如何使用tensorflow的Dataset API Iterator作为(递归)神经网络的输入？

Question

当使用tensorflow的Dataset API Iterator时，我的目标是定义一个对迭代器的get_next()张量操作的RNN作为它的输入(参见代码中的(1) )。

但是，简单地用dynamic_rnn作为输入定义get_next()会导致一个错误：ValueError: Initializer for variable rnn/basic_lstm_cell/kernel/ is from inside a control-flow construct, such as a loop or conditional. When creating a variable inside a loop or conditional, use a lambda as the initializer.

现在，我知道一种解决方法是为next_batch创建一个占位符，然后为张量创建一个占位符(因为您不能传递张量本身)，然后使用feed_dict传递它(参见代码中的X和(2) )。然而，如果我正确理解它，这不是一个有效的解决方案，因为我们首先评估，然后重新初始化张量。

有没有办法：

1. 直接在Iterator输出的顶部定义dynamic_rnn；

或者：

1. 以某种方式将现有的get_next()张量直接传递给占位符，即dynamic_rnn的输入。

完整的示例；(1)版本是我想要工作的，但它不工作，而(2)是起作用的解决方案。

import tensorflow as tf

from tensorflow.contrib.rnn import BasicLSTMCell
from tensorflow.python.data import Iterator

data = [ [[1], [2], [3]], [[4], [5], [6]], [[1], [2], [3]] ]
dataset = tf.data.Dataset.from_tensor_slices(data)
dataset = dataset.batch(2)
iterator = Iterator.from_structure(dataset.output_types,
                                   dataset.output_shapes)
next_batch = iterator.get_next()
iterator_init = iterator.make_initializer(dataset)

# (2):
X = tf.placeholder(tf.float32, shape=(None, 3, 1))

cell = BasicLSTMCell(num_units=8)

# (1):
# outputs, states = lstm_outputs, lstm_states = tf.nn.dynamic_rnn(cell, next_batch, dtype=tf.float32)

# (2):
outputs, states = lstm_outputs, lstm_states = tf.nn.dynamic_rnn(cell, X, dtype=tf.float32)

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    sess.run(iterator_init)

    # (1):
    # o, s = sess.run([outputs, states])
    # o, s = sess.run([outputs, states])

    # (2):
    o, s = sess.run([outputs, states], feed_dict={X: next_batch.eval()})
    o, s = sess.run([outputs, states], feed_dict={X: next_batch.eval()})

(使用tensorflow 1.4.0，Python3.6)

(非常感谢:)

Answer 1

原来这个神秘的错误很可能是tensorflow中的一个错误，参见https://github.com/tensorflow/tensorflow/issues/14729。更具体地说，错误实际上来自于输入错误的数据类型(在我前面的示例中，data数组包含int32值，但它应该包含浮点数)。

而不是获取ValueError: Initializer for variable rnn/basic_lstm_cell/kernel/ is from inside a control-flow construct错误，

tensorflow应返回：

TypeError: Tensors in list passed to 'values' of 'ConcatV2' Op have types [int32, float32] that don't all match. (见1)。

要解决这个问题，只需更改

data = [ [[1], [2], [3]], [[4], [5], [6]], [[1], [2], [3]] ]

至

data = np.array([[ [1], [2], [3]], [[4], [5], [6]], [[1], [2], [3]] ], dtype=np.float32)

然后，下列代码应正常工作：

import tensorflow as tf
import numpy as np

from tensorflow.contrib.rnn import BasicLSTMCell
from tensorflow.python.data import Iterator

data = np.array([[ [1], [2], [3]], [[4], [5], [6]], [[1], [2], [3]] ], dtype=np.float32)
dataset = tf.data.Dataset.from_tensor_slices(data)
dataset = dataset.batch(2)
iterator = Iterator.from_structure(dataset.output_types,
                                   dataset.output_shapes)
next_batch = iterator.get_next()
iterator_init = iterator.make_initializer(dataset)

# (2):
# X = tf.placeholder(tf.float32, shape=(None, 3, 1))

cell = BasicLSTMCell(num_units=8)

# (1):
outputs, states = lstm_outputs, lstm_states = tf.nn.dynamic_rnn(cell, next_batch, dtype=tf.float32)

# (2):
# outputs, states = lstm_outputs, lstm_states = tf.nn.dynamic_rnn(cell, X, dtype=tf.float32)

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    sess.run(iterator_init)

    # (1):
    o, s = sess.run([outputs, states])
    o, s = sess.run([outputs, states])

    # (2):
    # o, s = sess.run([outputs, states], feed_dict={X: next_batch.eval()})
    # o, s = sess.run([outputs, states], feed_dict={X: next_batch.eval()})