Mercurial > ift6266

comparison deep/stacked_dae/v_sylvain/sgd_optimization.py @ 384:8117c0e70db9

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Ajout de la faculte de prendre PNIST07
author SylvainPL <sylvain.pannetier.lebeuf@umontreal.ca>
date Tue, 27 Apr 2010 08:20:37 -0400
parents f24b10e43a6f
children 88cb95007670
comparison
equal deleted inserted replaced
383:5c3935aa3f8a 384:8117c0e70db9
154 f = open('params_pretrain.txt', 'w') 154 f = open('params_pretrain.txt', 'w')
155 cPickle.dump(self.parameters_pre,f,protocol=-1) 155 cPickle.dump(self.parameters_pre,f,protocol=-1)
156 f.close() 156 f.close()
157 157
158 158
159 def finetune(self,dataset,dataset_test,num_finetune,ind_test,special=0,decrease=0): 159 def finetune(self,dataset,dataset_test,num_finetune,ind_test,special=0,decrease=0,dataset_test2=None):
160 160
161 if special != 0 and special != 1: 161 if special != 0 and special != 1:
162 sys.exit('Bad value for variable special. Must be in {0,1}') 162 sys.exit('Bad value for variable special. Must be in {0,1}')
163 print "STARTING FINETUNING, time = ", datetime.datetime.now() 163 print "STARTING FINETUNING, time = ", datetime.datetime.now()
164 164
165 minibatch_size = self.hp.minibatch_size 165 minibatch_size = self.hp.minibatch_size
166 if ind_test == 0 or ind_test == 20: 166 if ind_test == 0 or ind_test == 20:
167 nom_test = "NIST" 167 nom_test = "NIST"
168 nom_train="P07" 168 nom_train="P07"
169 elif ind_test == 2:
170 nom_train = "PNIST07"
171 nom_test = "NIST"
172 nom_test2 = "P07"
169 else: 173 else:
170 nom_test = "P07" 174 nom_test = "P07"
171 nom_train = "NIST" 175 nom_train = "NIST"
172 176
173 177
236 self.series["training_error"].append((epoch, minibatch_index), cost_ij) 240 self.series["training_error"].append((epoch, minibatch_index), cost_ij)
237 241
238 if (total_mb_index+1) % validation_frequency == 0: 242 if (total_mb_index+1) % validation_frequency == 0:
239 #minibatch_index += 1 243 #minibatch_index += 1
240 #The validation set is always NIST (we want the model to be good on NIST) 244 #The validation set is always NIST (we want the model to be good on NIST)
241 if ind_test == 0 | ind_test == 20: 245 if ind_test == 0 | ind_test == 20 | ind_test == 2:
242 iter=dataset_test.valid(minibatch_size,bufsize=buffersize) 246 iter=dataset_test.valid(minibatch_size,bufsize=buffersize)
243 else: 247 else:
244 iter = dataset.valid(minibatch_size,bufsize=buffersize) 248 iter = dataset.valid(minibatch_size,bufsize=buffersize)
245 if self.max_minibatches: 249 if self.max_minibatches:
246 iter = itermax(iter, self.max_minibatches) 250 iter = itermax(iter, self.max_minibatches)
247 validation_losses = [validate_model(x,y) for x,y in iter] 251 validation_losses = [validate_model(x,y) for x,y in iter]
279 iter2 = dataset_test.test(minibatch_size,bufsize=buffersize) 283 iter2 = dataset_test.test(minibatch_size,bufsize=buffersize)
280 if self.max_minibatches: 284 if self.max_minibatches:
281 iter2 = itermax(iter2, self.max_minibatches) 285 iter2 = itermax(iter2, self.max_minibatches)
282 test_losses2 = [test_model(x,y) for x,y in iter2] 286 test_losses2 = [test_model(x,y) for x,y in iter2]
283 test_score2 = numpy.mean(test_losses2) 287 test_score2 = numpy.mean(test_losses2)
288
289 #test it on the third test set if there is one
290 iter3 = dataset_test2.test(minibatch_size, bufsize=buffersize)
291 if self.max_minibatches:
292 iter3 = itermax(iter3, self.max_minibatches)
293 test_losses3 = [test_model(x,y) for x,y in iter3]
294 test_score3 = numpy.mean(test_losses3)
284 295
285 self.series["test_error"].\ 296 self.series["test_error"].\
286 append((epoch, minibatch_index), test_score*100.) 297 append((epoch, minibatch_index), test_score*100.)
287 298
288 print((' epoch %i, minibatch %i, test error on dataset %s (train data) of best ' 299 print((' epoch %i, minibatch %i, test error on dataset %s (train data) of best '
292 303
293 print((' epoch %i, minibatch %i, test error on dataset %s of best ' 304 print((' epoch %i, minibatch %i, test error on dataset %s of best '
294 'model %f %%') % 305 'model %f %%') %
295 (epoch, minibatch_index+1,nom_test, 306 (epoch, minibatch_index+1,nom_test,
296 test_score2*100.)) 307 test_score2*100.))
308 print((' epoch %i, minibatch %i, test error on dataset %s of best '
309 'model %f %%') %
310 (epoch, minibatch_index+1,nom_test2,
311 test_score3*100.))
297 312
298 if patience <= total_mb_index: 313 if patience <= total_mb_index:
299 done_looping = True 314 done_looping = True
300 break #to exit the FOR loop 315 break #to exit the FOR loop
301 316
304 # useful when doing tests 319 # useful when doing tests
305 if self.max_minibatches and minibatch_index >= self.max_minibatches: 320 if self.max_minibatches and minibatch_index >= self.max_minibatches:
306 break 321 break
307 322
308 if decrease == 1: 323 if decrease == 1:
309 if (ind_test == 21 & epoch % 100 == 0) | ind_test == 20: 324 if (ind_test == 21 & epoch % 100 == 0) | ind_test == 20 | ind_test == 2:
310 learning_rate /= 2 #divide the learning rate by 2 for each new epoch of P07 (or 100 of NIST) 325 learning_rate /= 2 #divide the learning rate by 2 for each new epoch of P07 (or 100 of NIST)
311 326
312 self.series['params'].append((epoch,), self.classifier.all_params) 327 self.series['params'].append((epoch,), self.classifier.all_params)
313 328
314 if done_looping == True: #To exit completly the fine-tuning 329 if done_looping == True: #To exit completly the fine-tuning
322 337
323 print(('\nOptimization complete with best validation score of %f %%,' 338 print(('\nOptimization complete with best validation score of %f %%,'
324 'with test performance %f %% on dataset %s ') % 339 'with test performance %f %% on dataset %s ') %
325 (best_validation_loss * 100., test_score*100.,nom_train)) 340 (best_validation_loss * 100., test_score*100.,nom_train))
326 print(('The test score on the %s dataset is %f')%(nom_test,test_score2*100.)) 341 print(('The test score on the %s dataset is %f')%(nom_test,test_score2*100.))
342 print(('The test score on the %s dataset is %f')%(nom_test2,test_score3*100.))
327 343
328 print ('The finetuning ran for %f minutes' % ((end_time-start_time)/60.)) 344 print ('The finetuning ran for %f minutes' % ((end_time-start_time)/60.))
329 345
330 sys.stdout.flush() 346 sys.stdout.flush()
331 347
347 cPickle.dump(parameters_finetune,f,protocol=-1) 363 cPickle.dump(parameters_finetune,f,protocol=-1)
348 f.close() 364 f.close()
349 365
350 elif ind_test== 21: #To keep a track of the value of the parameters 366 elif ind_test== 21: #To keep a track of the value of the parameters
351 f = open('params_finetune_P07_then_NIST.txt', 'w') 367 f = open('params_finetune_P07_then_NIST.txt', 'w')
368 cPickle.dump(parameters_finetune,f,protocol=-1)
369 f.close()
370 elif ind_test == 2:
371 f = open('params_finetune_PNIST07.txt', 'w')
352 cPickle.dump(parameters_finetune,f,protocol=-1) 372 cPickle.dump(parameters_finetune,f,protocol=-1)
353 f.close() 373 f.close()
354 374
355 375
356 #Set parameters like they where right after pre-train or finetune 376 #Set parameters like they where right after pre-train or finetune
374 [self.classifier.x,self.classifier.y], self.classifier.errors) 394 [self.classifier.x,self.classifier.y], self.classifier.errors)
375 395
376 iter2 = dataset.train(self.hp.minibatch_size,bufsize=buffersize) 396 iter2 = dataset.train(self.hp.minibatch_size,bufsize=buffersize)
377 train_losses2 = [test_model(x,y) for x,y in iter2] 397 train_losses2 = [test_model(x,y) for x,y in iter2]
378 train_score2 = numpy.mean(train_losses2) 398 train_score2 = numpy.mean(train_losses2)
379 print "Training error is: " + str(train_score2) 399 print(('The training error is %f')%(train_score2*100.))
380 400
381 #To see the prediction of the model, the real answer and the image to judge 401 #To see the prediction of the model, the real answer and the image to judge
382 def see_error(self, dataset): 402 def see_error(self, dataset):
383 import pylab 403 import pylab
384 #The function to know the prediction 404 #The function to know the prediction