什么时候使用birnn的输出和状态？用于关系提取任务

Question

我正在研究关系提取。众所周知，tf.nn.bidirectional_dynamic_rnn的输出包括outputs和output_states。在阅读代码时，我发现有些使用outputs，而有些使用output_states。哪一个更好？我该如何选择？

情况1：

outputs, states = tf.nn.bidirectional_dynamic_rnn(fw_cell, bw_cell, x, sequence_length = sequence_length, dtype = tf.float32, scope = 'dynamic-bi-rnn')
fw_outputs, bw_outputs = outputs
outputs = tf.add(fw_outputs, bw_outputs)

案例2：

_, states = tf.nn.bidirectional_dynamic_rnn(fw_cell, bw_cell, x, sequence_length = sequence_length, dtype = tf.float32, scope = 'dynamic-bi-rnn')
fw_states, bw_states = states
if isinstance(fw_states, tuple):
    fw_states = fw_states[0]
    bw_states = bw_states[0]
return tf.concat([fw_states, bw_states], axis = 1)

请帮助我，非常感谢！