1

我正在实现一个具有多层双向 rnn 和注意力机制的 Seq2Seq 模型，在遵循本教程https://github.com/tensorflow/nmt时，我对如何在双向层之后正确操作 encoder_state 感到困惑。

引用教程“对于多个双向层，我们需要稍微操纵encoder_state，有关更多详细信息，请参见model.py，方法_build_bidirectional_rnn()。” 这是代码的相关部分（https://github.com/tensorflow/nmt/blob/master/nmt/model.py第 770 行）：

encoder_outputs, bi_encoder_state = (
            self._build_bidirectional_rnn(
            inputs=self.encoder_emb_inp,
            sequence_length=sequence_length,
            dtype=dtype,
            hparams=hparams,
            num_bi_layers=num_bi_layers,
            num_bi_residual_layers=num_bi_residual_layers))

if num_bi_layers == 1:
   encoder_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_state = tuple(encoder_state)

所以这就是我现在所拥有的：

def get_a_cell(lstm_size):
    lstm = tf.nn.rnn_cell.BasicLSTMCell(lstm_size)
    #drop = tf.nn.rnn_cell.DropoutWrapper(lstm, 
                       output_keep_prob=keep_prob)
    return lstm


encoder_FW = tf.nn.rnn_cell.MultiRNNCell(
    [get_a_cell(num_units) for _ in range(num_layers)])
encoder_BW = tf.nn.rnn_cell.MultiRNNCell(
    [get_a_cell(num_units) for _ in range(num_layers)])


bi_outputs, bi_encoder_state = tf.nn.bidirectional_dynamic_rnn(
encoder_FW, encoder_BW, encoderInput,
sequence_length=x_lengths, dtype=tf.float32)
encoder_output = tf.concat(bi_outputs, -1)

encoder_state = []

for layer_id in range(num_layers):
    encoder_state.append(bi_encoder_state[0][layer_id])  # forward
    encoder_state.append(bi_encoder_state[1][layer_id])  # backward
encoder_state = tuple(encoder_state)

#DECODER -------------------

decoder_cell = tf.nn.rnn_cell.MultiRNNCell([get_a_cell(num_units) for _ in range(num_layers)])

# Create an attention mechanism
attention_mechanism = tf.contrib.seq2seq.LuongAttention(num_units_attention, encoder_output ,memory_sequence_length=x_lengths)

decoder_cell = tf.contrib.seq2seq.AttentionWrapper(
              decoder_cell,attention_mechanism,
              attention_layer_size=num_units_attention)

decoder_initial_state = decoder_cell.zero_state(batch_size,tf.float32)
                        .clone(cell_state=encoder_state)

问题是我收到错误

The two structures don't have the same nested structure.

First structure: type=AttentionWrapperState 
str=AttentionWrapperState(cell_state=(LSTMStateTuple(c=, h=), 
LSTMStateTuple(c=, h=)), attention=, time=, alignments=, alignment_history=
(), attention_state=)

Second structure: type=AttentionWrapperState 
str=AttentionWrapperState(cell_state=(LSTMStateTuple(c=, h=), 
LSTMStateTuple(c=, h=), LSTMStateTuple(c=, h=), LSTMStateTuple(c=, h=)), 
attention=, time=, alignments=, alignment_history=(), attention_state=)

这对我来说有点道理，因为我们没有包括所有输出层，但（我猜）只包括最后一层。而对于状态，我们实际上是连接所有层。

所以正如我所期待的，当只连接最后一层状态时，如下所示：

encoder_state = []
encoder_state.append(bi_encoder_state[0][num_layers-1])  # forward
encoder_state.append(bi_encoder_state[1][num_layers-1])  # backward
encoder_state = tuple(encoder_state)

它运行没有错误。

据我所知，在将编码器状态传递到注意力层之前，没有任何部分代码会再次转换编码器状态。那么他们的代码是如何工作的呢？更重要的是，我的修复是否破坏了注意力机制的正确行为？

4

1 回答 1

0

这是问题所在：

只有编码器是双向的，但您为解码器提供双向状态（始终是单向的）。

解决方案：

您所要做的就是简单地连接状态，因此，您再次操纵“单向数据”！

encoder_state = []

for layer_id in range(num_layers):
    state_fw = bi_encoder_state[0][layer_id]
    state_bw = bi_encoder_state[1][layer_id]

    # Merging the fw state and the bw state
    cell_state = tf.concat([state_fw.c, state_bw.c], 1)
    hidden_state= tf.concat([state_fw.h, state_bw.h], 1)

    # This state as the same structure than an uni-directional encoder state
    state = tf.nn.rnn_cell.LSTMStateTuple(c=cell_state, h=hidden_state)

    encoder_state.append(state)

encoder_state = tuple(encoder_state)
于 2019-09-05T09:12:40.603 回答