def bidirectionnal_rnn(self, input_tensor, input_sequence_length):
    lstm_num_units = self.config.lstm_num_units
    print("rnn input tensor ===> ", input_tensor)
    with tf.variable_scope('lstm_layers'):
      fw_cell_list = [rnn.BasicLSTMCell(nh, forget_bias=1.0, name='fw_cell_%d'%(nh)) for nh in [lstm_num_units] * 2]
      bw_cell_list = [rnn.BasicLSTMCell(nh, forget_bias=1.0, name='bw_cell_%d'%(nh)) for nh in [lstm_num_units] * 2]

      stack_lstm_layer, _, _ = rnn.stack_bidirectional_dynamic_rnn(
          cells_fw=fw_cell_list, 
          cells_bw=bw_cell_list, 
          inputs=input_tensor, 
          sequence_length=input_sequence_length, 
          dtype=tf.float32)
      hidden_num = lstm_num_units * 2
      rnn_reshaped = tf.nn.dropout(stack_lstm_layer, keep_prob=self.keep_prob)
      w = tf.get_variable(initializer=tf.truncated_normal([hidden_num, self.num_classes], stddev=0.02), name="w")
      w_t = tf.tile(tf.expand_dims(w, 0),[self.batch_size,1,1])
      logits = tf.matmul(rnn_reshaped, w_t, name="nn_logits")
      self.logits = tf.identity(tf.transpose(logits, (1, 0, 2)), name='logits')
      return logits 

def _sequence_label(self, inputdata, name):
        with tf.variable_scope(name_or_scope=name):
            fw_cell_list = [tf.nn.rnn_cell.LSTMCell(nh, forget_bias=1.0) for
                            nh in [self._hidden_nums] * self._layers_nums]
            # Backward direction cells
            bw_cell_list = [tf.nn.rnn_cell.LSTMCell(nh, forget_bias=1.0) for
                            nh in [self._hidden_nums] * self._layers_nums]

            stack_lstm_layer, _, _ = rnn.stack_bidirectional_dynamic_rnn(fw_cell_list,
                bw_cell_list, inputdata, #sequence_length=CFG.ARCH.SEQ_LENGTH * np.ones(CFG.TRAIN.BATCH_SIZE),
            dtype=tf.float32)
            #stack_lstm_layer = self.dropout(inputdata=stack_lstm_layer, keep_prob=0.5,\
            #    is_training=self._is_training, name='sequence_drop_out')

            [batch_s, _, hidden_nums] = inputdata.get_shape().as_list()  # [batch, width, 2*n_hidden]

            shape = tf.shape(stack_lstm_layer)
            rnn_reshaped = tf.reshape(stack_lstm_layer, [shape[0] * shape[1], shape[2]])

            w = tf.get_variable(name='w',shape=[hidden_nums, self._num_classes],\
                initializer=tf.truncated_normal_initializer(stddev=0.02),trainable=True)
            # Doing the affine projection
            logits = tf.matmul(rnn_reshaped, w, name='logits')
            logits = tf.reshape(logits, [shape[0], shape[1], self._num_classes], name='logits_reshape')
            raw_pred = tf.argmax(tf.nn.softmax(logits), axis=2, name='raw_prediction')
            # Swap batch and batch axis
            rnn_out = tf.transpose(logits, [1, 0, 2], name='transpose_time_major')  # [width, batch, n_classes]

        return rnn_out, raw_pred 

def __sequence_label(self, inputdata):
        """
        Implement the sequence label part of the network
        :param inputdata:
        :return:
        """
        with tf.variable_scope('LSTMLayers'):
            # construct stack lstm rcnn layer
            # forward lstm cell
            fw_cell_list = [rnn.BasicLSTMCell(nh, forget_bias=1.0) for nh in [self.__hidden_nums, self.__hidden_nums]]
            # Backward direction cells
            bw_cell_list = [rnn.BasicLSTMCell(nh, forget_bias=1.0) for nh in [self.__hidden_nums, self.__hidden_nums]]

            stack_lstm_layer, _, _ = rnn.stack_bidirectional_dynamic_rnn(fw_cell_list, bw_cell_list, inputdata,
                                                                         dtype=tf.float32)

            if self.phase.lower() == 'train':
                stack_lstm_layer = self.dropout(inputdata=stack_lstm_layer, keep_prob=0.5)

            [batch_s, _, hidden_nums] = inputdata.get_shape().as_list()  # [batch, width, 2*n_hidden]
            rnn_reshaped = tf.reshape(stack_lstm_layer, [-1, hidden_nums])  # [batch x width, 2*n_hidden]

            w = tf.Variable(tf.truncated_normal([hidden_nums, self.__num_classes], stddev=0.1), name="w")
            # Doing the affine projection

            logits = tf.matmul(rnn_reshaped, w)

            logits = tf.reshape(logits, [batch_s, -1, self.__num_classes])

            raw_pred = tf.argmax(tf.nn.softmax(logits), axis=2, name='raw_prediction')

            # Swap batch and batch axis
            rnn_out = tf.transpose(logits, (1, 0, 2), name='transpose_time_major')  # [width, batch, n_classes]

        return rnn_out, raw_pred 

def __sequence_label(self, inputdata):
        """
        Implement the sequence label part of the network
        :param inputdata:
        :return:
        """
        with tf.variable_scope('LSTMLayers'):
            # construct stack lstm rcnn layer
            # forward lstm cell
            fw_cell_list = [rnn.BasicLSTMCell(nh, forget_bias=1.0) for nh in [self.__hidden_nums, self.__hidden_nums]]
            # Backward direction cells
            bw_cell_list = [rnn.BasicLSTMCell(nh, forget_bias=1.0) for nh in [self.__hidden_nums, self.__hidden_nums]]

            stack_lstm_layer, _, _ = rnn.stack_bidirectional_dynamic_rnn(fw_cell_list, bw_cell_list, inputdata,
                                                                         dtype=tf.float32)

            if self.phase.lower() == 'train':
                stack_lstm_layer = self.dropout(inputdata=stack_lstm_layer, keep_prob=0.5)

            [batch_s, _, hidden_nums] = inputdata.get_shape().as_list()  # [batch, width, 2*n_hidden]
            rnn_reshaped = tf.reshape(stack_lstm_layer, [-1, hidden_nums])  # [batch x width, 2*n_hidden]

            w = tf.Variable(tf.truncated_normal([hidden_nums, self.__num_classes], stddev=0.1), name="w")
            # Doing the affine projection

            logits = tf.matmul(rnn_reshaped, w)

            logits = tf.reshape(logits, [batch_s, -1, self.__num_classes])

            raw_pred = tf.argmax(tf.nn.softmax(logits), axis=2, name='raw_prediction')

            # Swap batch and batch axis
            rnn_out = tf.transpose(logits, (1, 0, 2), name='transpose_time_major')  # [width, batch, n_classes]

        return rnn_out, raw_pred 

def bi_gru(self, inputs):
        """build the bi-GRU network. 返回个所有层的隐含状态。"""
        cells_fw = [self.gru_cell() for _ in range(self.n_layer)]
        cells_bw = [self.gru_cell() for _ in range(self.n_layer)]
        initial_states_fw = [cell_fw.zero_state(self.batch_size, tf.float32) for cell_fw in cells_fw]
        initial_states_bw = [cell_bw.zero_state(self.batch_size, tf.float32) for cell_bw in cells_bw]
        outputs, _, _ = rnn.stack_bidirectional_dynamic_rnn(cells_fw, cells_bw, inputs,
                                                            initial_states_fw=initial_states_fw,
                                                            initial_states_bw=initial_states_bw, dtype=tf.float32)
        return outputs 

def _sequence_label(self, inputdata, name):
        """ Implements the sequence label part of the network

        :param inputdata:
        :param name:
        :return:
        """
        with tf.variable_scope(name_or_scope=name):
            # construct stack lstm rcnn layer
            # forward lstm cell
            fw_cell_list = [tf.nn.rnn_cell.LSTMCell(nh, forget_bias=1.0) for
                            nh in [self._hidden_nums] * self._layers_nums]
            # Backward direction cells
            bw_cell_list = [tf.nn.rnn_cell.LSTMCell(nh, forget_bias=1.0) for
                            nh in [self._hidden_nums] * self._layers_nums]

            stack_lstm_layer, _, _ = rnn.stack_bidirectional_dynamic_rnn(
                fw_cell_list, bw_cell_list, inputdata,
                dtype=tf.float32
            )
            stack_lstm_layer = self.dropout(
                inputdata=stack_lstm_layer,
                keep_prob=0.5,
                is_training=self._is_training,
                name='sequence_drop_out'
            )

            [batch_s, _, hidden_nums] = inputdata.get_shape().as_list()  # [batch, width, 2*n_hidden]

            shape = tf.shape(stack_lstm_layer)
            rnn_reshaped = tf.reshape(stack_lstm_layer, [shape[0] * shape[1], shape[2]])

            w = tf.get_variable(
                name='w',
                shape=[hidden_nums, self._num_classes],
                initializer=tf.truncated_normal_initializer(stddev=0.02),
                trainable=True
            )

            # Doing the affine projection
            logits = tf.matmul(rnn_reshaped, w, name='logits')

            logits = tf.reshape(logits, [shape[0], shape[1], self._num_classes], name='logits_reshape')

            raw_pred = tf.argmax(tf.nn.softmax(logits), axis=2, name='raw_prediction')

            # Swap batch and batch axis
            rnn_out = tf.transpose(logits, [1, 0, 2], name='transpose_time_major')  # [width, batch, n_classes]

        return rnn_out, raw_pred