[keras] Layers: fixed saving

keras.py

@@ -899,7 +899,7 @@ class DenseLayer(tf.keras.layers.Layer):
     """
 
     def __init__(self, nodes=0, activation=None, dropout=0.0, l2=0, batch_norm=False):
-        super().__init__()
+        super().__init__(name="DenseLayer")
         self.nodes = nodes
         self.activation = activation
         self.dropout = dropout
@@ -914,7 +914,7 @@ class DenseLayer(tf.keras.layers.Layer):
         parts.append(weights)
 
         if self.batch_norm:
-            dropout = 0.0
+            self.dropout = 0.0
             bn = tf.keras.layers.BatchNormalization()
             parts.append(bn)
 
@@ -922,9 +922,9 @@ class DenseLayer(tf.keras.layers.Layer):
         parts.append(act)
 
         if self.activation == "selu":
-            dropout = tf.keras.layers.AlphaDropout(dropout)
+            dropout = tf.keras.layers.AlphaDropout(self.dropout)
         else:
-            dropout = tf.keras.layers.Dropout(dropout)
+            dropout = tf.keras.layers.Dropout(self.dropout)
 
         parts.append(dropout)
         self.parts = parts
@@ -995,25 +995,25 @@ class FullyConnected(tf.keras.layers.Layer):
 
     """
 
-    def __init__(self, layers=0, **kwargs):
+    def __init__(self, layers=0, sub_kwargs=None, **kwargs):
         super().__init__(name="FullyConnected")
         self.layers = layers
-        self.kwargs = kwargs
+        self.sub_kwargs = kwargs if sub_kwargs is None else sub_kwargs
 
     def build(self, input_shape):
-        _layers = []
+        network_layers = []
         for layer in range(self.layers):
-            _layers.append(DenseLayer(**self.kwargs))
-        self._layers = _layers
+            network_layers.append(DenseLayer(**self.sub_kwargs))
+        self.network_layers = network_layers
 
     def call(self, input_tensor, training=False):
         x = input_tensor
-        for layer in self._layers:
+        for layer in self.network_layers:
             x = layer(x, training=training)
         return x
 
     def get_config(self):
-        return {"layers": self.layers}
+        return {"layers": self.layers, "sub_kwargs": self.sub_kwargs}
 
 
 class ResNet(tf.keras.layers.Layer):