Variable rnn/multi_rnn_cell/cell_0/basic_lstm_cell/kernel already exists问题解决方法

问题出现背景：博主在同一个tf.name_scope(name)下同时处理音频和文本的序列特征时，使用两次tf.nn.dynamic_rnn（）函数时出现该错误：

ValueError: Variable rnn/multi_rnn_cell/cell_0/basic_lstm_cell/kernel already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?

问题出现分析：同时调用两次RNN时，第一次调用RNN时没有任何问题，关键是出在第二次调用RNN时，由于第一次tf.nn.dynamic_rnn（）处理音频序列时会产生一个权重，而再次调用RNN，再次使用tf.nn.dynamic_rnn（）处理文本序列时，这时候系统里应该存在两个不同的lstm_cell模型，但是系统无法辨别出来，因此会出现kernel already exists。（此处参考https://blog.csdn.net/u013041398/article/details/74941991)

问题解决：需要用tf.variable_scope(name)来定义不同的作用范围。

贴上报错代码：

def dropout():
    if self.config.rnn == 'lstm':
        cell = tf.contrib.rnn.BasicLSTMCell(self.config.hidden_dim)
    else:
        cell = tf.contrib.rnn.GRUCell(self.config.hidden_dim)
    return tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=self.keep_prob)
with tf.name_scope("rnn"):

cells1 = [dropout() for _ in range(self.config.num_layers)]
rnn_cell1 = tf.contrib.rnn.MultiRNNCell(cells1, state_is_tuple=True)

audio_value, _ = tf.nn.dynamic_rnn(cell=rnn_cell1, inputs=self.audio_encoder_inputs, dtype=tf.float32)
audio_value = tf.transpose(audio_value, [1, 0, 2])
self.audio_encoder = tf.gather(audio_value, int(audio_value.get_shape()[0]) - 1)
self.final_audio_encoder = tf.concat([self.audio_encoder, self.encoder_prosody], axis=1)
cells2 = [dropout() for _ in range(self.config.num_layers)]
rnn_cell2 = tf.contrib.rnn.MultiRNNCell(cells2, state_is_tuple=True)
data = tf.nn.embedding_lookup(self.wordVectors, self.text_encoder_inputs) 
data = tf.cast(data, tf.float32) 
text_value, _ = tf.nn.dynamic_rnn(cell=rnn_cell2, inputs=data, dtype=tf.float32)

修改后的代码：

def dropout():
    with tf.variable_scope('lstmcell', reuse=True):
        if self.config.rnn == 'lstm':
            cell = tf.contrib.rnn.BasicLSTMCell(self.config.hidden_dim)
        else:
            cell = tf.contrib.rnn.GRUCell(self.config.hidden_dim)
    return tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=self.keep_prob)

with tf.name_scope("rnn"):
    with tf.variable_scope("audio_rnn"):
        cells1 = [dropout() for _ in range(self.config.num_layers)]
        rnn_cell1 = tf.contrib.rnn.MultiRNNCell(cells1, state_is_tuple=True)

        audio_value, _ = tf.nn.dynamic_rnn(cell=rnn_cell1, inputs=self.audio_encoder_inputs, dtype=tf.float32)
        audio_value = tf.transpose(audio_value, [1, 0, 2])
        self.audio_encoder = tf.gather(audio_value, int(audio_value.get_shape()[0]) - 1)
        self.final_audio_encoder = tf.concat([self.audio_encoder, self.encoder_prosody], axis=1)
        self.audio_batch_pred = tf.layers.dense(self.final_audio_encoder,
                                                (self.config.hidden_dim + self.config.AUDIO_PROSODY) // 2,
                                                name='final_audio_encoder')
    with tf.variable_scope("text_rnn"):
        cells2 = [dropout() for _ in range(self.config.num_layers)]
        rnn_cell2 = tf.contrib.rnn.MultiRNNCell(cells2, state_is_tuple=True)
        data = tf.nn.embedding_lookup(self.wordVectors, self.text_encoder_inputs)
        data = tf.cast(data, tf.float32)  
        text_value, _ = tf.nn.dynamic_rnn(cell=rnn_cell2, inputs=data, dtype=tf.float32)
        text_value = tf.transpose(text_value, [1, 0, 2])