添加L1/L2损失函数,以及解决报错
- 1.添加L1 loss
- 2.添加L2 loss
- 3.代码报错:AttributeError: 'NoneType' object has no attribute 'data'
1.添加L1 loss
# 方式1:添加到损失函数中
def l1_regularization(model, l1_alpha):
l1_loss = []
for module in model.modules():
if type(module) is nn.BatchNorm2d:
l1_loss.append(torch.abs(module.weight).sum())
return l1_alpha * sum(l1_loss)
# 方式2:添加到反向传播后的参数梯度中
def l1_regularization(model, l1_alpha):
for module in model.modules():
if type(module) is nn.BatchNorm2d:
module.weight.grad.data.add_(l1_alpha * torch.sign(module.weight.data))
2.添加L2 loss
方式1:添加到损失函数中
def l2_regularization(model, l2_alpha):
l2_loss = []
for module in model.modules():
if type(module) is nn.Conv2d:
l2_loss.append((module.weight ** 2).sum() / 2.0)
return l2_alpha * sum(l2_loss)
# 方式2:添加到反向传播后的参数梯度中
def l2_regularization(model, l2_alpha):
for module in model.modules():
if type(module) is nn.Conv2d:
module.weight.grad.data.add_(l2_alpha * module.weight.data)
3.代码报错:AttributeError: ‘NoneType’ object has no attribute ‘data’
我在for module in model.modules():
中,当循环遍历到module为BatchNorm2d时,获取 module.weight
报错。
解决办法:
检查,在定义nn.BatchNorm2d时,有没有将偏执项bias关闭,将affine设置为True,或者直接将这个参数设置删除(默认为True)。