如何在功能 API 中实现双向包装？

Question

Does the bidirectional layer connect encoder to decoder or decoder to decoder. This is the 3 parts of the encoder which feed to the decoders below.

#encoding layers
input_context = Input(shape = (maxLen, ), dtype = 'int32', name = 'input_context')
input_ctx_embed = embed_layer(input_context)
encoder_lstm, h1, c1 = LSTM(256, return_state = True, return_sequences = True)(input_ctx_embed)
encoder_lstm2,h2, c2 = LSTM(256, return_state = True, return_sequences = True)(encoder_lstm)
_,h3, c3 = LSTM(256, return_state = True)(encoder_lstm2)
encoder_states = [h1, c1, h2, c2,h3,c3]

#layers for the decoder
input_target = Input(shape = (maxLen, ), dtype = 'int32', name = 'input_target')
input_tar_embed = embed_layer(input_target)
# the decoder lstm uses the final states from the encoder lstm as the initial state
decoder_lstm, context_h, context_c = LSTM(256, return_state = True, return_sequences = True) 
         (input_tar_embed, initial_state = [h1, c1],)
decoder_lstm2, context_h2, context_c2 = LSTM(256, return_state = True, return_sequences = True) 
         (decoder_lstm, initial_state = [h2, c2],)
final, context_h3, context_c3 = LSTM(256, return_state = True, return_sequences = True) 
        (decoder_lstm2, initial_state = [h3, c3],)
dense_layer=Dense(vocab_size, activation = 'softmax')
output = TimeDistributed(dense_layer)(final)
#output=Dropout(0.3)(output)
model = Model([input_context, input_target], output)

Answer 1

Not sure where the bidirectional layer is, since in my opinion, if you would like to use keras.layers.LSTM() to build a Bidirectional RNN structure without using keras.layer.Bidirectional() , then there's one setting in keras.layers.LSTM() which is called go_backwards and its default is False , set it True makes the LSTM going backward. And if you are just asking where to put Bidirectional LSTM in a encoder-decoder structure, then my answer will be "you can put it wherever you want , if that way make your model better ."

If I mixed up anything, let me know.