changed weighted add

autokeras/graph.py

@@ -10,7 +10,7 @@
 from autokeras import constant
 from autokeras.layer_transformer import wider_bn, wider_next_conv, wider_next_dense, wider_pre_dense, wider_pre_conv, \
     deeper_conv_block, dense_to_deeper_block, add_noise
-from autokeras.layers import StubConcatenate, StubWeightedAdd, StubConv, is_layer, layer_width, \
+from autokeras.layers import StubConcatenate, StubAdd, StubConv, is_layer, layer_width, \
     to_real_layer
 from autokeras.stub import to_stub_model
 
@@ -410,7 +410,7 @@ def to_add_skip_model(self, start_id, end_id):
 
         # Add the weighted add layer.
         new_node_id = self._add_new_node()
-        layer = StubWeightedAdd()
+        layer = StubAdd()
         if self.weighted:
             layer.set_weights([np.float32(1.0)])
 
@@ -526,7 +526,7 @@ def produce_model(self):
             for u, layer_id in self.reverse_adj_list[v]:
                 layer = self.layer_list[layer_id]
 
-                if isinstance(layer, (StubWeightedAdd, StubConcatenate)):
+                if isinstance(layer, (StubAdd, StubConcatenate)):
                     edge_input_tensor = list(map(lambda x: node_list[x],
                                                  self.layer_id_to_input_node_ids[layer_id]))
                 else:

autokeras/layer_transformer.py

@@ -2,7 +2,7 @@
 
 from autokeras import constant
 from autokeras.layers import StubConv, StubBatchNormalization, StubActivation, StubDropout, StubDense, \
-    StubWeightedAdd
+    StubAdd
 
 NOISE_RATIO = 1e-4
 
@@ -237,10 +237,10 @@ def wider_weighted_add(layer, n_add, weighted=True):
         The wider weighted add layer
     """
     if not weighted:
-        return StubWeightedAdd()
+        return StubAdd()
 
     n_add += 0
-    new_layer = StubWeightedAdd()
+    new_layer = StubAdd()
     new_layer.set_weights(layer.get_weights())
     return new_layer
 

autokeras/layers.py

@@ -1,43 +1,10 @@
-from keras import backend
 from keras.engine import InputLayer
 from keras.layers import Add, Conv2D, Conv3D, Conv1D, Dense, BatchNormalization, Concatenate, Dropout, Activation, \
     Flatten, MaxPooling1D, MaxPooling2D, MaxPooling3D, GlobalAveragePooling1D, GlobalAveragePooling2D, \
     GlobalAveragePooling3D
 from keras.regularizers import l2
 
 
-class WeightedAdd(Add):
-    """Weighted Add class inherited from Add class
-
-    It's used to do add weights for data in Add layer
-
-    Attributes:
-        weights: backend variable
-        one: const 1.0
-        kernel: None
-        _trainable_weights: list that store weight
-    """
-
-    def __init__(self, **kwargs):
-        """Init Weighted add class"""
-        super(WeightedAdd, self).__init__(**kwargs)
-        self.weight = backend.variable(1.0)
-        self.one = backend.constant(1.0)
-        self.kernel = None
-        self._trainable_weights.append(self.weight)
-
-    def call(self, x, **kwargs):
-        """Override call function in Add and return new weights"""
-        a = backend.tf.scalar_mul(self.weight, x[0])
-        b = backend.tf.scalar_mul(backend.tf.subtract(self.one, self.weight), x[1])
-        c = backend.tf.add(a, b)
-        return c
-
-    def compute_output_shape(self, input_shape):
-        """Return output_shape"""
-        return input_shape
-
-
 class StubLayer:
     def __init__(self, input_node=None, output_node=None):
         self.input = input_node
@@ -94,7 +61,7 @@ class StubConcatenate(StubAggregateLayer):
     pass
 
 
-class StubWeightedAdd(StubAggregateLayer):
+class StubAdd(StubAggregateLayer):
     pass
 
 
@@ -142,8 +109,8 @@ def is_layer(layer, layer_type):
         return isinstance(layer, (StubBatchNormalization, BatchNormalization))
     if layer_type == 'Concatenate':
         return isinstance(layer, (StubConcatenate, Concatenate))
-    if layer_type == 'WeightedAdd':
-        return isinstance(layer, (StubWeightedAdd, WeightedAdd))
+    if layer_type == 'Add':
+        return isinstance(layer, (StubAdd, Add))
     if layer_type == 'Pooling':
         return isinstance(layer, StubPooling) or is_pooling_layer(layer)
     if layer_type == 'Dropout':
@@ -217,7 +184,7 @@ def to_real_layer(layer):
     if is_layer(layer, 'Concatenate'):
         return Concatenate()
     if is_layer(layer, 'WeightedAdd'):
-        return WeightedAdd()
+        return Add()
     if is_layer(layer, 'Dropout'):
         return Dropout(layer.rate)
     if is_layer(layer, 'Activation'):
@@ -234,7 +201,7 @@ def to_stub_layer(layer, input_id, output_id):
     elif is_layer(layer, 'Dense'):
         temp_stub_layer = StubDense(layer.units, layer.activation, input_id, output_id)
     elif is_layer(layer, 'WeightedAdd'):
-        temp_stub_layer = StubWeightedAdd(input_id, output_id)
+        temp_stub_layer = StubAdd(input_id, output_id)
     elif is_layer(layer, 'Concatenate'):
         temp_stub_layer = StubConcatenate(input_id, output_id)
     elif is_layer(layer, 'BatchNormalization'):

tests/test_layer_transformer.py

@@ -35,10 +35,10 @@ def test_wider_bn():
 
 
 def test_wider_weighted_add():
-    layer = StubWeightedAdd()
+    layer = StubAdd()
     layer.set_weights(get_add_skip_model().layers[13].get_weights())
     new_layer = wider_weighted_add(layer, 4)
-    assert isinstance(new_layer, StubWeightedAdd)
+    assert isinstance(new_layer, StubAdd)
 
 
 def test_wider_next_dense():