TensorFlow.data 批处理方法不批处理数据集中的文本行

Question

我正在遵循斯坦福大学张量流链接的张量流教程，我遇到了一堵砖墙。

我正在从具有以下行的文本文件中读取数据：

I use Tensorflow
You use PyTorch
Tensorflow is better
By a lot

使用一次性迭代器时，批处理方法工作正常

# Reading the file with tf.data
import tensorflow as tf
dataset = tf.data.TextLineDataset("file.txt")
iterator = dataset.make_one_shot_iterator() # iter can loop through data once
next_element = iterator.get_next() 

#---TRANSFORMING DATA---
# create batches 
batch_size = 2
dataset = dataset.batch(batch_size) 
# prefetching data (transfer from main storage to temp for faster loader)
dataset = dataset.prefetch(1)
iterator = dataset.make_one_shot_iterator()
next_element = iterator.get_next()
with tf.Session() as sess:
    for _ in range(4//batch_size):
        print(sess.run(next_element))

这将按预期返回（请注意，注释不是输出的一部分）：

[b'I use Tensorflow' b'You use PyTorch'] # first batch
[b'Tensorflow is better' b'By a lot'] # second batch

但是，当我使用可初始化的迭代器进行类似的练习时......

# Using initializable iterators
import tensorflow as tf
dataset = tf.data.TextLineDataset("file.txt")
iterator = dataset.make_initializable_iterator()
next_element = iterator.get_next()
init_op = iterator.initializer # this allows you to reset iterator --> you can iterate multiple times (epochs)
epochs = 2
batch_size = 2
num_examples = 4
# This code block doesn't seem to work

dataset = dataset.batch(batch_size) # doesn't batch up lines

dataset = dataset.prefetch(1)
with tf.Session() as sess:
    # Initialize the iterator
    for i in range(epochs):
        sess.run(init_op)
        for _ in range(num_examples//batch_size): # loops through all batches
            print(sess.run(next_element))
        print("n")

。我得到这个结果：

# first epoch
b'I use Tensorflow' #??
b'You use PyTorch'
# second epoch
b'I use Tensorflow'
b'You use PyTorch'

我所期望的：

# First epoch
[b'I use Tensorflow' b'You use PyTorch'] # first batch
[b'Tensorflow is better' b'By a lot'] # second batch
# Second epoch
[b'I use Tensorflow' b'You use PyTorch'] # first batch
[b'Tensorflow is better' b'By a lot'] # second batch

有人可以帮助我弄清楚我做错了什么吗？ 我已经检查了 batch（） 的文档，一切看起来都已检查出来。

谢谢。

Answer 1

在创建迭代器之前移动dataset.batch(batch_size)行。

Answer 2

也许，因为你在每个时代都叫sess.run(init_op)？在循环之前调用它一次。