从bidirectional_rnn切换到bidirectional_dynamic_rnn后，结果变得更糟

Question

基本上是尝试用bidirectional_dynamic_rnn (重塑输入)来替代bidirectional_rnn，但在分类任务上得到的结果要差得多。我做错了什么吗？重塑？

bidirectional_rnn版本(代码摘录)：

encoder_embedded_inputs = [embedding_ops.embedding_lookup(
                    W, encoder_input) for encoder_input in encoder_inputs]

encoder_outputs, encoder_state_fw, encoder_state_bw  = rnn.bidirectional_rnn(
                    encoder_cell_fw,
                    encoder_cell_bw,
                    encoder_embedded_inputs,
                    sequence_length=sequence_length,
                    dtype=dtype)

encoder_state = array_ops.concat(1, [array_ops.concat(
                1, encoder_state_fw), array_ops.concat(1, encoder_state_bw)])
top_states = [array_ops.reshape(e, [-1, 1, cell.output_size * 2])
                              for e in encoder_outputs]
attention_states = array_ops.concat(1, top_states)

分类准确率: 95%

bidirectional_dynamic_rnn版本(代码摘录)：

encoder_embedded_inputs = [embedding_ops.embedding_lookup(
                    W, encoder_input) for encoder_input in encoder_inputs]
emb_size = int(encoder_embedded_inputs[0].get_shape()[1])
enc_size = len(encoder_embedded_inputs)
birnn_inputs = tf.reshape(encoder_embedded_inputs, [-1,enc_size,emb_size])

encoder_outputs, encoder_states  = rnn.bidirectional_dynamic_rnn(
                    encoder_cell_fw,
                    encoder_cell_bw,
                    birnn_inputs,
                    sequence_length=sequence_length,
                    dtype=dtype)
encoder_state_fw, encoder_state_bw = encoder_states
encoder_state = array_ops.concat(1, [array_ops.concat(
                1, encoder_state_fw), array_ops.concat(1, encoder_state_bw)])

attention_states = tf.concat(2, encoder_outputs)

分类准确率: 70%

Answer 1

好吧，我发现tf.reshape不适合这个任务，我应该用tf.stack和tf.transpose来代替。因此，它基本上是在使用混乱的输入，并且不再能够学习。

错误：

birnn_inputs = tf.reshape(encoder_embedded_inputs, [-1,enc_size,emb_size])

右图：

birnn_inputs = tf.stack(encoder_embedded_inputs)
birnn_inputs = tf.transpose(birnn_inputs, [1,0,2])

所以现在它运行得很好。