Fix re: saving/loading models with frozen layers.

keras/engine/topology.py

@@ -299,10 +299,10 @@ def __init__(self, **kwargs):
 
         # These properties will be set upon call of self.build(),
         # which itself will be called upon self.add_inbound_node if necessary.
-        if not hasattr(self, 'trainable_weights'):
-            self.trainable_weights = []
-        if not hasattr(self, 'non_trainable_weights'):
-            self.non_trainable_weights = []
+        if not hasattr(self, '_trainable_weights'):
+            self._trainable_weights = []
+        if not hasattr(self, '_non_trainable_weights'):
+            self._non_trainable_weights = []
         if not hasattr(self, 'losses'):
             self.losses = []
         if not hasattr(self, 'constraints'):
@@ -418,9 +418,9 @@ def add_weight(self, shape, initializer, name=None,
         if constraint is not None:
             self.constraints[weight] = constraint
         if trainable:
-            self.trainable_weights.append(weight)
+            self._trainable_weights.append(weight)
         else:
-            self.non_trainable_weights.append(weight)
+            self._non_trainable_weights.append(weight)
         return weight
 
     def assert_input_compatibility(self, input):
@@ -1053,16 +1053,11 @@ def __init__(self, input_shape=None, batch_input_shape=None,
         self.uses_learning_phase = False
         self.trainable = False
         self.built = True
-        self.trainable_weights = []
-        self.non_trainable_weights = []
-
+        self._trainable_weights = []
+        self._non_trainable_weights = []
         self.inbound_nodes = []
         self.outbound_nodes = []
-
-        self.trainable_weights = []
-        self.non_trainable_weights = []
         self.constraints = {}
-
         self.sparse = sparse
 
         if not name:
@@ -1269,8 +1264,8 @@ def __init__(self, layers=None, mode='sum', concat_axis=-1,
         self.inbound_nodes = []
         self.outbound_nodes = []
         self.constraints = {}
-        self.trainable_weights = []
-        self.non_trainable_weights = []
+        self._trainable_weights = []
+        self._non_trainable_weights = []
         self.supports_masking = True
         self.uses_learning_phase = False
         self.input_spec = None  # Compatible with anything.

keras/models.py

@@ -153,7 +153,8 @@ def deserialize(obj):
         return model
     training_config = json.loads(training_config.decode('utf-8'))
     optimizer_config = training_config['optimizer_config']
-    optimizer = optimizer_from_config(optimizer_config, custom_objects=custom_objects)
+    optimizer = optimizer_from_config(optimizer_config,
+                                      custom_objects=custom_objects)
 
     # recover loss functions and metrics
     loss = deserialize(training_config['loss'])

keras/utils/layer_utils.py

@@ -36,7 +36,8 @@ def layer_from_config(config, custom_objects=None):
 
     arg_spec = inspect.getargspec(layer_class.from_config)
     if 'custom_objects' in arg_spec.args:
-        return layer_class.from_config(config['config'], custom_objects=custom_objects)
+        return layer_class.from_config(config['config'],
+                                       custom_objects=custom_objects)
     else:
         return layer_class.from_config(config['config'])