work report
https://blog.csdn.net/qq_42278791/article/details/94398159
code to live a link, a website of 19 years of target detection papers
Paper reading:
1.《Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift》
Mainly make up for evil →. →, I have learned too little before, BN has always had a little understanding, take a little time to learn specifically.
BN is mainly used to accelerate network convergence and prevent common problems of neural networks such as gradient disappearance.
In simple terms, transformation and reconstruction introduces learnable parameters γ and β, so that the network can learn to restore the feature distribution that the original network needs to learn, thereby reducing or avoiding the impact of normalization processing:
where the parameters γ and β are
The formula for the forward conduction process is: do
n’t say much about the formula, just look at the source code:
m = K.mean(X, axis=-1, keepdims=True)#计算均值
std = K.std(X, axis=-1, keepdims=True)#计算标准差
X_normed = (X - m) / (std + self.epsilon)#归一化
out = self.gamma * X_normed + self.beta#重构变换y=γ*x+β
The source code is only 4 lines of code, which is relatively simple.
Reference blog: https://blog.csdn.net/shuzfan/article/details/50723877