Question

根据Tensorflow documentation for raw RNN，您可以将cell_output设置为您自己的形状作为初始时间步，当time = 0时，为了发出您自己的输出。

目前，它只有在我们将initial（for time = 0）单元格输出设置为None且其余发出的单元格输出等于前一个单元格输出时才有效，但这不是我们想要的。我们希望发出自己的输出，而不是RNN单元的内部状态。

以下是相关的源代码：

    def get_loop_fn(self, encoder_final_state, eos_token, batch_size, WW, bb):
    def loop_fn(time, previous_output, previous_state, previous_loop_state):
        # inputs:  time, previous_cell_output, previous_cell_state, previous_loop_state
        # outputs: elements_finished, input, cell_state, output, loop_state

        if previous_state is None:  # time == 0
            assert previous_output is None
            return self.loop_fn_initial(encoder_final_state=encoder_final_state, eos_token=eos_token,
                                        batch_size=batch_size)
        else:
            return self.loop_fn_transition(previous_cell_output=previous_output, previous_cell_state=previous_state,
                                           batch_size=batch_size, WW=WW, bb=bb)

    return loop_fn

def loop_fn_initial(self, encoder_final_state, eos_token, batch_size):
    # https://www.tensorflow.org/api_docs/python/tf/nn/raw_rnn
    # here we have a static rnn case so all length so be taken into account,
    # so I guess they are all False always
    # From documentation: a boolean Tensor of shape [batch_size]
    initial_elements_finished = self.all_elems_non_finished(batch_size=batch_size)

    initial_input = eos_token
    initial_cell_state = encoder_final_state

    # initial_cell_output = None
    # give it the shape that we want:
    initial_cell_output = tf.constant(np.zeros(shape=(batch_size, self.TARGET_FEATURE_LEN)), dtype=tf.float32)

    initial_loop_state = None  # we don't need to pass any additional information

    return (initial_elements_finished,
            initial_input,
            initial_cell_state,
            initial_cell_output,
            initial_loop_state)

def loop_fn_transition(self, previous_cell_output, previous_cell_state, batch_size, WW, bb):
    """note that the matrix W is going to be shared among outputs"""
    print "previous cell output"
    print previous_cell_output
    print

    finished = self.all_elems_non_finished(batch_size=batch_size)  # (time >= decoder_lengths)
    # this operation produces boolean tensor of [batch_size] defining if corresponding sequence has ended

    # this is always false in our case so just comment next two lines
    # finished = tf.reduce_all(elements_finished)  # -> boolean scalar
    # input = tf.cond(finished, lambda: pad_step_embedded, get_next_input)
    next_input = tf.add(tf.matmul(previous_cell_output, WW), bb)

    next_cell_state = previous_cell_state
    emit_output = next_input,  # previous_cell_output

    next_loop_state = None  # we don't need to pass any additional information

    return (finished,
            next_input,
            next_cell_state,
            emit_output,
            next_loop_state)

这是我们在尝试创建图表时收到的错误消息：

ValueError：这两个结构没有相同的嵌套结构。第一种结构：，第二种结构:(，）。

欢迎任何想法

如何使原始rnn发出自定义输出而不是默认值？

0 个答案: