我试图基于谷歌的NMT模型（https://github.com/tensorflow/nmt）建立我自己的神经机器翻译器。 一切顺利，简单的编码器/解码器，添加了BeamSearch，然后是Attention。现在我试图在编码器上添加双向rnn并遇到这个错误：
ValueError：没有足够的值来解包（预期2，得到0）
传递seq2seq.TrainingHelper和seq2seq.BasicDecoder后，在seq2seq.dynamic_decode函数上生成此错误。有人可以帮我解决这个问题吗？
我基于代码的这一特定部分：
https://github.com/tensorflow/nmt/blob/master/nmt/model.py#L589
我的代码：
if encoder_type == "bidirectional":
num_bi_layers = int(n_layers / 2)
num_residual_layers = n_layers - 1
num_bi_residual_layers = int(num_residual_layers / 2)
cell_list = []
for i in range(n_layers):
cell = tf.contrib.rnn.BasicLSTMCell(hidden_units, forget_bias=forget_bias)
if (i >= n_layers - num_residual_layers):
cell = tf.contrib.rnn.ResidualWrapper(cell, residual_fn=None)
if dropout > 0.0:
cell = tf.contrib.rnn.DropoutWrapper(
cell=cell, input_keep_prob=(1.0 - dropout))
cell_list.append(cell)
if len(cell_list) == 1: # Single layer.
fw_cell = cell_list[0]
bw_cell = cell_list[0]
else: # Multi layers
fw_cell = tf.contrib.rnn.MultiRNNCell(cell_list)
bw_cell = tf.contrib.rnn.MultiRNNCell(cell_list)
bi_outputs, bi_state = tf.nn.bidirectional_dynamic_rnn(
fw_cell,
bw_cell,
encoder_inputs_embedded,
dtype=dtype,
sequence_length=encoder_inputs_length,
time_major=False,
swap_memory=True)
encoder_outputs, bi_encoder_state = tf.concat(bi_outputs, -1), bi_state
if num_bi_layers == 1:
encoder_last_state = bi_encoder_state
else:
# alternatively concat forward and backward states
encoder_state = []
for layer_id in range(num_bi_layers):
encoder_state.append(bi_encoder_state[0][layer_id]) # forward
encoder_state.append(bi_encoder_state[1][layer_id]) # backward
encoder_last_state = tuple(encoder_state)
我认为错误是在binnn状态提取中，但令我困惑的是源代码实际上有效，表明这是正确的。我按照README步骤执行它，并使用整个模型完成所有操作......
此外，您可以按照整个代码（Jupyter笔记本上的整个错误）：
https://github.com/denisb411/seq2seq-NMT-tensorflow/blob/master/seq2seq_tensorflow_1_4.ipynb
追溯：