TensorFlow中权重的随机初始化的方法

时间:2022-05-16 08:56:07

一开始没看懂stddev是什么参数,找了一下,在tensorflow/python/ops里有random_ops,其中是这么写的:

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
def random_normal(shape, mean=0.0, stddev=1.0, dtype=types.float32,
         seed=None, name=None):
 """Outputs random values from a normal distribution.
 
 Args:
  shape: A 1-D integer Tensor or Python array. The shape of the output tensor.
  mean: A 0-D Tensor or Python value of type `dtype`. The mean of the normal
   distribution.
  stddev: A 0-D Tensor or Python value of type `dtype`. The standard deviation
   of the normal distribution.
  dtype: The type of the output.
  seed: A Python integer. Used to create a random seed for the distribution.
   See
   [`set_random_seed`](../../api_docs/python/constant_op.md#set_random_seed)
   for behavior.
  name: A name for the operation (optional).
 
 Returns:
  A tensor of the specified shape filled with random normal values.
 """

也就是按照正态分布初始化权重,mean是正态分布的平均值,stddev是正态分布的标准差(standard deviation),seed是作为分布的random seed(随机种子,我百度了一下,跟什么伪随机数发生器还有关,就是产生随机数的),在mnist/concolutional中seed赋值为66478,挺有意思,不知道是什么原理。

后面还有truncated_normal的定义:

?
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
def truncated_normal(shape, mean=0.0, stddev=1.0, dtype=types.float32,
           seed=None, name=None):
 """Outputs random values from a truncated normal distribution.
 
 The generated values follow a normal distribution with specified mean and
 standard deviation, except that values whose magnitude is more than 2 standard
 deviations from the mean are dropped and re-picked.
 
 Args:
  shape: A 1-D integer Tensor or Python array. The shape of the output tensor.
  mean: A 0-D Tensor or Python value of type `dtype`. The mean of the
   truncated normal distribution.
  stddev: A 0-D Tensor or Python value of type `dtype`. The standard deviation
   of the truncated normal distribution.
  dtype: The type of the output.
  seed: A Python integer. Used to create a random seed for the distribution.
   See
   [`set_random_seed`](../../api_docs/python/constant_op.md#set_random_seed)
   for behavior.
  name: A name for the operation (optional).
 
 Returns:
  A tensor of the specified shape filled with random truncated normal values.
 """

截断正态分布,以前都没听说过。

TensorFlow还提供了平均分布等。

以上就是本文的全部内容,希望对大家的学习有所帮助,也希望大家多多支持服务器之家。

原文链接:https://www.cnblogs.com/dcsds1/p/5203764.html