新人上手TensorFlow 之 Normalization

程序员文章站 2022-07-16 18:45:53

...

上一篇转载自张俊林老师的博客，参考《batch normalization: accelerating deep network training by reducing internal》这篇论文，基本讲了一下，批处理归一化对于神经网络的意义所在及基本的原理和步骤。算是理论上的理解吧！这篇博客，我们来看一下，在TensorFlow中如何实现Normalization！

TensorFlow中的Normalization函数

在TensorFlow中，封装了多种种归一化的函数，分别是：

编号	TensorFlow函数	数学原理
1	tf.nn.l2_normalize	output = x / sqrt(max(sum(x**2), epsilon))
2	tf.nn.batch_normalization
3	tf.nn.moments	计算 mean & variance

具体见： https://www.tensorflow.org/api_guides/python/nn#Normalization
新人上手TensorFlow 之 Normalization

TensorFlow中Normalization函数示例

'''
Normalization in TensorFlow
'''
import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
import tensorflow as tf
import numpy as np

x = np.arange(10,dtype = np.float16)
print('x-orig:', x)

sess = tf.Session()

#l2-normalization
x_l2_norm = tf.nn.l2_normalize(x, dim = 0)
x_norm_res = sess.run(x_norm)
print('l2-norm:',x_norm_res)

# batch  normalization
## calculate the mean & variance 
x = tf.constant(x, dtype = tf.float32)
[x_mean, x_varia] = tf.nn.moments(x, axes=0)
offset = 0
scale = 0.1
vari_epsl = 0.0001
## calculate the batch normalization
x_bn = tf.nn.batch_normalization(x, x_mean, x_varia, offset,scale,vari_epsl)
print('mean & vari: ', sess.run( [x_mean, x_varia]))
x_bn_res = sess.run(x_bn)
print('x_bn:', x_bn_res)

sess.close()

运行结果：

x-orig: [ 0.  1.  2.  3.  4.  5.  6.  7.  8.  9.]
l2-norm: [ 0.          0.05923489  0.11846977  0.17770466  0.23693955  0.29617444
  0.35540932  0.41464421  0.4738791   0.53311396]
mean & vari:  [4.5, 8.25]
x_bn: [-0.15666895 -0.12185362 -0.0870383  -0.05222298 -0.01740766  0.01740767
  0.05222298  0.08703831  0.12185363  0.15666895]