python - Tensorflow 中的 Seq2Seq，但我收到 ValueError : Input 0 of layer gru_cell_3 is incompatible with the layer:

Question

我正在努力了解如何在 Tensorflow 中使用 RNN 来实现 Seq2Seq 模型，并且我到达了执行动态 RNN 的最后一步，我得到了动态动态解码步骤，并收到了错误:

“ValueError:层 gru_cell_3 的输入 0 与层不兼容:预期 ndim=2，发现 ndim=1。收到完整形状:[无]”

import tensorflow as tf

data_inputs = tf.placeholder(tf.float32,[None,102,300]) 

batch_lengths = tf.cast(tf.reduce_sum(tf.reduce_max(tf.sign(data_inputs),2),1),tf.int32)

encoder_cell_forward = tf.nn.rnn_cell.GRUCell(num_units = 150)

encoder_cell_backward = tf.nn.rnn_cell.GRUCell(num_units = 150)

_ , state = tf.nn.bidirectional_dynamic_rnn(
    encoder_cell_forward,encoder_cell_backward,
    data_inputs,sequence_length = batch_lengths,
    dtype = tf.float32 )

state = tf.concat(state,1)

decoder_cell = tf.nn.rnn_cell.GRUCell(num_units = 300)

helper = tf.contrib.seq2seq.TrainingHelper(state,batch_lengths)

projection_layer = tf.layers.Dense(
    units = 300,activation= None,trainable =True )

decoder = tf.contrib.seq2seq.BasicDecoder(
    decoder_cell, helper, state,
    output_layer=projection_layer)

final_outputs, final_state, final_sequence_lengths = tf.contrib.seq2seq.dynamic_decode(
    decoder,maximum_iterations= 102,impute_finished=False)

我在这里做错了什么？

Answer 1

我想通了，与训练助手一致:

tf.contrib.seq2seq.TrainingHelper(state,batch_lengths)

state 必须是您想要解码的序列批处理，如果您使用编码状态，则会引发错误。