ift6266: deep/stacked_dae/v_sylvain/sgd_optimization.py comparison

comparison deep/stacked_dae/v_sylvain/sgd_optimization.py @ 305:fe5d428c2acc

Ajout de commentaires ainsi que correction de bug pour sauvegarder les parametres

author	SylvainPL <sylvain.pannetier.lebeuf@umontreal.ca>
date	Wed, 31 Mar 2010 21:00:59 -0400
parents	1cc535f3e254
children	a5abd5c8b4b0

comparison

equal deleted inserted replaced

-:1e4bf5a5b46d
+:fe5d428c2acc
 def pretrain(self,dataset):
 print "STARTING PRETRAINING, time = ", datetime.datetime.now()
 sys.stdout.flush()
-un_fichier=int(819200.0/self.hp.minibatch_size) #Number of batches in a P07 batch
+un_fichier=int(819200.0/self.hp.minibatch_size) #Number of batches in a P07 file
 start_time = time.clock()
 ## Pre-train layer-wise
 for i in xrange(self.classifier.n_layers):
 # go through pretraining epochs
 done_looping = False
 epoch = 0
 total_mb_index = 0
 minibatch_index = -1
+parameters_finetune=[]
 while (epoch < num_finetune) and (not done_looping):
 epoch = epoch + 1
 for x,y in dataset.train(minibatch_size):
 self.series["training_error"].append((epoch, minibatch_index), cost_ij)
 if (total_mb_index+1) % validation_frequency == 0:
 #minibatch_index += 1
-#The validation set is always NIST
+#The validation set is always NIST (we want the model to be good on NIST)
-if ind_test == 0:
+if ind_test == 0 | ind_test == 20:
 iter=dataset_test.valid(minibatch_size)
 else:
 iter = dataset.valid(minibatch_size)
 if self.max_minibatches:
 iter = itermax(iter, self.max_minibatches)
 this_validation_loss = numpy.mean(validation_losses)
 self.series["validation_error"].\
 append((epoch, minibatch_index), this_validation_loss*100.)
-print('epoch %i, minibatch %i, validation error on %s : %f %%' % \
+print('epoch %i, minibatch %i, validation error on NIST : %f %%' % \
-(epoch, minibatch_index+1,nom_test, \
+(epoch, minibatch_index+1, \
 this_validation_loss*100.))
 # if we got the best validation score until now
 if this_validation_loss < best_validation_loss:
 #improve patience if loss improvement is good enough
 if this_validation_loss < best_validation_loss *  \
 improvement_threshold :
 patience = max(patience, total_mb_index * patience_increase)
-# save best validation score and iteration number
+# save best validation score, iteration number and parameters
 best_validation_loss = this_validation_loss
 best_iter = total_mb_index
+parameters_finetune=[copy(x.value) for x in self.classifier.params]
 # test it on the test set
 iter = dataset.test(minibatch_size)
 if self.max_minibatches:
 iter = itermax(iter, self.max_minibatches)
 (epoch, minibatch_index+1,nom_test,
 test_score2*100.))
 if patience <= total_mb_index:
 done_looping = True
-break
+break   #to exit the FOR loop
 sys.stdout.flush()
 # useful when doing tests
 if self.max_minibatches and minibatch_index >= self.max_minibatches:
 break
 self.series['params'].append((epoch,), self.classifier.all_params)
 if done_looping == True:    #To exit completly the fine-tuning
-break
+break   #to exit the WHILE loop
 end_time = time.clock()
 self.hp.update({'finetuning_time':end_time-start_time,\
 'best_validation_error':best_validation_loss,\
 'test_score':test_score,
 (best_validation_loss * 100., test_score*100.,nom_train))
 print(('The test score on the %s dataset is %f')%(nom_test,test_score2*100.))
 print ('The finetuning ran for %f minutes' % ((end_time-start_time)/60.))
+sys.stdout.flush()
 #Save a copy of the parameters in a file to be able to get them in the future
 if special == 1:    #To keep a track of the value of the parameters
-parameters_finetune=[copy(x.value) for x in self.classifier.params]
 f = open('params_finetune_stanford.txt', 'w')
 pickle.dump(parameters_finetune,f)
 f.close()
 elif ind_test== 0:    #To keep a track of the value of the parameters
-parameters_finetune=[copy(x.value) for x in self.classifier.params]
 f = open('params_finetune_P07.txt', 'w')
 pickle.dump(parameters_finetune,f)
 f.close()
 elif ind_test== 1:    #For the run with 2 finetunes. It will be faster.
-parameters_finetune=[copy(x.value) for x in self.classifier.params]
 f = open('params_finetune_NIST.txt', 'w')
 pickle.dump(parameters_finetune,f)
 f.close()
 elif ind_test== 20:    #To keep a track of the value of the parameters
-parameters_finetune=[copy(x.value) for x in self.classifier.params]
 f = open('params_finetune_NIST_then_P07.txt', 'w')
 pickle.dump(parameters_finetune,f)
 f.close()
-#Set parameters like they where right after pre-train
+#Set parameters like they where right after pre-train or finetune
 def reload_parameters(self,which):
 #self.parameters_pre=pickle.load('params_pretrain.txt')
 f = open(which)
 self.parameters_pre=pickle.load(f)

Mercurial > ift6266

comparison deep/stacked_dae/v_sylvain/sgd_optimization.py @ 305:fe5d428c2acc