# HG changeset patch
# User fsavard
# Date 1268408856 18000
# Node ID 851e7ad4a1430234c237a4ce0c11f6ffcb84d3e8
# Parent  acae439d6572160dfc84c607faef670553a223cc
Corrigé une erreur dans la formule de coût modifiée dans stacked_dae, et enlevé des timers dans sgd_optimization

diff -r acae439d6572 -r 851e7ad4a143 deep/stacked_dae/v2/nist_sda.py
--- a/deep/stacked_dae/v2/nist_sda.py	Fri Mar 12 10:31:10 2010 -0500
+++ b/deep/stacked_dae/v2/nist_sda.py	Fri Mar 12 10:47:36 2010 -0500
@@ -35,8 +35,8 @@
 TEST_CONFIG = False
 
 NIST_ALL_LOCATION = '/data/lisa/data/nist/by_class/all'
-JOBDB = 'postgres://ift6266h10@gershwin/ift6266h10_db/fsavard_sda4'
-EXPERIMENT_PATH = "ift6266.deep.stacked_dae.nist_sda.jobman_entrypoint"
+JOBDB = 'postgres://ift6266h10@gershwin/ift6266h10_sandbox_db/fsavard_sda_v2'
+EXPERIMENT_PATH = "ift6266.deep.stacked_dae.v2.nist_sda.jobman_entrypoint"
 
 REDUCE_TRAIN_TO = None
 MAX_FINETUNING_EPOCHS = 1000
@@ -74,8 +74,7 @@
 
 '''
 Function called by jobman upon launching each job
-Its path is the one given when inserting jobs:
-ift6266.deep.stacked_dae.nist_sda.jobman_entrypoint
+Its path is the one given when inserting jobs: see EXPERIMENT_PATH
 '''
 def jobman_entrypoint(state, channel):
     # record mercurial versions of each package
diff -r acae439d6572 -r 851e7ad4a143 deep/stacked_dae/v2/sgd_optimization.py
--- a/deep/stacked_dae/v2/sgd_optimization.py	Fri Mar 12 10:31:10 2010 -0500
+++ b/deep/stacked_dae/v2/sgd_optimization.py	Fri Mar 12 10:47:36 2010 -0500
@@ -104,9 +104,6 @@
         print "STARTING PRETRAINING, time = ", datetime.datetime.now()
         sys.stdout.flush()
 
-        time_acc_func = 0.0
-        time_acc_total = 0.0
-
         start_time = time.clock()  
         ## Pre-train layer-wise 
         for i in xrange(self.classifier.n_layers):
@@ -114,14 +111,7 @@
             for epoch in xrange(self.hp.pretraining_epochs_per_layer):
                 # go through the training set
                 for batch_index in xrange(self.n_train_batches):
-                    t1 = time.clock()
                     c = self.classifier.pretrain_functions[i](batch_index)
-                    t2 = time.clock()
-
-                    time_acc_func += t2 - t1
-
-                    if batch_index % 500 == 0:
-                        print "acc / total", time_acc_func / (t2 - start_time), time_acc_func
 
                     self.series["reconstruction_error"].append((epoch, batch_index), c)
                         
diff -r acae439d6572 -r 851e7ad4a143 deep/stacked_dae/v2/stacked_dae.py
--- a/deep/stacked_dae/v2/stacked_dae.py	Fri Mar 12 10:31:10 2010 -0500
+++ b/deep/stacked_dae/v2/stacked_dae.py	Fri Mar 12 10:47:36 2010 -0500
@@ -133,7 +133,7 @@
     #         used later when stacking dAs. 
     self.y   = T.nnet.sigmoid(T.dot(self.tilde_x, self.W      ) + self.b)
     # Equation (3)
-    self.z   = T.nnet.sigmoid(T.dot(self.y, self.W_prime) + self.b_prime)
+    #self.z   = T.nnet.sigmoid(T.dot(self.y, self.W_prime) + self.b_prime)
     # Equation (4)
     # note : we sum over the size of a datapoint; if we are using minibatches,
     #        L will  be a vector, with one entry per example in minibatch
@@ -142,9 +142,9 @@
 
     # bypassing z to avoid running to log(0)
     z_a = T.dot(self.y, self.W_prime) + self.b_prime
-    log_sigmoid = T.log(1) - T.log(1+T.exp(-z_a))
+    log_sigmoid = T.log(1.) - T.log(1.+T.exp(-z_a))
     # log(1-sigmoid(z_a))
-    log_1_sigmoid = -self.x - T.log(1+T.exp(-z_a))
+    log_1_sigmoid = -self.z_a - T.log(1.+T.exp(-z_a))
     self.L = -T.sum( self.x * (log_sigmoid) \
                     + (1.0-self.x) * (log_1_sigmoid), axis=1 )