tf.truncated_ Normal() usage

sketch

tf.truncated_normal(shape, mean=0.0, stddev=1.0, 
dtype=tf.float32, seed=None, name=None)

• This function produces a normal distribution ; This is a truncated function that produces a normal distribution , The resulting values follow a normal distribution with a specified mean and standard deviation , let me put it another way , If the difference between the generated value and the mean value is greater than twice the standard deviation, it will be discarded and re selected ; This function is compared with random data generated by general normal distribution , The difference between the random number generated by this function and the mean value will not exceed twice the standard deviation , But other general functions are possible .
• shape： Represents the dimension of the generated tensor
• mean： mean value
• stddev： Standard deviation
• dtype： Type of output
• seed： An integer , After setting , The random number generated each time is the same
• name： The name of the operation

Besides ： In the curve of normal distribution ：
Horizontal axis interval （μ-σ,μ+σ） The area inside is 68.268949%
Horizontal axis interval （μ-2σ,μ+2σ） The area inside is 95.449974%
Horizontal axis interval （μ-3σ,μ+3σ） The area inside is 99.730020%

X Fall in the (μ-3σ,μ+3σ) The probability is less than three thousandths , In practical problems, it is often thought that the corresponding event will not happen , Basically, you can put the interval (μ-3σ,μ+3σ) As a random variable X The actual possible value range , This is called the normal distribution "3σ" principle .
stay tf.truncated_normal If X The value of is in the interval (μ-2σ,μ+2σ) Otherwise, choose again . This ensures that the generated values are near the mean .

example

import tensorflow._api.v2.compat.v1 as tf
tf.disable_v2_behavior()

c = tf.truncated_normal(shape = [3,3],mean = 0 , stddev = 1)
with tf.Session() as sess:
    print(sess.run(c))

Output :

''' [[-0.04338798 0.9103105 -1.0928041 ] [ 0.1483252 -0.11247149 -1.2393879 ] [ 0.5323354 0.946114 -0.66076565]] '''