Merge pull request #28 from Yoctol/fix-sequential-add

remove the unexpected part in Sequential.add()

talos/module/sequential.py

@@ -2,18 +2,52 @@
 from tensorflow.python.keras.engine.network import Network
 from tensorflow.python.keras.engine.sequential import Sequential as keras_Sequential
 from tensorflow.python.util import tf_inspect
+from tensorflow.python.keras.engine import base_layer
+from tensorflow.python.training.checkpointable import base as checkpointable
 
 
 class Sequential(keras_Sequential):
 
-    # override: add **kwargs
+    # HACK override: remove adding InputLayer part!!
+    # https://github.com/tensorflow/tensorflow/blob/r1.11/tensorflow/python/keras/engine/sequential.py#L122-L188
+    # just copy/paste the source code and remove L142-L170
+    @checkpointable.no_automatic_dependency_tracking
+    def add(self, layer):
+        # source code L137
+        if not isinstance(layer, base_layer.Layer):
+            raise TypeError(
+                'The added layer must be '
+                'an instance of class Layer. '
+                f'Found: {layer}',
+            )
+        self.built = False
+
+        # source code L172
+        if self.outputs:
+            # If the model is being built continuously on top of an input layer:
+            # refresh its output.
+            output_tensor = layer(self.outputs[0])
+            if isinstance(output_tensor, list):
+                raise TypeError(
+                    'All layers in a Sequential model '
+                    'should have a single output tensor. '
+                    'For multi-output layers, '
+                    'use the functional API.',
+                )
+                self.outputs = [output_tensor]
+
+        self._layers.append(layer)
+        if self._layers:
+            self._track_layers(self._layers)
+
+    # HACK override: add **kwargs
     # https://github.com/tensorflow/tensorflow/blob/r1.11/tensorflow/python/keras/engine/sequential.py#L227-L233
     def call(self, inputs, training=None, mask=None, **kwargs):
         outputs, _ = self._call_and_compute_mask(
             inputs, training=training, mask=mask, **kwargs)
         return outputs
 
-    # override: add **kwargs
+    # HACK override: add **kwargs
     # https://github.com/tensorflow/tensorflow/blob/r1.11/tensorflow/python/keras/engine/sequential.py#L235-L257
     def _call_and_compute_mask(self, inputs, training=None, mask=None, **kwargs):
         if not self.built:

talos/module/tests/test_sequential.py

@@ -12,6 +12,34 @@ def graph():
         yield graph
 
 
+def test_build_sublayers_when_first_called(graph):
+    sequential = Sequential([
+        tf.keras.layers.Embedding(20, 10),
+        tf.keras.layers.LSTM(10, return_sequences=True),
+        tf.keras.layers.Dense(5),
+        tf.keras.layers.MaxPooling1D(),
+    ])
+    assert all(not layer.built for layer in sequential.layers)
+    inputs = tf.zeros([1, 3], dtype=tf.float32)
+    sequential(inputs)
+    assert all(layer.built for layer in sequential.layers)
+
+
+def test_context_manager_work_when_first_called(graph):
+    sequential = Sequential([
+        tf.keras.layers.Embedding(20, 10),
+        tf.keras.layers.LSTM(10, return_sequences=True),
+        tf.keras.layers.Dense(5),
+        tf.keras.layers.MaxPooling1D(),
+    ])
+    with tf.variable_scope('scope'):
+        inputs = tf.zeros([1, 3], dtype=tf.float32)
+        sequential(inputs)
+    variables = graph.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
+    assert all(var.graph is graph for var in variables)
+    assert all(var.name.startswith('scope') for var in variables)
+
+
 def test_additional_inputs(graph):
 
     class LayerNeedSeqlen(tf.keras.layers.Layer):