🐛 fix bugs caused by discrepancies between tf.keras & Keras

BrikerMan · Jun 28, 2019 · f15fd0a · f15fd0a
1 parent b76d798
commit f15fd0a
Show file tree

Hide file tree

Showing 3 changed files with 14 additions and 7 deletions.
diff --git a/kashgari/layers/att_wgt_avg_layer.py b/kashgari/layers/att_wgt_avg_layer.py
@@ -10,10 +10,10 @@
 import tensorflow as tf
 from tensorflow.python import keras
 from tensorflow.python.keras import backend as K
-from tensorflow.python.keras.engine.input_spec import InputSpec
 
 L = keras.layers
 initializers = keras.initializers
+InputSpec = L.InputSpec
 
 if tf.test.is_gpu_available(cuda_only=True):
     L.LSTM = L.CuDNNLSTM
@@ -35,10 +35,12 @@ def build(self, input_shape):
         self.input_spec = [InputSpec(ndim=3)]
         assert len(input_shape) == 3
 
-        self.W = self.add_weight(shape=(input_shape[2], 1),
+        self.W = self.add_weight(shape=(input_shape[2].value, 1),
                                  name='{}_w'.format(self.name),
-                                 initializer=self.init)
-        self.trainable_weights = [self.W]
+                                 initializer=self.init,
+                                 trainable=True
+                                 )
+        # self.trainable_weights = [self.W]
         super(AttentionWeightedAverageLayer, self).build(input_shape)
 
     def call(self, x, mask=None):
@@ -77,6 +79,11 @@ def compute_mask(self, inputs, input_mask=None):
         else:
             return None
 
+    def get_config(self):
+        config = {'return_attention': self.return_attention,}
+        base_config = super(AttentionWeightedAverageLayer, self).get_config()
+        return dict(list(base_config.items()) + list(config.items()))
+
 
 AttentionWeightedAverage = AttentionWeightedAverageLayer
 AttWgtAvgLayer = AttentionWeightedAverageLayer

diff --git a/kashgari/layers/kmax_pool_layer.py b/kashgari/layers/kmax_pool_layer.py
@@ -10,9 +10,9 @@
 import tensorflow as tf
 from tensorflow.python import keras
 from tensorflow.python.keras import backend as K
-from tensorflow.python.keras.engine.input_spec import InputSpec
 
 L = keras.layers
+InputSpec = L.InputSpec
 
 if tf.test.is_gpu_available(cuda_only=True):
     L.LSTM = L.CuDNNLSTM

diff --git a/kashgari/tasks/classification/models.py b/kashgari/tasks/classification/models.py
@@ -283,13 +283,13 @@ def build_model_arc(self):
         tensors_conv = [layer_conv(embed_tensor) for layer_conv in layers_conv]
         tensors_matrix_sensor = []
         for tensor_conv in tensors_conv:
+            tensor_sensors = []
             tensor_sensors = [layer_sensor(tensor_conv) for layer_sensor in layers_sensor]
-            # tensor_sensors = []
             # tensor_sensors.append(L.GlobalMaxPooling1D()(tensor_conv))
             # tensor_sensors.append(AttentionWeightedAverageLayer()(tensor_conv))
             # tensor_sensors.append(L.GlobalAveragePooling1D()(tensor_conv))
             tensors_matrix_sensor.append(tensor_sensors)
-        tensors_views = [layer_view(tensors) for tensors in zip(*tensors_matrix_sensor)]
+        tensors_views = [layer_view(list(tensors)) for tensors in zip(*tensors_matrix_sensor)]
         tensor = layer_allviews(tensors_views)
         # tensors_v_cols = [L.concatenate(tensors, **config['v_col3']) for tensors
         #                   in zip(*tensors_matrix_sensor)]