Changeset 10 for trunk/bin/train
- Timestamp:
- 07/13/15 19:16:30 (9 years ago)
- File:
-
- 1 edited
Legend:
- Unmodified
- Added
- Removed
-
trunk/bin/train
r9 r10 26 26 27 27 from nanownlib import * 28 from nanownlib.stats import * 29 from nanownlib.parallel import WorkerThreads 28 30 import nanownlib.storage 29 from nanownlib.stats import boxTest,multiBoxTest,subsample,bootstrap,bootstrap2,trimean,midhinge,midhingeTest,samples2Distributions,samples2MeanDiffs 30 from nanownlib.parallel import WorkerThreads 31 31 32 32 33 … … 35 36 #parser.add_argument('-c', dest='cases', type=str, default='{"short":10000,"long":1010000}', 36 37 # help='JSON representation of echo timing cases. Default: {"short":10000,"long":1010000}') 38 parser.add_argument('--retrain', action='append', default=[], help='Force a classifier to be retrained. May be specified multiple times.') 39 parser.add_argument('--retest', action='append', default=[], help='Force a classifier to be retested. May be specified multiple times.') 37 40 parser.add_argument('session_data', default=None, 38 41 help='Database file storing session information') 39 42 options = parser.parse_args() 40 43 41 42 43 def trainBoxTest(db, unusual_case, greater, subseries_size): 44 44 45 def trainBoxTest(db, unusual_case, greater, num_observations): 46 db.resetOffsets() 47 45 48 def trainAux(low,high,num_trials): 46 49 estimator = functools.partial(multiBoxTest, {'low':low, 'high':high}, greater) 47 estimates = bootstrap3(estimator, db, 'train', unusual_case, subseries_size, num_trials)48 null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, subseries_size, num_trials)50 estimates = bootstrap3(estimator, db, 'train', unusual_case, num_observations, num_trials) 51 null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials) 49 52 50 53 bad_estimates = len([e for e in estimates if e != 1]) … … 116 119 117 120 num_trials = 500 118 widths = [good_width+(x/100.0) for x in range(- 60,75,5) if good_width+(x/100.0) > 0.0]121 widths = [good_width+(x/100.0) for x in range(-70,75,5) if good_width+(x/100.0) > 0.0] 119 122 performance = [] 120 123 for width in widths: … … 124 127 job_id,errors = wt.resultq.get() 125 128 fp,fn = errors 126 performance.append(((fp+fn)/2.0, job_id, fn, fp)) 129 #performance.append(((fp+fn)/2.0, job_id, fn, fp)) 130 performance.append((abs(fp-fn), job_id, fn, fp)) 127 131 performance.sort() 128 132 #pprint.pprint(performance) … … 133 137 wt.stop() 134 138 params = json.dumps({"low":best_low,"high":best_low+best_width}) 135 return {'algorithm':"boxtest", 139 return {'trial_type':"train", 140 'num_observations':num_observations, 141 'num_trials':num_trials, 136 142 'params':params, 137 'sample_size':subseries_size,138 'num_trials':num_trials,139 'trial_type':"train",140 143 'false_positives':performance[0][3], 141 144 'false_negatives':performance[0][2]} 142 145 143 146 144 def trainMidhinge(db, unusual_case, greater, subseries_size): 145 147 def trainSummary(summaryFunc, db, unusual_case, greater, num_observations): 148 db.resetOffsets() 149 stest = functools.partial(summaryTest, summaryFunc) 150 146 151 def trainAux(distance, threshold, num_trials): 147 estimator = functools.partial( midhingeTest, {'distance':distance,'threshold':threshold}, greater)148 estimates = bootstrap3(estimator, db, 'train', unusual_case, subseries_size, num_trials)149 null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, subseries_size, num_trials)152 estimator = functools.partial(stest, {'distance':distance,'threshold':threshold}, greater) 153 estimates = bootstrap3(estimator, db, 'train', unusual_case, num_observations, num_trials) 154 null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials) 150 155 151 156 bad_estimates = len([e for e in estimates if e != 1]) … … 158 163 #determine expected delta based on differences 159 164 mean_diffs = [s['unusual_case']-s['other_cases'] for s in db.subseries('train', unusual_case)] 160 threshold = trimean(mean_diffs)/2.0165 threshold = summaryFunc(mean_diffs)/2.0 161 166 #print("init_threshold:", threshold) 162 167 … … 181 186 num_trials = 500 182 187 performance = [] 183 for t in range( 50,154,4):188 for t in range(80,122,2): 184 189 wt.addJob(threshold*(t/100.0), (good_distance,threshold*(t/100.0),num_trials)) 185 190 wt.wait() … … 187 192 job_id,errors = wt.resultq.get() 188 193 fp,fn = errors 189 performance.append(((fp+fn)/2.0, job_id, fn, fp)) 194 #performance.append(((fp+fn)/2.0, job_id, fn, fp)) 195 performance.append((abs(fp-fn), job_id, fn, fp)) 190 196 performance.sort() 191 197 #pprint.pprint(performance) … … 217 223 job_id,errors = wt.resultq.get() 218 224 fp,fn = errors 219 performance.append(((fp+fn)/2.0, job_id, fn, fp)) 225 #performance.append(((fp+fn)/2.0, job_id, fn, fp)) 226 performance.append((abs(fp-fn), job_id, fn, fp)) 220 227 performance.sort() 221 228 #pprint.pprint(performance) … … 225 232 wt.stop() 226 233 params = json.dumps({'distance':best_distance,'threshold':best_threshold}) 227 return {'algorithm':"midhinge", 234 return {'trial_type':"train", 235 'num_observations':num_observations, 236 'num_trials':num_trials, 228 237 'params':params, 229 'sample_size':subseries_size,230 'num_trials':num_trials,231 'trial_type':"train",232 238 'false_positives':performance[0][3], 233 239 'false_negatives':performance[0][2]} 234 240 235 241 236 classifiers = {'boxtest':{'train':trainBoxTest, 'test':multiBoxTest}, 237 'midhinge':{'train':trainMidhinge, 'test':midhinge}} 242 def trainKalman(db, unusual_case, greater, num_observations): 243 db.resetOffsets() 244 245 def trainAux(params, num_trials): 246 estimator = functools.partial(kalmanTest, params, greater) 247 estimates = bootstrap3(estimator, db, 'train', unusual_case, num_observations, num_trials) 248 null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials) 249 250 bad_estimates = len([e for e in estimates if e != 1]) 251 bad_null_estimates = len([e for e in null_estimates if e != 0]) 252 253 false_negatives = 100.0*bad_estimates/num_trials 254 false_positives = 100.0*bad_null_estimates/num_trials 255 return false_positives,false_negatives 256 257 mean_diffs = [s['unusual_case']-s['other_cases'] for s in db.subseries('train', unusual_case)] 258 good_threshold = kfilter({},mean_diffs)['est'][-1]/2.0 259 260 wt = WorkerThreads(2, trainAux) 261 num_trials = 200 262 performance = [] 263 for t in range(90,111): 264 params = {'threshold':good_threshold*(t/100.0)} 265 wt.addJob(good_threshold*(t/100.0), (params,num_trials)) 266 wt.wait() 267 while not wt.resultq.empty(): 268 job_id,errors = wt.resultq.get() 269 fp,fn = errors 270 #performance.append(((fp+fn)/2.0, job_id, fn, fp)) 271 performance.append((abs(fp-fn), job_id, fn, fp)) 272 performance.sort() 273 #pprint.pprint(performance) 274 best_threshold = performance[0][1] 275 #print("best_threshold:", best_threshold) 276 params = {'threshold':best_threshold} 277 278 wt.stop() 279 280 return {'trial_type':"train", 281 'num_observations':num_observations, 282 'num_trials':num_trials, 283 'params':json.dumps(params), 284 'false_positives':performance[0][3], 285 'false_negatives':performance[0][2]} 286 287 288 #determine expected delta based on differences 289 classifiers = {'boxtest':{'train':trainBoxTest, 'test':multiBoxTest, 'train_results':[]}, 290 'midsummary':{'train':functools.partial(trainSummary, midsummary), 'test':midsummaryTest, 'train_results':[]}, 291 #'ubersummary':{'train':functools.partial(trainSummary, ubersummary), 'test':ubersummaryTest, 'train_results':[]}, 292 'quadsummary':{'train':functools.partial(trainSummary, quadsummary), 'test':quadsummaryTest, 'train_results':[]}, 293 'kalman':{'train':trainKalman, 'test':kalmanTest, 'train_results':[]}, 294 #'_trimean':{'train':None, 'test':trimeanTest, 'train_results':[]}, 295 } 238 296 239 297 … … 242 300 import cProfile 243 301 244 def trainClassifier(db, unusual_case, greater, trainer): 302 def trainClassifier(db, unusual_case, greater, classifier, retrain=False): 303 if retrain: 304 print("Dropping stored training results...") 305 db.deleteClassifierResults(classifier, 'train') 306 307 trainer = classifiers[classifier]['train'] 245 308 threshold = 5.0 # in percent 246 size = 4000 309 num_obs = 1000 310 max_obs = int(db.populationSize('train')/5) 247 311 result = None 248 while size < db.populationSize('train')/5: 249 size = min(size*2, int(db.populationSize('train')/5)) 250 result = trainer(db,unusual_case,greater,size) 312 while num_obs < max_obs: 313 num_obs = min(int(num_obs*1.5), max_obs) 314 result = db.fetchClassifierResult(classifier, 'train', num_obs) 315 if result != None: 316 train_time = "(stored)" 317 else: 318 start = time.time() 319 result = trainer(db,unusual_case,greater,num_obs) 320 result['classifier'] = classifier 321 train_time = "%f" % (time.time()-start) 322 251 323 error = statistics.mean([result['false_positives'],result['false_negatives']]) 252 print("subseries size: %d | error: %f | false_positives: %f | false_negatives: %f" 253 % (size,error,result['false_positives'],result['false_negatives'])) 324 print("number of observations: %d | error: %f | false_positives: %f | false_negatives: %f | train time: %s | params: %s" 325 % (num_obs, error, result['false_positives'],result['false_negatives'], train_time, result['params'])) 326 db.addClassifierResults(result) 327 classifiers[classifier]['train_results'].append(result) 328 254 329 if error < threshold: 255 330 break 256 if result != None:257 db.addClassifierResults(result)258 331 259 332 return result 333 334 335 336 def testClassifier(db, unusual_case, greater, classifier, retest=False): 337 target_error = 5.0 # in percent 338 num_trials = 1000 339 max_obs = int(db.populationSize('test')/5) 340 341 tester = classifiers[classifier]['test'] 342 343 def testAux(params, num_trials, num_observations): 344 estimator = functools.partial(tester, params, greater) 345 estimates = bootstrap3(estimator, db, 'test', unusual_case, num_observations, num_trials) 346 null_estimates = bootstrap3(estimator, db, 'train_null', unusual_case, num_observations, num_trials) 347 348 bad_estimates = len([e for e in estimates if e != 1]) 349 bad_null_estimates = len([e for e in null_estimates if e != 0]) 350 351 false_negatives = 100.0*bad_estimates/num_trials 352 false_positives = 100.0*bad_null_estimates/num_trials 353 print("testAux:", num_observations, false_positives, false_negatives, params) 354 return false_positives,false_negatives 355 356 357 if retest: 358 print("Dropping stored test results...") 359 db.deleteClassifierResults(classifier, 'test') 360 361 362 test_results = [] 363 lte = math.log(target_error/100.0) 364 for tr in classifiers[classifier]['train_results']: 365 db.resetOffsets() 366 params = json.loads(tr['params']) 367 num_obs = tr['num_observations'] 368 369 print("initial test") 370 fp,fn = testAux(params, num_trials, num_obs) 371 error = (fp+fn)/2.0 372 print("walking up") 373 while (error > target_error) and (num_obs < max_obs): 374 increase_factor = 1.5 * lte/math.log(error/100.0) # don't ask how I came up with this 375 #print("increase_factor:", increase_factor) 376 num_obs = min(int(increase_factor*num_obs), max_obs) 377 fp,fn = testAux(params, num_trials, num_obs) 378 error = (fp+fn)/2.0 379 380 print("walking down") 381 while (num_obs > 0): 382 current_best = (num_obs,error,params,fp,fn) 383 num_obs = int(0.95*num_obs) 384 fp,fn = testAux(params, num_trials, num_obs) 385 error = (fp+fn)/2.0 386 if error > target_error: 387 break 388 389 test_results.append(current_best) 390 391 test_results.sort() 392 best_obs,error,best_params,fp,fn = test_results[0] 393 394 return {'classifier':classifier, 395 'trial_type':"test", 396 'num_observations':best_obs, 397 'num_trials':num_trials, 398 'params':best_params, 399 'false_positives':fp, 400 'false_negatives':fn} 260 401 261 402 … … 268 409 print(":", end-start) 269 410 270 271 for c,funcs in classifiers.items(): 411 for c in sorted(classifiers.keys()): 412 if classifiers[c]['train'] == None: 413 continue 272 414 start = time.time() 273 415 print("Training %s..." % c) 274 result = trainClassifier(db, unusual_case, greater, funcs['train'])416 result = trainClassifier(db, unusual_case, greater, c, c in options.retrain) 275 417 print("%s result:" % c) 276 418 pprint.pprint(result) 277 419 print("completed in:", time.time()-start) 278 420 279 sys.exit(0) 280 281 start = time.time() 282 results = trainBoxTest(db, unusual_case, greater, 6000) 283 #db.addClassifierResults(results) 284 print("multi box test result:") 285 pprint.pprint(results) 286 print(":", time.time()-start) 421 db.clearCache() 422 423 for c in sorted(classifiers.keys()): 424 start = time.time() 425 print("Testing %s..." % c) 426 result = testClassifier(db, unusual_case, greater, c, c in options.retest) 427 print("%s result:" % c) 428 pprint.pprint(result) 429 classifiers[c]['test_error'] = (result['false_positives']+result['false_negatives'])/2.0 430 print("completed in:", time.time()-start)
Note: See TracChangeset
for help on using the changeset viewer.