https://www.jb51.net/article/178976.htm

直接看代码例子，有详细注释！！

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
import tensorflow as tf
import numpy as np
 
 
d = np.arange(0,60).reshape([6, 10])
 
# 将array转化为tensor
data = tf.data.Dataset.from_tensor_slices(d)
 
# 从data数据集中按顺序抽取buffer_size个样本放在buffer中，然后打乱buffer中的样本
# buffer中样本个数不足buffer_size，继续从data数据集中安顺序填充至buffer_size，
# 此时会再次打乱
data = data.shuffle(buffer_size=3)
 
# 每次从buffer中抽取4个样本
data = data.batch(4)
 
# 将data数据集重复，其实就是2个epoch数据集
data = data.repeat(2)
 
# 构造获取数据的迭代器
iters = data.make_one_shot_iterator()
 
# 每次从迭代器中获取一批数据
batch = iters.get_next()
 
sess = tf.Session()
 
sess.run(batch)
# 数据集完成遍历完之后，继续抽取的话会报错：OutOfRangeError
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
In [21]: d
Out[21]:
array([[ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
  [10, 11, 12, 13, 14, 15, 16, 17, 18, 19],
  [20, 21, 22, 23, 24, 25, 26, 27, 28, 29],
  [30, 31, 32, 33, 34, 35, 36, 37, 38, 39],
  [40, 41, 42, 43, 44, 45, 46, 47, 48, 49],
  [50, 51, 52, 53, 54, 55, 56, 57, 58, 59]])
In [22]: sess.run(batch)
Out[22]:
array([[ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
  [30, 31, 32, 33, 34, 35, 36, 37, 38, 39],
  [20, 21, 22, 23, 24, 25, 26, 27, 28, 29],
  [10, 11, 12, 13, 14, 15, 16, 17, 18, 19]])
 
In [23]: sess.run(batch)
Out[23]:
array([[40, 41, 42, 43, 44, 45, 46, 47, 48, 49],
  [50, 51, 52, 53, 54, 55, 56, 57, 58, 59]])

从输出结果可以看出：

shuffle是按顺序将数据放入buffer里面的；

当repeat函数在shuffle之后的话，是将一个epoch的数据集抽取完毕，再进行下一个epoch的。

那么，当repeat函数在shuffle之前会怎么样呢？如下：

1
2
3
4
5
data = data.repeat(2)
 
data = data.shuffle(buffer_size=3)
 
data = data.batch(4)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
In [25]: sess.run(batch)
Out[25]:
array([[10, 11, 12, 13, 14, 15, 16, 17, 18, 19],
  [20, 21, 22, 23, 24, 25, 26, 27, 28, 29],
  [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
  [40, 41, 42, 43, 44, 45, 46, 47, 48, 49]])
 
In [26]: sess.run(batch)
Out[26]:
array([[50, 51, 52, 53, 54, 55, 56, 57, 58, 59],
  [ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
  [30, 31, 32, 33, 34, 35, 36, 37, 38, 39],
  [30, 31, 32, 33, 34, 35, 36, 37, 38, 39]])
 
In [27]: sess.run(batch)
Out[27]:
array([[10, 11, 12, 13, 14, 15, 16, 17, 18, 19],
  [50, 51, 52, 53, 54, 55, 56, 57, 58, 59],
  [20, 21, 22, 23, 24, 25, 26, 27, 28, 29],
  [40, 41, 42, 43, 44, 45, 46, 47, 48, 49]])

可以看出，其实它就是先将数据集复制一遍，然后把两个epoch当成同一个新的数据集，一直shuffle和batch下去。