张量流中两种RNN实现的区别是什么?

Nil*_*Cao 5 python deep-learning lstm tensorflow recurrent-neural-network

我在tensorflow中找到了两种RNN实现.

第一个实现是这个(从第124行到第129行).它使用循环来定义RNN中的每个输入步骤.

with tf.variable_scope("RNN"):
      for time_step in range(num_steps):
        if time_step > 0: tf.get_variable_scope().reuse_variables()
        (cell_output, state) = cell(inputs[:, time_step, :], state)
        outputs.append(cell_output)
        states.append(state)
Run Code Online (Sandbox Code Playgroud)

第二个实现是这个(从第51行到第70行).它不使用任何循环来定义RNN中的每个输入步骤.

def RNN(_X, _istate, _weights, _biases):

    # input shape: (batch_size, n_steps, n_input)
    _X = tf.transpose(_X, [1, 0, 2])  # permute n_steps and batch_size
    # Reshape to prepare input to hidden activation
    _X = tf.reshape(_X, [-1, n_input]) # (n_steps*batch_size, n_input)
    # Linear activation
    _X = tf.matmul(_X, _weights['hidden']) + _biases['hidden']

    # Define a lstm cell with tensorflow
    lstm_cell = rnn_cell.BasicLSTMCell(n_hidden, forget_bias=1.0)
    # Split data because rnn cell needs a list of inputs for the RNN inner loop
    _X = tf.split(0, n_steps, _X) # n_steps * (batch_size, n_hidden)

    # Get lstm cell output
    outputs, states = rnn.rnn(lstm_cell, _X, initial_state=_istate)

    # Linear activation
    # Get inner loop last output
    return tf.matmul(outputs[-1], _weights['out']) + _biases['out']
Run Code Online (Sandbox Code Playgroud)


在第一个实现中,我发现输入单元到隐藏单元之间没有权重矩阵,只定义隐藏单元到输出单元之间的权重矩阵(从132到133行).

output = tf.reshape(tf.concat(1, outputs), [-1, size])
        softmax_w = tf.get_variable("softmax_w", [size, vocab_size])
        softmax_b = tf.get_variable("softmax_b", [vocab_size])
        logits = tf.matmul(output, softmax_w) + softmax_b
Run Code Online (Sandbox Code Playgroud)

但是在第二种实现中,定义了两个权重矩阵(从第42行到第47行).

weights = {
    'hidden': tf.Variable(tf.random_normal([n_input, n_hidden])), # Hidden layer weights
    'out': tf.Variable(tf.random_normal([n_hidden, n_classes]))
}
biases = {
    'hidden': tf.Variable(tf.random_normal([n_hidden])),
    'out': tf.Variable(tf.random_normal([n_classes]))
}
Run Code Online (Sandbox Code Playgroud)

我想知道为什么?

Ara*_*lla 3

我注意到的区别是,第二个实现中的代码使用 tf.nn.rnn,它获取每个时间步的输入列表并生成每个时间步的输出列表。

(输入:长度为 T 的输入列表,每个输入都是形状为 [batch_size, input_size] 的张量。)

因此,如果您检查第 62 行第二个实现中的代码,输入数据将被整形为 n_steps * (batch_size, n_hidden)

# Split data because rnn cell needs a list of inputs for the RNN inner loop
_X = tf.split(0, n_steps, _X) # n_steps * (batch_size, n_hidden)
Run Code Online (Sandbox Code Playgroud)

第一个实现中,它们循环遍历 n_time_steps 并提供输入并获取相应的输出并存储在输出列表中。

第 113 行到 117 行的代码片段

outputs = []
    state = self._initial_state
    with tf.variable_scope("RNN"):
      for time_step in range(num_steps):
        if time_step > 0: tf.get_variable_scope().reuse_variables()
        (cell_output, state) = cell(inputs[:, time_step, :], state)
        outputs.append(cell_output)
Run Code Online (Sandbox Code Playgroud)

接下来回答你的第二个问题:

如果您仔细观察这两种实现中输入馈送到 RNN 的方式。

在第一个实现中,输入的形状已经是batch_size x num_steps(这里num_steps是隐藏大小):

self._input_data = tf.placeholder(tf.int32, [batch_size, num_steps])
Run Code Online (Sandbox Code Playgroud)

而在第二个实现中,初始输入的形状为(batch_size x n_steps x n_input)。因此需要一个权重矩阵来转换为形状(n_steps x batch_size xhidden_​​size):

    # Input shape: (batch_size, n_steps, n_input)
    _X = tf.transpose(_X, [1, 0, 2])  # Permute n_steps and batch_size
    # Reshape to prepare input to hidden activation
    _X = tf.reshape(_X, [-1, n_input]) # (n_steps*batch_size, n_input)
    # Linear activation
    _X = tf.matmul(_X, _weights['hidden']) + _biases['hidden']
    # Split data because rnn cell needs a list of inputs for the RNN inner loop
    _X = tf.split(0, n_steps, _X) # n_steps * (batch_size, n_hidden)
Run Code Online (Sandbox Code Playgroud)

我希望这是有帮助的...