remove input shape of first conv layer in networks because unnecessary

add selu activation to deeper network designs

models/pauls_networks.py

@@ -52,8 +52,7 @@ def get_model(cnnDropout, flow_features, domain_features, window_size, domain_le
     # CNN processing a small slides of flow windows
     y = Conv1D(cnn_dims,
                kernel_size,
-               activation='relu',
+               activation='relu'
-               input_shape=(window_size, domain_features + flow_features))(merged)
     # remove temporal dimension by global max pooling
     y = GlobalMaxPooling1D()(y)
     y = Dropout(cnnDropout)(y)
@@ -78,8 +77,7 @@ def get_new_model(dropout, flow_features, domain_features, window_size, domain_l
     # CNN processing a small slides of flow windows
     y = Conv1D(cnn_dims,
                kernel_size,
-               activation='relu',
+               activation='relu')(merged)
-               input_shape=(window_size, domain_features + flow_features))(merged)
     # remove temporal dimension by global max pooling
     y = GlobalMaxPooling1D()(y)
     y = Dropout(dropout)(y)

models/renes_networks.py

@@ -1,11 +1,26 @@
+from collections import namedtuple
+
 import keras
+from keras.activations import elu
 from keras.engine import Input, Model as KerasModel
-from keras.layers import Embedding, Conv1D, GlobalMaxPooling1D, Dense, Dropout, TimeDistributed, MaxPool1D, \
+from keras.layers import Conv1D, Dense, Dropout, Embedding, GlobalAveragePooling1D, GlobalMaxPooling1D, MaxPool1D, \
-    GlobalAveragePooling1D
+    TimeDistributed
 
 import dataset
 
-from collections import namedtuple
+
+def selu(x):
+    """Scaled Exponential Linear Unit. (Klambauer et al., 2017)
+    # Arguments
+        x: A tensor or variable to compute the activation function for.
+    # References
+        - [Self-Normalizing Neural Networks](https://arxiv.org/abs/1706.02515)
+    # copied from keras.io
+    """
+    alpha = 1.6732632423543772848170429916717
+    scale = 1.0507009873554804934193349852946
+    return scale * elu(x, alpha)
+
 
 Model = namedtuple("Model", ["in_domains", "in_flows", "out_client", "out_server"])
 
@@ -13,9 +28,9 @@ Model = namedtuple("Model", ["in_domains", "in_flows", "out_client", "out_server
 def get_embedding(embedding_size, input_length, filter_size, kernel_size, hidden_dims, drop_out=0.5):
     x = y = Input(shape=(input_length,))
     y = Embedding(input_dim=dataset.get_vocab_size(), output_dim=embedding_size)(y)
-    y = Conv1D(filter_size, kernel_size=5, activation='relu')(y)
+    y = Conv1D(filter_size, kernel_size=5, activation=selu)(y)
-    y = Conv1D(filter_size, kernel_size=3, activation='relu')(y)
+    y = Conv1D(filter_size, kernel_size=3, activation=selu)(y)
-    y = Conv1D(filter_size, kernel_size=3, activation='relu')(y)
+    y = Conv1D(filter_size, kernel_size=3, activation=selu)(y)
     y = GlobalAveragePooling1D()(y)
     y = Dense(hidden_dims, activation="relu")(y)
     return KerasModel(x, y)
@@ -28,17 +43,17 @@ def get_model(cnnDropout, flow_features, domain_features, window_size, domain_le
     ipt_flows = Input(shape=(window_size, flow_features), name="ipt_flows")
     merged = keras.layers.concatenate([encoded, ipt_flows], -1)
     # CNN processing a small slides of flow windows
-    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation='relu', padding="same",
+    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation=selu, padding="same",
                input_shape=(window_size, domain_features + flow_features))(merged)
     y = MaxPool1D(pool_size=3, strides=1)(y)
-    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation='relu', padding="same")(y)
+    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation=selu, padding="same")(y)
     y = MaxPool1D(pool_size=3, strides=1)(y)
-    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation='relu', padding="same")(y)
+    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation=selu, padding="same")(y)
     # remove temporal dimension by global max pooling
     y = GlobalMaxPooling1D()(y)
     y = Dropout(cnnDropout)(y)
-    y = Dense(dense_dim, activation='relu')(y)
+    y = Dense(dense_dim, activation=selu)(y)
-    y = Dense(dense_dim // 2, activation='relu')(y)
+    y = Dense(dense_dim // 2, activation=selu)(y)
     out_client = Dense(1, activation='sigmoid', name="client")(y)
     out_server = Dense(1, activation='sigmoid', name="server")(y)
 
@@ -49,22 +64,35 @@ def get_new_model(dropout, flow_features, domain_features, window_size, domain_l
                   dense_dim, cnn, model_output="both"):
     ipt_domains = Input(shape=(window_size, domain_length), name="ipt_domains")
     ipt_flows = Input(shape=(window_size, flow_features), name="ipt_flows")
-    encoded = TimeDistributed(cnn)(ipt_domains)
+    encoded = TimeDistributed(cnn, name="domain_cnn")(ipt_domains)
     merged = keras.layers.concatenate([encoded, ipt_flows], -1)
     y = Dense(dense_dim, activation="relu")(merged)
     out_server = Dense(1, activation="sigmoid", name="server")(y)
+    merged = keras.layers.concatenate([merged, y], -1)
     # CNN processing a small slides of flow windows
-    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation='relu', padding="same",
+    y = Conv1D(filters=cnn_dims,
-               input_shape=(window_size, domain_features + flow_features))(y)
+               kernel_size=kernel_size,
-    y = MaxPool1D(pool_size=3, strides=1)(y)
+               activation=selu,
-    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation='relu', padding="same")(y)
+               padding="same",
-    y = MaxPool1D(pool_size=3, strides=1)(y)
+               input_shape=(window_size, domain_features + flow_features))(merged)
-    y = Conv1D(filters=cnn_dims, kernel_size=kernel_size, activation='relu', padding="same")(y)
+    y = MaxPool1D(pool_size=3,
+                  strides=1)(y)
+    y = Conv1D(filters=cnn_dims,
+               kernel_size=kernel_size,
+               activation=selu,
+               padding="same")(y)
+    y = MaxPool1D(pool_size=3,
+                  strides=1)(y)
+    y = Conv1D(filters=cnn_dims,
+               kernel_size=kernel_size,
+               activation=selu,
+               padding="same")(y)
     # remove temporal dimension by global max pooling
     y = GlobalMaxPooling1D()(y)
     y = Dropout(dropout)(y)
-    y = Dense(dense_dim, activation='relu')(y)
+    y = Dense(dense_dim,
-
+              activation=selu,
+              name="dense_client")(y)
     out_client = Dense(1, activation='sigmoid', name="client")(y)
 
     return Model(ipt_domains, ipt_flows, out_client, out_server)