脚本之家,脚本语言编程技术及教程分享平台!
分类导航

Python|VBS|Ruby|Lua|perl|VBA|Golang|PowerShell|Erlang|autoit|Dos|bat|

服务器之家 - 脚本之家 - Python - 使用tensorflow DataSet实现高效加载变长文本输入

使用tensorflow DataSet实现高效加载变长文本输入

2020-04-08 19:09lyg5623 Python

今天小编就为大家分享一篇使用tensorflow DataSet实现高效加载变长文本输入,具有很好的参考价值,希望对大家有所帮助。一起跟随小编过来看看吧

DataSet是tensorflow 1.3版本推出的一个high-level的api,在1.3版本还只是处于测试阶段,1.4版本已经正式推出。

在网上搜了一遍,发现关于使用DataSet加载文本的资料比较少,官方举的例子只是csv格式的,要求csv文件中所有样本必须具有相同的维度,也就是padding必须在写入csv文件之前做掉,这会增加文件的大小。

经过一番折腾试验,这里给出一个DataSet+TFRecords加载变长样本的范例。

首先先把变长的数据写入到TFRecords文件:

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
def writedata():
 xlist = [[1,2,3],[4,5,6,8]]
 ylist = [1,2]
 #这里的数据只是举个例子来说明样本的文本长度不一样,第一个样本3个词标签1,第二个样本4个词标签2
 writer = tf.python_io.TFRecordWriter("train.tfrecords")
 for i in range(2):
  x = xlist[i]
  y = ylist[i]
  example = tf.train.Example(features=tf.train.Features(feature={
   "y": tf.train.Feature(int64_list=tf.train.Int64List(value=[y])),
   'x': tf.train.Feature(int64_list=tf.train.Int64List(value=x))
  }))
  writer.write(example.SerializeToString())
 writer.close()

然后用DataSet加载:

?
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
feature_names = ['x']
 
def my_input_fn(file_path, perform_shuffle=False, repeat_count=1):
 def parse(example_proto):
  features = {"x": tf.VarLenFeature(tf.int64),
    "y": tf.FixedLenFeature([1], tf.int64)}
  parsed_features = tf.parse_single_example(example_proto, features)
  x = tf.sparse_tensor_to_dense(parsed_features["x"])
  x = tf.cast(x, tf.int32)
  x = dict(zip(feature_names, [x]))
  y = tf.cast(parsed_features["y"], tf.int32)
  return x, y
 
 dataset = (tf.contrib.data.TFRecordDataset(file_path)
    .map(parse))
 if perform_shuffle:
  dataset = dataset.shuffle(buffer_size=256)
 dataset = dataset.repeat(repeat_count)
 dataset = dataset.padded_batch(2, padded_shapes=({'x':[6]},[1])) #batch size为2,并且x按maxlen=6来做padding
 iterator = dataset.make_one_shot_iterator()
 batch_features, batch_labels = iterator.get_next()
 return batch_features, batch_labels
 
next_batch = my_input_fn('train.tfrecords', True)
init = tf.initialize_all_variables()
with tf.Session() as sess:
 sess.run(init)
 for i in range(1):
  xs, y =sess.run(next_batch)
  print(xs['x'])
  print(y)

注意变长的数据TFRecords解析要用VarLenFeature,然后用sparse_tensor_to_dense转换。

以上这篇使用tensorflow DataSet实现高效加载变长文本输入就是小编分享给大家的全部内容了,希望能给大家一个参考,也希望大家多多支持服务器之家。

原文链接:https://blog.csdn.net/lyg5623/article/details/78456915

延伸 · 阅读

精彩推荐