[keras] adds improvements for DenseLayer

keras.py

@@ -975,13 +975,20 @@ class DenseLayer(tf.keras.layers.Layer):
 
     def build(self, input_shape):
         parts = []
+        if self.activation == "selu":
+            kernel_initializer = tf.variance_scaling_initializer(factor=1.0, mode="FAN_IN")
+        else:
+            kernel_initializer = "glorot_uniform"
 
         l2 = tf.keras.regularizers.l2(self.l2)
-        weights = tf.keras.layers.Dense(self.nodes, kernel_regularizer=l2)
+        weights = tf.keras.layers.Dense(
+            self.nodes,
+            kernel_regularizer=l2,
+            kernel_initializer=kernel_initializer,
+        )
         parts.append(weights)
 
-        if self.batch_norm:
-            self.dropout = 0.0
+        if self.batch_norm and not self.activation == "selu":
             bn = tf.keras.layers.BatchNormalization()
             parts.append(bn)