Tensorflow dynamic_rnn 输入排名错误

Question

所以，我正在尝试使用 tensorflow 中的 rnn 来生成文本。但是，一旦我从 static_rnn 切换到 dynamic_rnn，我就得到了这个错误：

File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/tensor_shape.py", line 654, in with_rank_at_least
    raise ValueError("Shape %s must have rank at least %d" % (self, rank))
ValueError: Shape (100, 5) must have rank at least 3

这是产生错误的代码部分：

inputs_series = self.input_layer()
with tf.variable_scope(constants.HIDDEN):
    self.hidden_state_placeholder = tf.placeholder(
        dtype=tf.float32, 
        shape=[self.settings.train.batch_size, self.settings.rnn.hidden_size],
        name="hidden_state_placeholder")
    cell = tf.contrib.rnn.GRUCell(self.settings.rnn.hidden_size)
    states_series, self.current_state = tf.nn.dynamic_rnn(
        cell=cell, 
        inputs=inputs_series,
        initial_state=self.hidden_state_placeholder)

inputs_series的shape为：(10,5,100),对应(截断的文字长度,batch size,类的个数)

hidden_state_placeholder 的形状是 (5, 100) for (batch size, hidden state size)，但即使我不提供初始状态，错误仍然存在。

tensorflow版本是1.3，如果有帮助。

如有任何见解，我们将不胜感激！

Answer 1

默认情况下，time_major == False 在 tf.nn.dynamic_rnn，但你的 inputs_series 是 time_major == True。所以也许将最后一条语句更改为

states_series, self.current_state = tf.nn.dynamic_rnn(
    cell=cell, 
    inputs=inputs_series,
    initial_state=self.hidden_state_placeholder,
    time_major=True)

Tensorflow dynamic_rnn 输入排名错误

Tensorflow dynamic_rnn input rank error

tensorflow

recurrent-neural-network