# HG changeset patch
# User Xavier Glorot <glorotxa@iro.umontreal.ca>
# Date 1250177967 14400
# Node ID ecae21a7262ef03d0645e1c8cd4b4ae508c6f758
# Parent  3a4bc4a0dbf4ff9bdf1ade2c4a0d8a4be6eea566
Added a rescalwsaturation instance method for DAA inputs groups

diff -r 3a4bc4a0dbf4 -r ecae21a7262e pylearn/algorithms/sandbox/DAA_inputs_groups.py
--- a/pylearn/algorithms/sandbox/DAA_inputs_groups.py	Wed Aug 12 19:19:24 2009 -0400
+++ b/pylearn/algorithms/sandbox/DAA_inputs_groups.py	Thu Aug 13 11:39:27 2009 -0400
@@ -540,7 +540,7 @@
         
         # facultative methods
         if self.debugmethod:
-            self.activation = [None] * (self.depth)
+            self.activation = [None] * (self.depth+1)
             self.representation = [None] * (self.depth)
             self.recactivation = [None] * (self.depth)
             self.reconstruction = [None] * (self.depth)
@@ -706,6 +706,7 @@
         self.NLL = theano.Method(self.inputs[-1],self.daaig[-1]._xent)
         
         if self.debugmethod:
+            self.activation[-1] = theano.Method(self.inputs[-2],self.daaig[-1].linear_output)
             self.compute_localcost[-1] = theano.Method(self.inputs[-1],self.localcost[-1])
             self.compute_localgradients[-1] = theano.Method(self.inputs[-1],self.localgradients[-1])
             self.compute_globalcost[-1] = theano.Method(self.inputs[-1],self.globalcost[-1])
@@ -827,6 +828,23 @@
     def _instance_nll(self,inst,inputs,target):
         return numpy.sum(inst.NLL(*(inputs+[target]))) / float(len(target))
     
+    #try--------------------------------------------------------------------
+    def _instance_rescalwsaturation(self,inst,inputs):
+        sat = [None]*(self.depth+1)
+        for i in range(self.depth+1):
+            sat[i] = inst.hidsaturation(i,inputs[min(i,self.depth)])
+        
+        for i in range(self.depth-1):
+            if sat[i+1] > max(sat[:i+1]):
+                inst.daaig[i+1].wenc = inst.daaig[i+1].wenc/sat[i+1]*max(sat[:i+1])
+                inst.daaig[i+1].benc = inst.daaig[i+1].benc/sat[i+1]*max(sat[:i+1])
+                sat[i+1] = max(sat[:i+1])
+        if sat[-1]>max(sat[:-1]):
+            inst.daaig[-1].w = inst.daaig[i+1].w/sat[-1]*max(sat[:-1])
+            inst.daaig[-1].b = inst.daaig[i+1].b/sat[-1]*max(sat[:-1])
+
+    #-----------------------------------------------------------------------
+    
     def _instance_unsupgrad(self,inst,inputs,layer,param_name):
         inst.noiseseed(0)
         gradin = inst.compute_localgradients_in[layer](*inputs)