尝试通过tensorflow中的bidirectional_rnn连接正向状态和反向状态输出

Question

我希望通过tf.nn.bidirectional_rnn（）连接最终的前向状态和后向状态输出，并将其用作后续softmax层的输入。但我失败了。 显示错误的代码在这里：

PROFESSOR(NAME,OFFICE,DEPARTMENT,AGE)
COURSE(CNO,TITLE,DEPARTMENT)
ENROLL(CNO,SEMESTER,INSTRUCTOR_NAME,ENROLLMENT)

，输出如下：

import tensorflow as tf

MAX_SENTENCE_LENGTH = 50


class RNNModel(object):
    '''RNNModel'''
    def __init__(self, config):
        self.batch_size = batch_size = config.batch_size
        self.num_steps = num_steps = config.num_steps
        size = config.hidden_size
        label_size = config.label_size
        embedding_size = config.embedding_size

        self._input_data = tf.placeholder(tf.int32, [None, num_steps], name="input_data")
        self._targets = tf.placeholder(tf.float32, [None, label_size], name="targets")

        lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(size, forget_bias=0.0, state_is_tuple=True)
        cell_fw = tf.nn.rnn_cell.MultiRNNCell([lstm_cell] * config.num_layers, state_is_tuple=True)
        cell_bw = tf.nn.rnn_cell.MultiRNNCell([lstm_cell] * config.num_layers, state_is_tuple=True)

        self._initial_state_fw = cell_fw.zero_state(batch_size, tf.float16)
        self._initial_state_bw = cell_bw.zero_state(batch_size, tf.float16)

        with tf.device('/cpu:0'):
            self._embedding = embedding = tf.get_variable("embedding", [config.vocab_size, embedding_size],
                                                          dtype=tf.float16, trainable=False)
            inputs = tf.nn.embedding_lookup(embedding, self._input_data)

        # embedding size to hidden size
        inputs2 = [tf.squeeze(s, squeeze_dims=[1]) for s in tf.split(1, num_steps, inputs)]
        # outputs, state = tf.nn.rnn(cell, inputs2, initial_state=self._initial_state)
        outputs, state_fw, state_bw = tf.nn.bidirectional_rnn(
            cell_fw, cell_bw, inputs2, initial_state_fw=self._initial_state_fw, initial_state_bw=self._initial_state_bw)
        # # use the mean of outputs in all time steps as output
        # output = tf.reduce_mean(outputs, 0)
        # use the final state as output
        c_fw, _ = state_fw[-1]
        c_bw, _ = state_bw[-1]
        print(c_fw)
        output = tf.concat(1, [c_fw, c_bw])

        softmax_w = tf.get_variable("softmax_w", [size * 2, label_size], dtype=tf.float16)
        softmax_b = tf.get_variable("softmax_b", [label_size], dtype=tf.float16)
        self.logits = logits = tf.add(tf.matmul(output, softmax_w), softmax_b, name="predict")
        loss = tf.nn.softmax_cross_entropy_with_logits(logits, self._targets)
        self._cost = cost = tf.reduce_mean(loss) + 0.5 * config.l2 * tf.nn.l2_loss(softmax_w)
        self._final_state_fw = state_fw
        self._final_state_bw = state_bw

        self._lr = tf.Variable(0.0, dtype=tf.float16, trainable=False)
        tvars = tf.trainable_variables()
        grads, _ = tf.clip_by_global_norm(tf.gradients(cost, tvars), config.max_grad_norm)
        optimizer = tf.train.GradientDescentOptimizer(self.lr)
        self._train_op = optimizer.apply_gradients(zip(grads, tvars))


class Config(object):
    '''Small Config'''
    init_scale = 0.1
    learning_rate = 1.0
    max_grad_norm = 5
    num_layers = 3
    num_steps = MAX_SENTENCE_LENGTH
    hidden_size = 200
    lr_decay = 0.5
    batch_size = 20
    vocab_size = 1570974
    label_size = 22
    embedding_size = 300
    l2 = 0.001

if __name__ == "__main__":
    config = Config()
    model = RNNModel(config)

我想知道导致此错误的原因，以及如何实现我的想法。 谢谢你的帮助！