如何使用 tensorflow 的 Dataset API Iterator 作为(循环)神经网络的输入? [英] How to use tensorflow's Dataset API Iterator as an input of a (recurrent) neural network?

问题描述

当使用 tensorflow 的 Dataset API Iterator 时，我的目标是定义一个 RNN，该 RNN 将迭代器的 get_next() 张量作为其输入(参见 (1) in代码).

When using the tensorflow's Dataset API Iterator, my goal is to define an RNN that operates on the iterator's get_next() tensors as its input (see (1) in the code).

然而，简单地将 dynamic_rnn 定义为 get_next() 作为其输入会导致错误:ValueError: Initializer for variable rnn/basic_lstm_cell/kernel/是从控制流结构内部，例如循环或条件.在循环或条件中创建变量时，使用 lambda 作为初始值设定项.

However, simply defining the dynamic_rnn with get_next() as its input results in an error: 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简单地为 next_batch 创建一个占位符然后 eval() 张量(因为你不能传递张量本身)和使用 feed_dict 传递它(参见代码中的 X 和 (2)).但是，如果我理解正确，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.

Now I know that one workaround is to simply create a placeholder for next_batch and then eval() the tensor (because you can't pass the tensor itself) and pass it using feed_dict (see X and (2) in the code). However, if I understand it correctly, this is not an efficient solution as we first evaluate and then reinitialize the tensor.

有没有办法:

1. 直接在迭代器的输出之上定义dynamic_rnn；

或:

1. 不知何故直接将现有的 get_next() 张量传递给作为 dynamic_rnn 输入的占位符?

1. Somehow directly pass the existing get_next() tensor to the placeholder that is the input of dynamic_rnn?

完整的工作示例；(1) 版本是我想要的，但它没有，而 (2) 是可行的解决方法.

Full working example; the (1) version is what I would like to work but it doesn't, while (2) is the workaround that does work.

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，Python 3.6.)

(Using tensorflow 1.4.0, Python 3.6.)

非常感谢:)

推荐答案

原来这个神秘的错误很可能是 tensorflow 中的一个 bug，参见 https://github.com/tensorflow/tensorflow/issues/14729.更具体地说，错误实际上是由于提供了错误的数据类型(在我上面的示例中，data 数组包含 int32 值，但它应该包含浮点数).

Turns out the mysterious error is likely a bug in tensorflow, see https://github.com/tensorflow/tensorflow/issues/14729. More specifically, the error really comes from feeding a wrong data type (in my example above, the data array contains int32 values but it should contain floats).

不是得到 ValueError: Initializer for variable rnn/basic_lstm_cell/kernel/is from inside a control-flow structure 错误，
张量流应该返回:
TypeError: 列表中的张量传递给 'ConcatV2' Op 的 'values' 的类型 [int32, float32] 并不完全匹配.(请参阅 1).

Instead of getting the ValueError: Initializer for variable rnn/basic_lstm_cell/kernel/ is from inside a control-flow construct error,
tensorflow should return:
TypeError: Tensors in list passed to 'values' of 'ConcatV2' Op have types [int32, float32] that don't all match. (see 1).

要解决此问题，只需更改
data = [ [[1], [2], [3]], [[4], [5], [6]], [[1], [2], [3]] ]
到
data = np.array([[ [1], [2], [3]], [[4], [5], [6]], [[1], [2], [3]] ], dtype=np.float32)

To fix this problem, simply change
data = [ [[1], [2], [3]], [[4], [5], [6]], [[1], [2], [3]] ]
to
data = np.array([[ [1], [2], [3]], [[4], [5], [6]], [[1], [2], [3]] ], dtype=np.float32)

然后下面的代码应该可以正常工作:

and then the following code shall work properly:

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()})

这篇关于如何使用 tensorflow 的 Dataset API Iterator 作为(循环)神经网络的输入?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！