python-将Keras中嵌入层输出的形状从3D减少到2D

我有一个数据集，其中每个示例有10个句子（config.max\u路径），每个句子有10个单词（config.max\u路径长度）。我正在做一个基于单词的嵌入。下面是代码：

def create_model(vocab_size):
    config = Config()
    path_input = Input((config.max_paths,config.max_path_length), dtype=tf.int32)
    
    #embedding layer
    nodes_embedded = Embedding(vocab_size+2, config.embedding_size, trainable = True, name='node_embedding')(path_input)
    
    # path embeddings from node embeddings
    # convert nodes_embedded to path embeddings to get paths_embedded
    
    subtree_vectors = TimeDistributed(
            Dense(config.embedding_size, use_bias=False, activation='tanh'))(nodes_embedded)
    attention_vectors = Dense(1,)(subtree_vectors)
    attention_weights = Softmax()(attention_vectors)
    code_vectors = K.sum(subtree_vectors * attention_weights, axis=1)
    output_class = Dense(config.num_classes, use_bias=False, activation='softmax')(code_vectors)
    
    model = Model(inputs=path_input, outputs=output_class)
    return model

当前的模型摘要：

 Layer (type)                   Output Shape         Param #     Connected to                     
==================================================================================================
 input_6 (InputLayer)           [(None, 15, 10)]     0           []                               
                                                                                                  
 node_embedding (Embedding)     (None, 15, 10, 10)   690         ['input_6[0][0]']                
                                                                                                  
 time_distributed_5 (TimeDistri  (None, 15, 10, 10)  100         ['node_embedding[0][0]']         
 buted)                                                                                           
                                                                                                  
 dense_14 (Dense)               (None, 15, 10, 1)    11          ['time_distributed_5[0][0]']     
                                                                                                  
 softmax_4 (Softmax)            (None, 15, 10, 1)    0           ['dense_14[0][0]']               
                                                                                                  
 tf.math.multiply_4 (TFOpLambda  (None, 15, 10, 10)  0           ['time_distributed_5[0][0]',     
 )                                                                'softmax_4[0][0]']              
                                                                                                  
 tf.math.reduce_sum_4 (TFOpLamb  (None, 10, 10)      0           ['tf.math.multiply_4[0][0]']     
 da)                                                                                              
                                                                                                  
 dense_15 (Dense)               (None, 10, 2)        20          ['tf.math.reduce_sum_4[0][0]'] 

我想对单词嵌入进行一些汇总或求和，以获得句子嵌入的简化形状，其中模型可能如下所示：

 Layer (type)                   Output Shape                              
==================================================================================================
 input_6 (InputLayer)           [(None, 15, 10)]                                  
                                                                                                  
 node_embedding (Embedding)     (None, 15, 10, 10)  

 path_embedding               (None, 15, 10) (Want to add this layer somehow)                
                                                                                                  
 time_distributed_5 (TimeDistri  (None, 15, 10)         
 buted)                                                                                           
                                                                                                  
 dense_14 (Dense)               (None, 15, 1)      
                                                                                                  
 softmax_4 (Softmax)            (None, 15, 1)            
                                                                                                  
 tf.math.multiply_4 (TFOpLambda  (None, 15, 10)            
 )                                                                           
                                                                                                  
 tf.math.reduce_sum_4 (TFOpLamb  (None, 10)         
 da)                                                                                              
                                                                                                  
 dense_15 (Dense)               (None, 2)   

我该怎么做？