Batch Normalization

太过爱你忘了你带给我的痛 2023-07-02 13:28 9阅读 0赞

# ![20191009191333910.png][] #

# [日萌社][Link 1] #

[人工智能AI：Keras PyTorch MXNet TensorFlow PaddlePaddle 深度学习实战（不定时更新）][AI_Keras PyTorch MXNet TensorFlow PaddlePaddle]

--------------------

[tensorflow 2.0 画出model网络模型的拓扑图][tensorflow 2.0 _model]

[tensorflow 2.0 的回调函数callbacks（TensorBoard、ModelCheckpoint）][tensorflow 2.0 _callbacks_TensorBoard_ModelCheckpoint]

[TensorBoard视觉化网络每层权重分布、视觉化网络层结构][TensorBoard]

[MSE(均方误差)、MAE(平均绝对误差)][MSE_MAE]

[from\_logits][from_logits]

[CE(Cross-Entropy)、BCE(Binary Cross-Entropy 即Sigmoid+CE)、CCE(Categorical Cross-Entropy 即Softmax+CE)][CE_Cross-Entropy_BCE_Binary Cross-Entropy _Sigmoid_CE_CCE_Categorical Cross-Entropy _Softmax_CE]

[对连续值/离散值进行预处理的两种方式（标准化/one-hot化）、反标准化/逆标准化、字符串预处理][one-hot]

[损失函数loss、指标函数metrics][loss_metrics]

[激活函数 Sigmoid、Tanh、ReLU、Softmax][Sigmoid_Tanh_ReLU_Softmax]

[Batch Normalization][]

[反向传播、梯度下降、学习率、优化器optimizers（GD、SGD、Momentum、AdaGrad、Adam）][optimizers_GD_SGD_Momentum_AdaGrad_Adam]

[权重初始化对于激活函数的选择：随机常态分布/随机正态分布初始化(标准差为0.01或1)、Xavier(Glorot)初始化、He初始化][0.01_1_Xavier_Glorot_He]

[图像增强(IA)、数据预处理][IA]

[混淆矩阵tf.math.confusion\_matrix][tf.math.confusion_matrix]

[使用预训练网络训练的两种方式：Keras Applications、TensorFlow Hub][Keras Applications_TensorFlow Hub]

--------------------

1.Batch Normalization 作用
    	强制让每一层的输出分布变得均匀，用于对每一个batch批次数据进行计算mean平均值和std标准差，并最后累加起来，
    	作为对网络每一层的输出都做标准化，还有额外对输出进行缩放(Scale)、偏移(Shift)的参数。
    	Batch Normalization使得可以同时使用较大的学习率进行训练，不仅可以时的网络收敛速度加快，更一定程度上减少梯度消失/梯度爆炸的出现，
    	并可实现减少使用Dropout/正则化Regularization(L1/L2)的同时也同样达到避免过拟合的问题，
    	并且Batch Normalization也不需要依赖权重初始化方法。
    
    2.Batch Normalization 使用
    	layers.BatchNormalization()：
    		一般配置在卷积层layers.Conv2D和激活函数之间，用于作为对卷积运算输出进行标准化之后使得输出分布变得均匀，
    		然后再输入到激活函数中，使得激活函数的输出分布也同样变得比较均匀，从而达到降低激活函数中出现梯度消失/梯度爆炸的问题。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0_size_16_color_FFFFFF_t_70][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0_size_16_color_FFFFFF_t_70 1][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0_size_16_color_FFFFFF_t_70 2][]![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0_size_16_color_FFFFFF_t_70 3][]

[20191009191333910.png]: https://img-blog.csdnimg.cn/20191009191333910.png
[Link 1]: http://www.rimengshe.com/
[AI_Keras PyTorch MXNet TensorFlow PaddlePaddle]: https://blog.csdn.net/zimiao552147572/article/details/88867161
[tensorflow 2.0 _model]: https://azusa.blog.csdn.net/article/details/104092884
[tensorflow 2.0 _callbacks_TensorBoard_ModelCheckpoint]: https://azusa.blog.csdn.net/article/details/104100362
[TensorBoard]: https://azusa.blog.csdn.net/article/details/104132982
[MSE_MAE]: https://azusa.blog.csdn.net/article/details/104094483
[from_logits]: https://azusa.blog.csdn.net/article/details/104095436
[CE_Cross-Entropy_BCE_Binary Cross-Entropy _Sigmoid_CE_CCE_Categorical Cross-Entropy _Softmax_CE]: https://azusa.blog.csdn.net/article/details/104095759
[one-hot]: https://azusa.blog.csdn.net/article/details/104099669
[loss_metrics]: https://azusa.blog.csdn.net/article/details/104099800
[Sigmoid_Tanh_ReLU_Softmax]: https://azusa.blog.csdn.net/article/details/104108314
[Batch Normalization]: https://azusa.blog.csdn.net/article/details/104115714
[optimizers_GD_SGD_Momentum_AdaGrad_Adam]: https://azusa.blog.csdn.net/article/details/104106429
[0.01_1_Xavier_Glorot_He]: https://azusa.blog.csdn.net/article/details/104110857
[IA]: https://azusa.blog.csdn.net/article/details/104114416
[tf.math.confusion_matrix]: https://azusa.blog.csdn.net/article/details/104125078
[Keras Applications_TensorFlow Hub]: https://azusa.blog.csdn.net/article/details/104134536
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20200130155626275.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/20200130155623474.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20200130155701332.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20200130155703902.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9henVzYS5ibG9nLmNzZG4ubmV0,size_16,color_FFFFFF,t_70