocation and free() for track data
mas01mc@292 300 Uns32T trackID = 0;
mas01mc@292 301
mas01mc@292 302 VERB_LOG(1, "indexing tracks...");
mas01mc@292 303
mas01mc@324 304 int trackfd = dbfid;
mas01mc@292 305 for(trackID = start_track ; trackID < end_track ; trackID++ ){
mas01mc@324 306 if( dbH->flags & O2_FLAG_LARGE_ADB ){
mas01mc@324 307 char* prefixedString = new char[O2_MAXFILESTR];
mas01mc@324 308 char* tmpStr = prefixedString;
mas01mc@324 309 // Open and check dimensions of feature file
mas01mc@324 310 strncpy(prefixedString, featureFileNameTable+trackID*O2_FILETABLE_ENTRY_SIZE, O2_MAXFILESTR);
mas01mc@324 311 prefix_name((char ** const) &prefixedString, adb_feature_root);
mas01mc@324 312 if(prefixedString!=tmpStr)
mas01mc@324 313 delete[] tmpStr;
mas01mc@324 314 initInputFile(prefixedString, false); // nommap, file pointer at correct position
mas01mc@324 315 trackfd = infid;
mas01mc@324 316 }
mas01mc@324 317 read_data(trackfd, trackID, &fvp, &nfv); // over-writes fvp and nfv
mas01mc@292 318 *fvpp = fvp; // Protect memory allocation and free() for track data
mas01mc@324 319
mas01mc@324 320 if( dbH->flags & O2_FLAG_LARGE_ADB )
mas01mc@324 321 // Load power and calculate power and l2norm sequence sums
mas01mc@324 322 init_track_aux_data(trackID, fvp, sNormpp, snPtrp, sPowerp, spPtrp);
mas01mc@324 323
mas01mc@292 324 if(!index_insert_track(trackID, fvpp, snPtrp, spPtrp))
mas01mc@292 325 break;
mas01mc@324 326 if ( dbH->flags & O2_FLAG_LARGE_ADB ){
mas01mc@324 327 close(infid);
mas01mc@324 328 delete[] *sNormpp;
mas01mc@324 329 delete[] *sPowerp;
mas01mc@324 330 *sNormpp = *sPowerp = *snPtrp = *snPtrp = 0;
mas01mc@324 331 }
mas01mc@324 332 } // end for(trackID = start_track ; ... )
mas01mc@292 333 std::cout << "finished inserting." << endl;
mas01mc@292 334 }
mas01mc@292 335
mas01mc@292 336 int audioDB::index_insert_track(Uns32T trackID, double** fvpp, double** snpp, double** sppp){
mas01mc@292 337 // Loop over the current input track's vectors
mas01cr@305 338 Uns32T numVecs = 0;
mas01cr@305 339 if (trackTable[trackID] > O2_MAXTRACKLEN) {
mas01cr@305 340 if (O2_MAXTRACKLEN < sequenceLength - 1) {
mas01cr@305 341 numVecs = 0;
mas01cr@305 342 } else {
mas01cr@305 343 numVecs = O2_MAXTRACKLEN - sequenceLength + 1;
mas01cr@305 344 }
mas01cr@305 345 } else {
mas01cr@305 346 if (trackTable[trackID] < sequenceLength - 1) {
mas01cr@305 347 numVecs = 0;
mas01cr@305 348 } else {
mas01cr@305 349 numVecs = trackTable[trackID] - sequenceLength + 1;
mas01cr@305 350 }
mas01cr@305 351 }
mas01mc@292 352
mas01mc@324 353 Uns32T numVecsAboveThreshold = 0, collisionCount = 0;
mas01mc@324 354 if(numVecs){
mas01mc@324 355 vv = index_initialize_shingles(numVecs);
mas01mc@324 356
mas01mc@324 357 for( Uns32T pointID = 0 ; pointID < numVecs; pointID++ )
mas01mc@324 358 index_make_shingle(vv, pointID, *fvpp, dbH->dim, sequenceLength);
mas01mc@324 359
mas01mc@324 360 numVecsAboveThreshold = index_norm_shingles(vv, *snpp, *sppp);
mas01mc@324 361 collisionCount = index_insert_shingles(vv, trackID, *sppp);
mas01mc@324 362 }
mas01mc@292 363 float meanCollisionCount = numVecsAboveThreshold?(float)collisionCount/numVecsAboveThreshold:0;
mas01mc@292 364
mas01mc@292 365 /* index_norm_shingles() only goes as far as the end of the
mas01mc@292 366 sequence, which is right, but the space allocated is for the
mas01mc@292 367 whole track. */
mas01mc@292 368
mas01mc@292 369 /* But numVecs will be <trackTable[track] if trackTable[track]>O2_MAXTRACKLEN
mas01mc@292 370 * So let's be certain the pointers are in the correct place
mas01mc@292 371 */
mas01mc@292 372
mas01mc@324 373 if( !(dbH->flags & O2_FLAG_LARGE_ADB) ){
mas01mc@324 374 *snpp += trackTable[trackID];
mas01mc@324 375 *sppp += trackTable[trackID];
mas01mc@324 376 *fvpp += trackTable[trackID] * dbH->dim;
mas01mc@324 377 }
mas01mc@292 378
mas01mc@292 379 std::cout << " n=" << trackTable[trackID] << " n'=" << numVecsAboveThreshold << " E[#c]=" << lsh->get_mean_collision_rate() << " E[#p]=" << meanCollisionCount << endl;
mas01mc@292 380 std::cout.flush();
mas01mc@292 381 return true;
mas01mc@292 382 }
mas01mc@292 383
mas01mc@292 384 Uns32T audioDB::index_insert_shingles(vector<vector<float> >* vv, Uns32T trackID, double* spp){
mas01mc@292 385 Uns32T collisionCount = 0;
mas01mc@292 386 cout << "[" << trackID << "]" << fileTable+trackID*O2_FILETABLE_ENTRY_SIZE;
mas01mc@324 387 for( Uns32T pointID=0 ; pointID < (*vv).size(); pointID+=sequenceHop){
mas01mc@324 388 if(!use_absolute_threshold || (use_absolute_threshold && (*spp >= absolute_threshold)))
mas01mc@324 389 collisionCount += lsh->insert_point((*vv)[pointID], index_from_trackInfo(trackID, pointID, lsh_n_point_bits));
mas01mc@324 390 spp+=sequenceHop;
mas01mc@311 391 }
mas01mc@292 392 return collisionCount;
mas01mc@292 393 }
mas01mc@292 394
mas01mc@292 395 /********************* LSH shingle construction ***************************/
mas01mc@292 396
mas01mc@292 397 // Construct shingles out of a feature matrix
mas01mc@292 398 // inputs:
mas01mc@292 399 // idx is vector index in feature matrix
mas01mc@292 400 // fvp is base feature matrix pointer double* [numVecs x dbH->dim]
mas01mc@292 401 //
mas01mc@292 402 // pre-conditions:
mas01mc@292 403 // dbH->dim
mas01mc@292 404 // sequenceLength
mas01mc@292 405 // idx < numVectors - sequenceLength + 1
mas01mc@292 406 //
mas01mc@292 407 // post-conditions:
mas01mc@292 408 // (*vv)[idx] contains a shingle with dbH->dim*sequenceLength float values
mas01mc@292 409
mas01mc@292 410 void audioDB::index_make_shingle(vector<vector<float> >* vv, Uns32T idx, double* fvp, Uns32T dim, Uns32T seqLen){
mas01mc@292 411 assert(idx<(*vv).size());
mas01mc@292 412 vector<float>::iterator ve = (*vv)[idx].end();
mas01mc@292 413 vi=(*vv)[idx].begin(); // shingle iterator
mas01mc@292 414 // First feature vector in shingle
mas01mc@292 415 if(idx==0){
mas01mc@292 416 while(vi!=ve)
mas01mc@292 417 *vi++ = (float)(*fvp++);
mas01mc@292 418 }
mas01mc@292 419 // Not first feature vector in shingle
mas01mc@292 420 else{
mas01mc@292 421 vector<float>::iterator ui=(*vv)[idx-1].begin() + dim; // previous shingle iterator
mas01mc@292 422 // Previous seqLen-1 dim-vectors
mas01mc@292 423 while(vi!=ve-dim)
mas01mc@292 424 *vi++=*ui++;
mas01mc@292 425 // Move data pointer to next feature vector
mas01mc@292 426 fvp += ( seqLen + idx - 1 ) * dim ;
mas01mc@292 427 // New d-vector
mas01mc@292 428 while(vi!=ve)
mas01mc@292 429 *vi++ = (float)(*fvp++);
mas01mc@292 430 }
mas01mc@292 431 }
mas01mc@292 432
mas01mc@292 433 // norm shingles
mas01mc@292 434 // in-place norming, no deletions
mas01mc@292 435 // If using power, return number of shingles above power threshold
mas01mc@292 436 int audioDB::index_norm_shingles(vector<vector<float> >* vv, double* snp, double* spp){
mas01mc@292 437 int z = 0; // number of above-threshold shingles
mas01mc@292 438 float l2norm;
mas01mc@292 439 double power;
mas01mc@292 440 float oneOverRadius = 1./(float)sqrt(radius); // Passed radius is really radius^2
mas01mc@292 441 float oneOverSqrtl2NormDivRad = oneOverRadius;
mas01mc@292 442 if(!spp)
mas01mc@292 443 error("LSH indexing and query requires a power feature using -w or -W");
mas01mc@292 444 Uns32T shingleSize = sequenceLength*dbH->dim;
mas01mc@292 445 for(Uns32T a=0; a<(*vv).size(); a++){
mas01mc@292 446 l2norm = (float)(*snp++);
mas01mc@292 447 if(audioDB::normalizedDistance)
mas01mc@292 448 oneOverSqrtl2NormDivRad = (1./l2norm)*oneOverRadius;
mas01mc@292 449
mas01mc@292 450 for(Uns32T b=0; b < shingleSize ; b++)
mas01mc@292 451 (*vv)[a][b]*=oneOverSqrtl2NormDivRad;
mas01mc@292 452
mas01mc@292 453 power = *spp++;
mas01mc@292 454 if(use_absolute_threshold){
mas01mc@292 455 if ( power >= absolute_threshold )
mas01mc@292 456 z++;
mas01mc@292 457 }
mas01mc@292 458 else
mas01mc@292 459 z++;
mas01mc@292 460 }
mas01mc@292 461 return z;
mas01mc@292 462 }
mas01mc@292 463
mas01mc@292 464
mas01mc@292 465 /*********************** LSH retrieval ****************************/
mas01mc@292 466
mas01mc@292 467
mas01mc@292 468 // return true if indexed query performed else return false
mas01mc@292 469 int audioDB::index_init_query(const char* dbName){
mas01mc@292 470
mas01mc@292 471 if(!(index_exists(dbName, radius, sequenceLength)))
mas01mc@292 472 return false;
mas01mc@292 473
mas01mc@292 474 char* indexName = index_get_name(dbName, radius, sequenceLength);
mas01mc@292 475
mas01mc@292 476 // Test to see if file exists
mas01mc@292 477 if((lshfid = open (indexName, O_RDONLY)) < 0){
mas01mc@292 478 delete[] indexName;
mas01mc@292 479 return false;
mas01mc@292 480 }
mas01mc@292 481
mas01mc@308 482 lsh = index_allocate(indexName, false); // Get the header only here
mas01mc@292 483 sequenceLength = lsh->get_lshHeader()->dataDim / dbH->dim; // shingleDim / vectorDim
mas01mc@292 484
mas01mc@311 485 if(lsh!=SERVER_LSH_INDEX_SINGLETON){
mas01mc@308 486 if( fabs(radius - lsh->get_radius())>fabs(O2_DISTANCE_TOLERANCE))
mas01mc@308 487 printf("*** Warning: adb_radius (%f) != lsh_radius (%f) ***\n", radius, lsh->get_radius());
mas01mc@324 488 printf("INDEX: dim %d\n", (int)dbH->dim);
mas01mc@308 489 printf("INDEX: R %f\n", lsh->get_radius());
mas01mc@308 490 printf("INDEX: seqlen %d\n", sequenceLength);
mas01mc@308 491 printf("INDEX: w %f\n", lsh->get_lshHeader()->get_binWidth());
mas01mc@308 492 printf("INDEX: k %d\n", lsh->get_lshHeader()->get_numFuns());
mas01mc@308 493 printf("INDEX: L (m*(m-1))/2 %d\n", lsh->get_lshHeader()->get_numTables());
mas01mc@308 494 printf("INDEX: N %d\n", lsh->get_lshHeader()->get_numRows());
mas01mc@324 495 printf("INDEX: s %d\n", index_to_trackID(lsh->get_maxp(), lsh_n_point_bits));
mas01mc@308 496 printf("INDEX: Opened LSH index file %s\n", indexName);
mas01mc@308 497 fflush(stdout);
mas01mc@308 498 }
mas01mc@292 499
mas01mc@292 500 // Check to see if we are loading hash tables into core, and do so if true
mas01mc@292 501 if((lsh->get_lshHeader()->flags&O2_SERIAL_FILEFORMAT2) || lsh_in_core){
mas01mc@308 502 if(SERVER_LSH_INDEX_SINGLETON)
mas01mc@308 503 fprintf(stderr,"INDEX: using persistent hash tables: %s\n", lsh->get_indexName());
mas01mc@308 504 else
mas01mc@308 505 printf("INDEX: loading hash tables into core %s\n", (lsh->get_lshHeader()->flags&O2_SERIAL_FILEFORMAT2)?"FORMAT2":"FORMAT1");
mas01mc@308 506 lsh = index_allocate(indexName, true);
mas01mc@292 507 }
mas01mc@292 508
mas01mc@292 509 delete[] indexName;
mas01mc@292 510 return true;
mas01mc@292 511 }
mas01mc@292 512
mas01mc@292 513 // *Static* approximate NN point reporter callback method for lshlib
mas01mc@292 514 void audioDB::index_add_point_approximate(void* instancePtr, Uns32T pointID, Uns32T qpos, float dist){
mas01mc@292 515 assert(instancePtr); // We need an instance for this callback
mas01mc@292 516 audioDB* myself = (audioDB*) instancePtr; // Use explicit cast to recover "this" instance
mas01mc@324 517 Uns32T trackID = index_to_trackID(pointID, myself->lsh_n_point_bits);
mas01mc@324 518 Uns32T spos = index_to_trackPos(pointID, myself->lsh_n_point_bits);
mas01mc@292 519 // Skip identity in query_from_key
mas01mc@292 520 if( !myself->query_from_key || (myself->query_from_key && ( trackID != myself->query_from_key_index )) )
mas01mc@292 521 myself->reporter->add_point(trackID, qpos, spos, dist);
mas01mc@292 522 }
mas01mc@292 523
mas01mc@292 524 // *Static* exact NN point reporter callback method for lshlib
mas01mc@292 525 // Maintain a queue of points to pass to query_points() for exact evaluation
mas01mc@292 526 void audioDB::index_add_point_exact(void* instancePtr, Uns32T pointID, Uns32T qpos, float dist){
mas01mc@292 527 assert(instancePtr); // We need an instance for this callback
mas01mc@292 528 audioDB* myself = (audioDB*) instancePtr; // Use explicit cast to recover "this" instance
mas01mc@324 529 Uns32T trackID = index_to_trackID(pointID, myself->lsh_n_point_bits);
mas01mc@324 530 Uns32T spos = index_to_trackPos(pointID, myself->lsh_n_point_bits);
mas01mc@292 531 // Skip identity in query_from_key
mas01mc@292 532 if( !myself->query_from_key || (myself->query_from_key && ( trackID != myself->query_from_key_index )) )
mas01mc@292 533 myself->index_insert_exact_evaluation_queue(trackID, qpos, spos);
mas01mc@292 534 }
mas01mc@292 535
mas01mc@292 536 void audioDB::initialize_exact_evalutation_queue(){
mas01mc@292 537 if(exact_evaluation_queue)
mas01mc@292 538 delete exact_evaluation_queue;
mas01mc@292 539 exact_evaluation_queue = new priority_queue<PointPair, std::vector<PointPair>, std::less<PointPair> >;
mas01mc@292 540 }
mas01mc@292 541
mas01mc@292 542 void audioDB::index_insert_exact_evaluation_queue(Uns32T trackID, Uns32T qpos, Uns32T spos){
mas01mc@292 543 PointPair p(trackID, qpos, spos);
mas01mc@292 544 exact_evaluation_queue->push(p);
mas01mc@292 545 }
mas01mc@292 546
mas01mc@292 547 // return 0: if index does not exist
mas01mc@292 548 // return nqv: if index exists
mas01mc@292 549 int audioDB::index_query_loop(const char* dbName, Uns32T queryIndex) {
mas01mc@292 550
mas01mc@324 551 unsigned int numVectors = 0;
mas01mc@324 552 double *query = 0, *query_data = 0;
mas01mc@324 553 double *qNorm = 0, *qnPtr = 0, *qPower = 0, *qpPtr = 0;
mas01mc@324 554 double meanQdur = 0;
mas01mc@292 555 void (*add_point_func)(void*,Uns32T,Uns32T,float);
mas01mc@292 556
mas01mc@292 557 // Set the point-reporter callback based on the value of lsh_exact
mas01mc@292 558 if(lsh_exact){
mas01mc@292 559 initialize_exact_evalutation_queue();
mas01mc@292 560 add_point_func = &index_add_point_exact;
mas01mc@292 561 }
mas01mc@292 562 else
mas01mc@292 563 add_point_func = &index_add_point_approximate;
mas01mc@292 564
mas01mc@292 565 if(!index_init_query(dbName)) // sets-up LSH index structures for querying
mas01mc@292 566 return 0;
mas01mc@292 567
mas01mc@292 568 char* database = index_get_name(dbName, radius, sequenceLength);
mas01mc@292 569
mas01mc@292 570 if(query_from_key)
mas01mc@292 571 set_up_query_from_key(&query_data, &query, &qNorm, &qnPtr, &qPower, &qpPtr, &meanQdur, &numVectors, queryIndex);
mas01mc@292 572 else
mas01mc@292 573 set_up_query(&query_data, &query, &qNorm, &qnPtr, &qPower, &qpPtr, &meanQdur, &numVectors); // get query vectors
mas01mc@292 574
mas01mc@292 575 VERB_LOG(1, "retrieving tracks...");
mas01mc@292 576
mas01mc@292 577 assert(pointNN>0 && pointNN<=O2_MAXNN);
mas01mc@292 578 assert(trackNN>0 && trackNN<=O2_MAXNN);
mas01mc@292 579
mas01mc@292 580 gettimeofday(&tv1, NULL);
mas01mc@292 581 // query vector index
mas01mc@292 582 Uns32T Nq = (numVectors>O2_MAXTRACKLEN?O2_MAXTRACKLEN:numVectors) - sequenceLength + 1;
mas01mc@292 583 vv = index_initialize_shingles(Nq); // allocate memory to copy query vectors to shingles
mas01mc@292 584 cout << "Nq=" << Nq; cout.flush();
mas01mc@292 585 // Construct shingles from query features
mas01mc@292 586 for( Uns32T pointID = 0 ; pointID < Nq ; pointID++ )
mas01mc@292 587 index_make_shingle(vv, pointID, query, dbH->dim, sequenceLength);
mas01mc@292 588
mas01mc@292 589 // Normalize query vectors
mas01mc@292 590 Uns32T numVecsAboveThreshold = index_norm_shingles( vv, qnPtr, qpPtr );
mas01mc@292 591 cout << " Nq'=" << numVecsAboveThreshold << endl; cout.flush();
mas01mc@292 592
mas01mc@292 593 // Nq contains number of inspected points in query file,
mas01mc@292 594 // numVecsAboveThreshold is number of points with power >= absolute_threshold
mas01mc@292 595 double* qpp = qpPtr; // Keep original qpPtr for possible exact evaluation
mas01mc@292 596 if(usingQueryPoint && numVecsAboveThreshold){
mas01mc@292 597 if((lsh->get_lshHeader()->flags&O2_SERIAL_FILEFORMAT2) || lsh_in_core)
mas01mc@292 598 lsh->retrieve_point((*vv)[0], queryPoint, add_point_func, (void*)this);
mas01mc@292 599 else
mas01mc@292 600 lsh->serial_retrieve_point(database, (*vv)[0], queryPoint, add_point_func, (void*)this);
mas01mc@292 601 }
mas01mc@292 602 else if(numVecsAboveThreshold)
mas01mc@292 603 for( Uns32T pointID = 0 ; pointID < Nq; pointID++ )
mas01mc@292 604 if(!use_absolute_threshold || (use_absolute_threshold && (*qpp++ >= absolute_threshold)))
mas01mc@292 605 if((lsh->get_lshHeader()->flags&O2_SERIAL_FILEFORMAT2) || lsh_in_core)
mas01mc@292 606 lsh->retrieve_point((*vv)[pointID], pointID, add_point_func, (void*)this);
mas01mc@292 607 else
mas01mc@292 608 lsh->serial_retrieve_point(database, (*vv)[pointID], pointID, add_point_func, (void*)this);
mas01mc@292 609
mas01mc@292 610 if(lsh_exact)
mas01mc@292 611 // Perform exact distance computation on point pairs in exact_evaluation_queue
mas01mc@292 612 query_loop_points(query, qnPtr, qpPtr, meanQdur, numVectors);
mas01mc@292 613
mas01mc@292 614 gettimeofday(&tv2,NULL);
mas01mc@292 615 VERB_LOG(1,"elapsed time: %ld msec\n",
mas01mc@292 616 (tv2.tv_sec*1000 + tv2.tv_usec/1000) -
mas01mc@292 617 (tv1.tv_sec*1000 + tv1.tv_usec/1000))
mas01mc@292 618
mas01mc@292 619 // Close the index file
mas01mc@292 620 close(lshfid);
mas01mc@292 621
mas01mc@292 622 // Clean up
mas01mc@292 623 if(query_data)
mas01mc@292 624 delete[] query_data;
mas01mc@292 625 if(qNorm)
mas01mc@292 626 delete[] qNorm;
mas01mc@292 627 if(qPower)
mas01mc@292 628 delete[] qPower;
mas01mc@292 629 if(database)
mas01mc@292 630 delete[] database;
mas01mc@292 631
mas01mc@292 632 return Nq;
mas01mc@292 633 }
mas01mc@292 634