Pytorch如何约束和限制权重/偏执的范围

方法一:

首先编写模型结构:

class Model(nn.Module):
    def __init__(self):
        super(Model,self).__init__()
        self.l1=nn.Linear(100,50)
        self.l2=nn.Linear(50,10)
        self.l3=nn.Linear(10,1)
        self.sig=nn.Sigmoid()
    
    def forward(self,x):
        x=self.l1(x)
        x=self.l2(x)
        x=self.l3(x)
        x=self.sig(x)
        return(x

然后编写限制权重范围的类:

class weightConstraint(object):
    def __init__(self):
        pass
    
    def __call__(self,module):
        if hasattr(module,'weight'):
            print("Entered")
            w=module.weight.data
            w=w.clamp(0.5,0.7) #将参数范围限制到0.5-0.7之间
            module.weight.data=w

最后实例化这个类,对权重进行限制:

# Applying the constraints to only the last layer
constraints=weightConstraint()
model=Model()

#for i in .....模型训练代码这里请自己补充,
loss = criterion(out, var_y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
model._modules['l3'].apply(constraints)

方法二:

在模型train的时候,对参数的范围进行限制:

loss = criterion(out, var_y)
optimizer.zero_grad()
loss.backward()
optimizer.step()

for p in net.parameters():
    p.data.clamp_(0, 99)

将权重和偏执的范围限制到0-99之间。

仅限制权重的范围,不限制偏执的范围:

for p in net.parameters():
        p.register_hook(lambda grad: torch.clamp(grad, 0,10))

 

posted @ 2022-02-04 22:04  Geeksongs  阅读(2978)  评论(0编辑  收藏  举报

Coded by Geeksongs on Linux

All rights reserved, no one is allowed to pirate or use the document for other purposes.