目录

1. introduction

2.操作走一走

1. introduction

• 在实际应用中&＃xff0c;采用自适应的模式&＃xff0c;&＃xff0c;即在批量归一化前面加一个参数γ。即&＃xff1a;
• BN(X)&＃61;γ * (x*u)/σ &＃43; β
• 在实际应用中&＃xff0c;批量归一化收敛速度非常快&＃xff0c;具有很强的泛化能力&＃xff0c;某些情况下可以完全替代正则化与dropout。
• 依然按数据维度BatchNorm1d&＃xff0c;BatchNorm2d&＃xff0c;BatchNorm3d

2.操作走一走

依然采用 &＃61;7&＃61; 中的双半圆数据&＃xff0c;对构建模型时&＃xff08;2.3节&＃xff09;添加self.BN&＃61;nn,BatchModrm1d(hiddendim)

class Logic_Dropout_Net(LogicNet):def __init__(self, inputdim, hiddendim, outputdim):super(Logic_Dropout_Net,self).__init__(inputdim, hiddendim, outputdim)self.BN &＃61;nn.BatchNorm1d(hiddendim) #定义批量归一化&＃xff08;BN&＃xff09;层def forward(self,x):x &＃61; self.Linear1(x)x &＃61; torch.tanh(x)x &＃61; nn.functional.dropout(x, p&＃61;0.07, traing&＃61;self.self.training)x &＃61; self.Linear2(x)return x
model &＃61; Logic_Dropout_Net(inputdim&＃61;2, hiddendim&＃61;500, outputdim&＃61;2)