L1 L2正则化

更小权重矩阵的神经网络导致更简单的模型。所以在损失函数后面添加一个数，让损失函数变小。

损失函数太大了，让损失函数的值变小一点点。

import torch
import torch.nn as nn

class RegL1Loss(nn.Module):
  def __init__(self,lam,params):
    super(RegL1Loss,self).__init__()
    self.lam = lam
    self.params = list(params)
    self.n_param = len(self.params)

  def forward(self):
    ls = 0.0
    for param in self.params:
      # L1正则化
      ls +=torch.mean(torch.abs(param))
      # L2正则化 torch.mean(torch.pow(param))
    return self.lam*ls/self.n_param


class Network(nn.Module):
  def __init__(self):
    super(Network,self).__init__()
    self.model = nn.Sequential(
      nn.Linear(19,128),
      nn.ReLU(),
      nn.Linear(128,256),
      nn.ReLU(),
      nn.Linear(256,3)
    )

  def forward(self,x):
    return self.model(x)
  
if __name__=='__main__':
  net = Network()
  loss_fn = nn.MSELoss()
  reg_l1_loss_fn = RegL1Loss(lam=0.1,params=net.parameters())
  # _loss = loss_fn(None,None) #获取任何损失(基于预测和实际值)
  _ll_loss = reg_l1_loss_fn() #求解惩罚项损失
  print(_ll_loss)
  # _loss = _loss+_ll_loss