I rerun the Sda tests as a sanity check to all the changes that have been made .. and updated the results. I did not get the time though to look into running the SdA with parameters such that you get relly good result (like the ones used by Isabelle )

Razvan Pascanu · Razvan Pascanu · commit 3a3fb4d60cb3 · 2010-02-24T08:35:48.000-05:00
diff --git a/code/SdA.py b/code/SdA.py
@@ -279,7 +279,7 @@ def test_score():
 
 
 
-def test_SdA( finetune_lr = 0.1, pretraining_epochs = 2, \
+def test_SdA( finetune_lr = 0.1, pretraining_epochs = 15, \
               pretrain_lr = 0.1, training_epochs = 1000, \
               dataset='mnist.pkl.gz'):
     """
@@ -322,7 +322,7 @@ def test_SdA( finetune_lr = 0.1, pretraining_epochs = 2, \
     print '... building the model'
     # construct the stacked denoising autoencoder class
     sda = SdA( numpy_rng = numpy_rng, n_ins = 28*28, 
-                      hidden_layers_sizes = [100,100,100],
+                      hidden_layers_sizes = [1000,1000,1000],
                       n_outs = 10)
     
 
diff --git a/doc/SdA.txt b/doc/SdA.txt
@@ -427,9 +427,9 @@ The user can run the code by calling:
   python code/SdA.py
 
 By default the code runs 15 pre-training epochs for each layer, with 
-a corruption level of 0.1 and a learning rate of 0.1. Pre-training takes
-78.88 minutes. Fine-tuning is completed after 32 epochs in 65.89 
-minutes and results  in a validation score of 1.7 %, with a test 
-performace of 1.65 %.
+a corruption level of 0.2 and a learning rate of 0.1. Pre-training takes
+80.63 minutes. Fine-tuning is completed after 48 epochs in 97.18 
+minutes and results  in a validation score of 1.63 %, with a test 
+performace of 1.68 %.
 
 
