無法合並多個輸入 tf.keras 模型/錯誤：圖形斷開連接：無法獲得張量 Tensor 的值

Question

我不明白為什么在編譯 tf.keras 模型時連接層不能與year_input一起使用。

細節：

• 要連接的 3 個層是tf.float32類型。
• 將 Keras Functional API 與 Tensorflow 2.1.0 GPU 結合使用。
• 這 3 個層的類型為： <tf.Tensor 'year_input_ll_12:0' shape=(None, 5) dtype=float32> 。
• 如果我創建一個沒有year_input的模型，模型將正確編譯。

def create_system_classifier_model(df, pretrain_model, create_date_df, 
                                   output_cat_nbr, spec_max_length, heading_max_length):

    heading_input = tf.keras.layers.Input((heading_max_length,), name="heading_input", dtype=tf.int32)
    spec_input = tf.keras.layers.Input((spec_max_length,), name="spec_input", dtype=tf.int32)
    year_input_ts = tf.keras.layers.Input((5,), name="year_input_ll", dtype=tf.float32)

    spec_embedding = pretrain_model(heading_input)[0]
    heading_embedding = pretrain_model(spec_input)[0]

    heading_pool_ts = tf.keras.layers.GlobalAveragePooling1D()(spec_embedding)
    spec_pool_ts = tf.keras.layers.GlobalAveragePooling1D()(heading_embedding)

    concat = tf.keras.layers.concatenate([heading_pool_ts, spec_pool_ts, year_input_ts], axis=1)

    dense_ts_1 = tf.keras.layers.Dense(768, activation='relu', name='dense_ts_1')(concat)
    dense_ts_2 = tf.keras.layers.Dense(768, activation='relu', name='dense_ts_2')(dense_ts_1)
    dense_ts_3 = tf.keras.layers.Dense(768, activation='relu', name='dense_ts_3')(dense_ts_2)
    drop_ts = tf.keras.layers.Dropout(0.2)(dense_ts_3)
    output_ts = tf.keras.layers.Dense(output_cat_nbr, activation='sigmoid')(drop_ts)

    model = tf.keras.models.Model(inputs=[heading_input, spec_input, year_input], outputs=output_ts)

    return model

函數調用

system_classifier_model = create_system_classifier_model(df, 
                                                         distil_bert_model, 
                                                         create_date_df,
                                                         len(system_cat_ls),
                                                         spec_max_length, 
                                                         heading_max_length)

錯誤

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-171-367b40edc46e> in <module>()
      4                                                          len(system_cat_ls),
      5                                                          spec_max_length,
----> 6                                                          heading_max_length)

5 frames
<ipython-input-170-6b485af7c7be> in create_system_classifier_model(df, pretrain_model, create_date_df, output_cat_nbr, spec_max_length, heading_max_length)
     20     output_ts = tf.keras.layers.Dense(output_cat_nbr, activation='sigmoid')(drop_ts)
     21 
---> 22     model = tf.keras.models.Model(inputs=[heading_input, spec_input], outputs=output_ts)
     23     model = tf.keras.models.Model(inputs=[heading_input, spec_input, year_input], outputs=output_ts)
     24 

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py in __init__(self, *args, **kwargs)
    144 
    145   def __init__(self, *args, **kwargs):
--> 146     super(Model, self).__init__(*args, **kwargs)
    147     _keras_api_gauge.get_cell('model').set(True)
    148     # initializing _distribution_strategy here since it is possible to call

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/network.py in __init__(self, *args, **kwargs)
    167         'inputs' in kwargs and 'outputs' in kwargs):
    168       # Graph network
--> 169       self._init_graph_network(*args, **kwargs)
    170     else:
    171       # Subclassed network

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/training/tracking/base.py in _method_wrapper(self, *args, **kwargs)
    455     self._self_setattr_tracking = False  # pylint: disable=protected-access
    456     try:
--> 457       result = method(self, *args, **kwargs)
    458     finally:
    459       self._self_setattr_tracking = previous_value  # pylint: disable=protected-access

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/network.py in _init_graph_network(self, inputs, outputs, name, **kwargs)
    322     # Keep track of the network's nodes and layers.
    323     nodes, nodes_by_depth, layers, _ = _map_graph_network(
--> 324         self.inputs, self.outputs)
    325     self._network_nodes = nodes
    326     self._nodes_by_depth = nodes_by_depth

/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/network.py in _map_graph_network(inputs, outputs)
   1674                              'The following previous layers '
   1675                              'were accessed without issue: ' +
-> 1676                              str(layers_with_complete_input))
   1677         for x in nest.flatten(node.output_tensors):
   1678           computable_tensors.add(id(x))

ValueError: Graph disconnected: cannot obtain value for tensor Tensor("year_input_ll_11:0", shape=(None, 5), dtype=float32) at layer "year_input_ll". The following previous layers were accessed without issue: ['spec_input', 'heading_input', 'tf_distil_bert_model_2', 'tf_distil_bert_model_2', 'global_average_pooling1d_72', 'global_average_pooling1d_73']

Answer 1

此代碼有效：

def create_system_classifier_model(df, pretrain_model, create_date_df, 
                                   output_cat_nbr, spec_max_length, heading_max_length):

    heading_input = tf.keras.layers.Input((heading_max_length,), name="heading_input", dtype=tf.int32)
    spec_input = tf.keras.layers.Input((spec_max_length,), name="spec_input", dtype=tf.int32)
    year_input_ts = tf.keras.layers.Input((5,), name="year_input_ll", dtype=tf.float32)

    spec_embedding = pretrain_model(heading_input)[0]
    heading_embedding = pretrain_model(spec_input)[0]

    heading_pool_ts = tf.keras.layers.GlobalAveragePooling1D()(spec_embedding)
    spec_pool_ts = tf.keras.layers.GlobalAveragePooling1D()(heading_embedding)

    concat = tf.keras.layers.concatenate([heading_pool_ts, spec_pool_ts, year_input_ts], axis=1)

    dense_ts_1 = tf.keras.layers.Dense(768, activation='relu', name='dense_ts_1')(concat)
    dense_ts_2 = tf.keras.layers.Dense(768, activation='relu', name='dense_ts_2')(dense_ts_1)
    dense_ts_3 = tf.keras.layers.Dense(768, activation='relu', name='dense_ts_3')(dense_ts_2)
    drop_ts = tf.keras.layers.Dropout(0.2)(dense_ts_3)
    output_ts = tf.keras.layers.Dense(output_cat_nbr, activation='sigmoid')(drop_ts)

    model = tf.keras.models.Model(inputs=[heading_input, spec_input, year_input], outputs=output_ts)

    return model