Gan

Ok, so I think that the modifications I made here can be merged ...

Two main changes as compared to the master:

1. Added some layers
2. Modified the way variable name and scope is handled in layers (and in SequenceNetwork as well)

Change 1 should by no means affect what has been done until now (on the master). However, change 2 may require a bit of adaptation, especially when defining architectures and layers ...

@tiago.pereira I'll let you check and then we can discuss

Also, I think that this will be irrelevant when we'll go to tf-slim anyway !

bob/learn/tensorflow/layers/Layer.py

@@ -76,8 +76,18 @@ class Layer(object):
             phase_train:
 
         """
-
-        return tf.contrib.layers.batch_norm(x, decay=1.9, updates_collections=None, epsilon=1e-5, scale=True, is_training=phase_train, scope=scope)
+        
+        # XXX THAT'S UGLY ...
+        reuse = False
+        for var in tf.global_variables():
+          if scope in var.name and ('beta' in var.name or 'gamma' in var.name):
+            print "{0} already exists".format(var.name)
+            reuse = True 
+
+        if reuse:
+          return tf.contrib.layers.batch_norm(x, decay=0.9, updates_collections=None, epsilon=1e-5, scale=True, scope=scope, is_training=phase_train, reuse=True)
+        else:  
+          return tf.contrib.layers.batch_norm(x, decay=0.9, updates_collections=None, epsilon=1e-5, scale=True, scope=scope, is_training=phase_train, reuse=None)
 
         from tensorflow.python.ops import control_flow_ops