如何从 LSTM 层获取输入的嵌入作为输出

Question

我创建了两个相同的编码器来训练两个不同的文本 job_description 和 resume，以余弦相似度作为损失函数。

encoder_input1 = Input(shape= (None,),name='encoder1')
embed_dim = 200
embedded_seq_encoder1 = Embedding(input_dim = 9000, output_dim = embed_dim)(encoder_input1)
encoder1 = LSTM(256, return_state=True)
encoder_output1, state_h,state_c = encoder1(embedded_seq_encoder1)




encoder_input2 = Input(shape= (None,),name='encoder2')
embed_dim = 200
embedded_seq_encoder2 = Embedding(input_dim = 9000, output_dim = embed_dim)(encoder_input2)
encoder2 = LSTM(256, return_state=True)
encoder_output2, state_h,state_c = encoder2(embedded_seq_encoder2)

encoder_dense = tf.keras.layers.Dense(1, activation='softmax', name='Final-Output-Dense')
encoder_outputs = encoder_dense(merged_Encoder_layer)

model = Model([encoder_input1,encoder_input2],[encoder_outputs])
model.compile(optimizer='rmsprop', loss = tf.keras.losses.BinaryCrossentropy())
model.summary()

然后我将两个编码器都转换为自定义层，我计划用这些编码器和一些密集层构建一个顺序模型

class Custom_Encoder_Layer1(tf.keras.layers.Layer):

    def __init__(self,shape,embed_dim,input_dim,n_units):
        super(Custom_Encoder_Layer1, self).__init__()
        self.shape = shape
        self.embed_dim = embed_dim
        self.input_dim = input_dim
        self.n_units = n_units
    def Encoder1(self):

        encoder_input1 = Input(shape= shape,name='encoder_layer1')
        embed_dim = embed_dim
        embedded_seq_encoder1 = Embedding(input_dim = input_dim, output_dim = embed_dim)(encoder_input1)
        encoder1 = LSTM(n_units, return_state=True)
        encoder_output1, state_h,state_c = encoder1(embedded_seq_encoder1)
        return encoder_input1,encoder_output1



class Custom_Encoder_Layer2(tf.keras.Model):

    def __init__(self,shape,embed_dim,input_dim,n_units):
        super(Custom_Encoder_Layer2, self).__init__()
        self.shape = shape
        self.embed_dim = embed_dim
        self.input_dim = input_dim
        self.n_units = n_units

    def Encoder2(self):

        encoder_input2 = Input(shape= shape,name='encoder_layer2')
        embed_dim = embed_dim
        embedded_seq_encoder2 = Embedding(input_dim = input_dim, output_dim = embed_dim)(encoder_input2)
        encoder2 = LSTM(n_units, return_state=True)
        encoder_output2, state_h,state_c = encoder2(embedded_seq_encoder2)
        return encoder_input2,encoder_output2

model = tf.keras.models.Sequential([

    Custom_Encoder_Layer1((None,),200,9000,256),
    Model.layers[4],
    Custom_Encoder_Layer2((None,),200,9000,256),
    Model.layers[5],
    Model.layers[7]  

])

现在我的问题是如何训练这个序列模型并获得输入文本的嵌入