Question

我正在使用支持GPU的Tensorflow 1.1.0，我有这个功能：

def get_init_cell(batch_size, rnn_size, keep_prob=0.75, layers=2):
    """
    Create an RNN Cell and initialize it.
    :param batch_size: Size of batches
    :param rnn_size: Size of RNNs
    :return: Tuple (cell, initialize state)
    """
    # Basic LSTM cell
    # lstm = tf.contrib.rnn.BasicLSTMCell(rnn_size)
    lstm = tf.contrib.rnn.BasicLSTMCell(rnn_size, forget_bias=0.0, state_is_tuple=True, reuse=tf.get_variable_scope().reuse)

    # Add drop to the cell
    drop = tf.contrib.rnn.DropoutWrapper(lstm, output_keep_prob=keep_prob)

    # Stack multiple LSTM layers
    rnn_cell = tf.contrib.rnn.MultiRNNCell([drop for _ in range(layers)], state_is_tuple=True)

    # Getting an initial state of zeros
    initial_state = rnn_cell.zero_state(batch_size, tf.float32)

    # Set the name
    initial_state = tf.identity(initial_state, 'initial_state')

    return rnn_cell, initial_state

如果我使用lstm = tf.contrib.rnn.BasicLSTMCell(rnn_size)，之前的代码与Tensorflow 1.0.0完美配合，但现在我改变了我获取lstm单元格的方式。我收到了这个错误：

ValueError：尝试重用RNNCell 与其不同的变量范围第一次使用。细胞的首次使用是在范围内 'rnn / multi_rnn_cell / cell_0 / basic_lstm_cell'，这个尝试是与范围'rnn / multi_rnn_cell / cell_1 / basic_lstm_cell'。请创建一个如果您希望它使用不同的集合，则单元格的新实例重量。如果在使用之前： MultiRNNCell（[BasicLSTMCell（...）] * num_layers），更改为： MultiRNNCell（[BasicLSTMCell（...）for _ in range（num_layers）]）。如果在你使用相同的单元格实例之前，前进和双向RNN的反向小区，只需创建两个实例（一个前进，后退一个）。 2017年5月，我们将开始转换此单元格的行为以使用现有存储的权重，如果 any，当使用scope = None调用它时（可以导致静默模型）降级，所以这个错误将一直存在。）

我用这个更改了get_init_cell：

def lstm_cell(rnn_size, keep_prob=0.75):
    # Basic LSTM cell
    lstm = tf.contrib.rnn.BasicLSTMCell(rnn_size, reuse=tf.get_variable_scope().reuse)

    # Add drop to the cell
    return tf.contrib.rnn.DropoutWrapper(lstm, output_keep_prob=keep_prob)

def get_init_cell(batch_size, rnn_size, keep_prob=0.75, layers=2):
    """
    Create an RNN Cell and initialize it.
    :param batch_size: Size of batches
    :param rnn_size: Size of RNNs
    :return: Tuple (cell, initialize state)
    """    
    # Stack multiple LSTM layers
    rnn_cell = tf.contrib.rnn.MultiRNNCell([lstm_cell(rnn_size, keep_prob) for _ in range(layers)])

    # Getting an initial state of zeros
    initial_state = rnn_cell.zero_state(batch_size, tf.float32)

    # Set the name
    initial_state = tf.identity(initial_state, 'initial_state')

    return rnn_cell, initial_state

我仍然遇到同样的错误。

我也尝试过同样的错误：

def get_init_cell(batch_size, rnn_size, keep_prob=0.75, layers=2):
    """
    Create an RNN Cell and initialize it.
    :param batch_size: Size of batches
    :param rnn_size: Size of RNNs
    :return: Tuple (cell, initialize state)
    """    

    def lstm_cell():
        # Basic LSTM cell
        return tf.contrib.rnn.BasicLSTMCell(rnn_size, reuse=tf.get_variable_scope().reuse)        

    def attn_cell():
        return tf.contrib.rnn.DropoutWrapper(lstm_cell(), output_keep_prob=keep_prob)

    # Stack multiple LSTM layers
    rnn_cell = tf.contrib.rnn.MultiRNNCell([attn_cell() for _ in range(layers)], state_is_tuple=True)

    # Getting an initial state of zeros
    initial_state = rnn_cell.zero_state(batch_size, tf.float32)

    # Set the name
    initial_state = tf.identity(initial_state, 'initial_state')

    return rnn_cell, initial_state

有什么想法吗？

尝试使用不同的变量范围重用RNNCell

0 个答案: