在PyTorch的nn.BatchNorm2d中,affine参数决定是否在批归一化(Batch Normalization)过程中引入可学习的缩放和平移参数。
BN层的公式如下,
affine参数决定是否在批归一化之后应用一个可学习的线性变换,即缩放和平移。具体来说,如果 affine=True,批归一化层会有两个额外的可学习参数:缩放参数γ(初始值为1)和平移参数β(初始值为0),归一化后的输出变为,
若affine=False,则γ和β都为None。相关的源码定义如下,
self.affine = affine
if self.affine:self.weight = Parameter(torch.empty(num_features, **factory_kwargs))self.bias = Parameter(torch.empty(num_features, **factory_kwargs))
else:self.register_parameter("weight", None)self.register_parameter("bias", None)if self.affine:init.ones_(self.weight)init.zeros_(self.bias)
以下是简单的代码示例,
import torch
import torch.nn as nnbn_affine = nn.BatchNorm2d(num_features=3, affine=True)
print("bn_affine γ:", bn_affine.weight)
print("bn_affine β:", bn_affine.bias)
"""
bn_affine.weight
Parameter containing:
tensor([1., 1., 1.], requires_grad=True)bn_affine.bias
Parameter containing:
tensor([0., 0., 0.], requires_grad=True)
"""bn_no_affine = nn.BatchNorm2d(num_features=3, affine=False)
print("bn_no_affine γ:", bn_affine.weight)
print("bn_no_affine β:", bn_affine.bias)
"""
None
None
"""