# HG changeset patch
# User SylvainPL <sylvain.pannetier.lebeuf@umontreal.ca>
# Date 1272039123 14400
# Node ID d391ad815d89c97669a3414bad9d41d81892bed3
# Parent  f24b10e43a6f7df1248184e766b24ef92a2c4d36
Correction d'un bug avec la fonction de log-likelihood pour utilisation de non-linearite de sortie sigmoides

diff -r f24b10e43a6f -r d391ad815d89 deep/stacked_dae/v_sylvain/stacked_dae.py
--- a/deep/stacked_dae/v_sylvain/stacked_dae.py	Fri Apr 23 11:39:55 2010 -0400
+++ b/deep/stacked_dae/v_sylvain/stacked_dae.py	Fri Apr 23 12:12:03 2010 -0400
@@ -28,8 +28,8 @@
         self.b = theano.shared( value=numpy.zeros((n_out,), 
                                             dtype = theano.config.floatX) )
         # compute vector of class-membership. This is a sigmoid instead of
-        #a softmax to be able to classify as nothing later
-##        self.p_y_given_x = T.nnet.softmax(T.dot(input, self.W)+self.b)
+        #a softmax to be able later to classify as nothing
+##        self.p_y_given_x = T.nnet.softmax(T.dot(input, self.W)+self.b) #row-wise
         self.p_y_given_x = T.nnet.sigmoid(T.dot(input, self.W)+self.b)
         
         # compute prediction as class whose probability is maximal in 
@@ -41,7 +41,13 @@
         
 
     def negative_log_likelihood(self, y):
-       return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]),y])
+##        return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]),y])
+        return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]),y]+T.sum(T.log(1-self.p_y_given_x), axis=1)-T.log(1-self.p_y_given_x)[T.arange(y.shape[0]),y])
+    
+    
+##    def kullback_leibler(self,y):
+##        return -T.mean(T.log(1/float(self.p_y_given_x))[T.arange(y.shape[0]),y])
+
 
     def errors(self, y):
         # check if y has same dimension of y_pred 
@@ -187,15 +193,14 @@
     #changed so it remain the same as when sigmoid is used
     self.y   = (T.tanh(T.dot(self.tilde_x, self.W ) + self.b)+1.0)/2.0
     
-    z_a = T.dot(self.y, self.W_prime) + self.b_prime
-    self.z =  (T.tanh(z_a )+1.0) / 2.0
+    self.z =  (T.tanh(T.dot(self.y, self.W_prime) + self.b_prime)+1.0) / 2.0
     #To ensure to do not have a log(0) operation
     if self.z <= 0:
         self.z = 0.000001
     if self.z >= 1:
         self.z = 0.999999
         
-    self.L = - T.sum( self.x*T.log(self.z) + (1-self.x)*T.log(1-self.z), axis=1 )
+    self.L = - T.sum( self.x*T.log(self.z) + (1.0-self.x)*T.log(1.0-self.z), axis=1 )
     
     self.cost = T.mean(self.L)