python - Tensorflow LSTM 抛出 ValueError : Shape () must have rank at least 2

Question

尝试运行时，抛出以下异常(ValueError)

ValueError: Shape () must have rank at least 2

这是针对以下行抛出的:

states_series, current_state = tf.contrib.rnn.static_rnn(cell, inputs_series, init_state)

这里定义了cell:

cell = tf.contrib.rnn.BasicLSTMCell(state_size, state_is_tuple=True)

查看 RNN 的规则和 Tesor_shape ，我可以看出这是某种张量维度形状问题。据我所知，它无法将 BasicLSTMCell 视为 2 阶矩阵？

完整错误:

/Library/Frameworks/Python.framework/Versions/3.6/bin/python3.6 /Users/glennhealy/PycharmProjects/firstRNNTest/LSTM-RNN.py
/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/importlib/_bootstrap.py:219: RuntimeWarning: compiletime version 3.5 of module 'tensorflow.python.framework.fast_tensor_util' does not match runtime version 3.6
  return f(*args, **kwds)
Traceback (most recent call last):
  File "/Users/glennhealy/PycharmProjects/firstRNNTest/LSTM-RNN.py", line 42, in <module>
    states_series, current_state = tf.contrib.rnn.static_rnn(cell, inputs_series, init_state)
  File "/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/tensorflow/python/ops/rnn.py", line 1181, in static_rnn
    input_shape = first_input.get_shape().with_rank_at_least(2)
  File "/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/tensorflow/python/framework/tensor_shape.py", line 670, in with_rank_at_least
    raise ValueError("Shape %s must have rank at least %d" % (self, rank))
ValueError: Shape () must have rank at least 2

Process finished with exit code 1

代码:

state_size = 4
cell = tf.contrib.rnn.BasicLSTMCell(state_size, state_is_tuple=True)
states_series, current_state = tf.contrib.rnn.static_rnn(cell, inputs_series, init_state)

tensorflow 1.2.1 Python 3.6 NumPy

更新更多信息:

考虑到@Maxim 给出的建议，我可以看出问题出在我的 input_series 上，这导致了形状问题，但是，我似乎无法理解他的建议。

一些有助于解决问题的更多信息，看看我是否能理解如何解决这个问题:

以下是否可以替代我的 BatchY 和 BatchX 占位符？

X = tf.placeholder(dtype=tf.float32, shape=[None, n_steps, n_inputs])
X_seqs = tf.unstack(tf.transpose(X, perm=[1, 0, 2]))
basic_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=n_neurons)
output_seqs, states = tf.nn.static_rnn(basic_cell, X_seqs,         dtype=tf.float32)

那么，我是否必须对以下内容进行更改以反射(reflect)以下内容的语法？

batchX_placeholder = tf.placeholder(tf.int32, [batch_size,      truncated_backprop_length])
batchY_placeholder = tf.placeholder(tf.float32, [batch_size,    truncated_backprop_length])

#unpacking the columns:
labels_series = tf.unstack(batchY_placeholder, axis=1)
inputs_series = tf.split(1, truncated_backprop_length, batchX_placeholder)

#Forward pass
cell = tf.contrib.rnn.BasicLSTMCell(state_size, state_is_tuple=True)
states_series, current_state = tf.contrib.rnn.static_rnn(cell, inputs_series, init_state)

losses = [tf.nn.sparse_softmax_cross_entropy_with_logits(logits, labels) for logits, labels in zip(logits_series,labels_series)]
total_loss = tf.reduce_mean(losses)

Answer 1

是的，问题出在 inputs_series 上。根据错误，它是一个形状为 () 的张量，即只是一个数字。

来自 tf.nn.static_rnn文档:

inputs: A length T list of inputs, each a Tensor of shape [batch_size, input_size], or a nested tuple of such elements.

在大多数情况下，您希望inputs为[seq_length, None, input_size]，其中:

• seq_length 是序列长度，或 LSTM 单元的数量。
• None 代表批量大小(任意)。
• input_size 是每个单元格的特征数。

因此请确保您的占位符(以及由此转换而来的 inputs_series)具有适当的形状。示例:

X = tf.placeholder(dtype=tf.float32, shape=[None, n_steps, n_inputs])
X_seqs = tf.unstack(tf.transpose(X, perm=[1, 0, 2]))
basic_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=n_neurons)
output_seqs, states = tf.nn.static_rnn(basic_cell, X_seqs, dtype=tf.float32)

更新:

这是 split 张量的错误方式:

# WRONG!
inputs_series = tf.split(1, truncated_backprop_length, batchX_placeholder)

你应该这样做(注意参数的顺序):

inputs_series = tf.split(batchX_placeholder, truncated_backprop_length, axis=1)