# HG changeset patch
# User SylvainPL <sylvain.pannetier.lebeuf@umontreal.ca>
# Date 1271029972 14400
# Node ID 18dc860a4ef4fd8cd915c1c711a3fc705525810c
# Parent  54ad8a091783a04005d3ee0efd1963613712ecf5
Rajout de la capacite de faire decroitre le taux d'apprentissage si demande

diff -r 54ad8a091783 -r 18dc860a4ef4 deep/stacked_dae/v_sylvain/sgd_optimization.py
--- a/deep/stacked_dae/v_sylvain/sgd_optimization.py	Sun Apr 11 19:52:44 2010 -0400
+++ b/deep/stacked_dae/v_sylvain/sgd_optimization.py	Sun Apr 11 19:52:52 2010 -0400
@@ -145,7 +145,7 @@
         f.close()
 
 
-    def finetune(self,dataset,dataset_test,num_finetune,ind_test,special=0):
+    def finetune(self,dataset,dataset_test,num_finetune,ind_test,special=0,decrease=0):
         
         if special != 0 and special != 1:
             sys.exit('Bad value for variable special. Must be in {0,1}')
@@ -200,16 +200,20 @@
         epoch = 0
 
         total_mb_index = 0
-        minibatch_index = -1
+        minibatch_index = 0
         parameters_finetune=[]
+        learning_rate = self.hp.finetuning_lr  #The initial finetune lr
+
 
         while (epoch < num_finetune) and (not done_looping):
             epoch = epoch + 1
 
             for x,y in dataset.train(minibatch_size,bufsize=buffersize):
                 minibatch_index += 1
+                
+                
                 if special == 0:
-                    cost_ij = self.classifier.finetune(x,y)
+                    cost_ij = self.classifier.finetune(x,y,learning_rate)
                 elif special == 1:
                     cost_ij = self.classifier.finetune2(x,y)
                 total_mb_index += 1
@@ -285,7 +289,10 @@
                 # useful when doing tests
                 if self.max_minibatches and minibatch_index >= self.max_minibatches:
                     break
-
+            
+            if decrease == 1:
+                learning_rate /= 2 #divide the learning rate by 2 for each new epoch
+            
             self.series['params'].append((epoch,), self.classifier.all_params)
 
             if done_looping == True:    #To exit completly the fine-tuning
@@ -341,21 +348,19 @@
                 self.classifier.params[idx].value=theano._asarray(copy(x),dtype=theano.config.floatX)
             else:
                 self.classifier.params[idx].value=copy(x)
-                
-    #Calculate error over the training set (or a part of)           
-    def training_error(self,data):
+
+    def training_error(self,dataset):
         # create a function to compute the mistakes that are made by the model
         # on the validation set, or testing set
         test_model = \
             theano.function(
                 [self.classifier.x,self.classifier.y], self.classifier.errors)
                 
-        iter2 = data.train(self.hp.minibatch_size,bufsize=buffersize)
+        iter2 = dataset.train(self.hp.minibatch_size,bufsize=buffersize)
         train_losses2 = [test_model(x,y) for x,y in iter2]
         train_score2 = numpy.mean(train_losses2)
         print "Training error is: " + str(train_score2)
 
-