comparison scripts/deepmlp.py @ 22:cb47cbc95a21

I fixed a bug in the computation of L1 and L2 regularizations
author Razvan Pascanu <r.pascanu@gmail.com>
date Fri, 29 Jan 2010 11:01:39 -0500
parents afdd41db8152
children
comparison
equal deleted inserted replaced
21:afdd41db8152 22:cb47cbc95a21
100 # symbolic form 100 # symbolic form
101 self.y_pred = T.argmax( self.p_y_given_x, axis =1) 101 self.y_pred = T.argmax( self.p_y_given_x, axis =1)
102 102
103 # L1 norm ; one regularization option is to enforce L1 norm to 103 # L1 norm ; one regularization option is to enforce L1 norm to
104 # be small 104 # be small
105 self.L1=abs(self.W[i]).sum() 105 self.L1=abs(self.W[0]).sum()
106 self.L2_sqr=abs(self.W[i]).sum() 106 self.L2_sqr=abs(self.W[0]).sum()
107 for i in range(1,n_layer+1): 107 for i in range(1,n_layer+1):
108 self.L1 += abs(self.W[i]).sum() 108 self.L1 += abs(self.W[i]).sum()
109 # square of L2 norm ; one regularization option is to enforce 109 # square of L2 norm ; one regularization option is to enforce
110 # square of L2 norm to be small 110 # square of L2 norm to be small
111 for i in range(n_layer+1): 111 for i in range(n_layer+1):