Merge branch 'lukas_issue361_feat_custom-dense-layers-in-rnn' into...

Merge branch 'lukas_issue361_feat_custom-dense-layers-in-rnn' into 'lukas_issue362_feat_branched-rnn'

Lukas issue361 feat custom dense layers in rnn

See merge request !389

mlair/model_modules/recurrent_networks.py

@@ -33,7 +33,7 @@ class RNN(AbstractModelClass):  # pragma: no cover
     def __init__(self, input_shape: list, output_shape: list, activation="relu", activation_output="linear",
                  activation_rnn="tanh", dropout_rnn=0,
                  optimizer="adam", n_layer=1, n_hidden=10, regularizer=None, dropout=None, layer_configuration=None,
-                 batch_normalization=False, rnn_type="lstm", add_dense_layer=False, **kwargs):
+                 batch_normalization=False, rnn_type="lstm", add_dense_layer=False, dense_layer_configuration=None, **kwargs):
         """
         Sets model and loss depending on the given arguments.
 
@@ -64,6 +64,15 @@ class RNN(AbstractModelClass):  # pragma: no cover
             is added if set to false. (Default false)
         :param rnn_type: define which kind of recurrent network should be applied. Chose from either lstm or gru. All
             units will be of this kind. (Default lstm)
+        :param add_dense_layer: set True to use additional dense layers between last recurrent layer and output layer. 
+            If no further specification is made on the exact dense_layer_configuration, a single layer as added with n 
+            neurons where n is equal to min(n_previous_layer, n_output**2). If set to False, the output layer directly 
+            follows after the last recurrent layer.
+        :param dense_layer_configuration: specify the number of dense layers and the number of neurons given as list
+            where each element corresponds to the number of neurons to add. The position / length of the list specifies
+            the number of layers to add. The last layer is followed by the output layer. In case a value is given for
+            the number of neurons that is less than the number of output neurons, the addition of dense layers is 
+            stopped immediately.
         """
 
         assert len(input_shape) == 1
@@ -80,6 +89,7 @@ class RNN(AbstractModelClass):  # pragma: no cover
         self.optimizer = self._set_optimizer(optimizer.lower(), **kwargs)
         self.bn = batch_normalization
         self.add_dense_layer = add_dense_layer
+        self.dense_layer_configuration = dense_layer_configuration or []
         self.layer_configuration = (n_layer, n_hidden) if layer_configuration is None else layer_configuration
         self.RNN = self._rnn.get(rnn_type.lower())
         self._update_model_name(rnn_type)
@@ -119,9 +129,22 @@ class RNN(AbstractModelClass):  # pragma: no cover
                 x_in = self.dropout(self.dropout_rate)(x_in)
 
         if self.add_dense_layer is True:
-            x_in = keras.layers.Dense(min(self._output_shape ** 2, conf[-1]), name=f"Dense_{len(conf) + 1}",
-                                      kernel_initializer=self.kernel_initializer, )(x_in)
-            x_in = self.activation(name=f"{self.activation_name}_{len(conf) + 1}")(x_in)
+            if len(self.dense_layer_configuration) == 0:
+                x_in = keras.layers.Dense(min(self._output_shape ** 2, conf[-1]), name=f"Dense_{len(conf) + 1}",
+                                          kernel_initializer=self.kernel_initializer, )(x_in)
+                x_in = self.activation(name=f"{self.activation_name}_{len(conf) + 1}")(x_in)
+                if self.dropout is not None:
+                    x_in = self.dropout(self.dropout_rate)(x_in)
+            else:
+                for layer, n_hidden in enumerate(self.dense_layer_configuration):
+                    if n_hidden < self._output_shape:
+                        break
+                    x_in = keras.layers.Dense(n_hidden, name=f"Dense_{len(conf) + layer + 1}",
+                                              kernel_initializer=self.kernel_initializer, )(x_in)
+                    x_in = self.activation(name=f"{self.activation_name}_{len(conf) + layer + 1}")(x_in)
+                    if self.dropout is not None:
+                        x_in = self.dropout(self.dropout_rate)(x_in)
+
         x_in = keras.layers.Dense(self._output_shape)(x_in)
         out = self.activation_output(name=f"{self.activation_output_name}_output")(x_in)
         self.model = keras.Model(inputs=x_input, outputs=[out])