Adam和AdamW的区别

Adam 与 Adamw的区别

一句话版本

Adamw 即 Adam + weight decate ,效果与 Adam + L2正则化相同,但是计算效率更高,因为L2正则化需要在loss中加入正则项,之后再算梯度,最后在反向传播,而Adamw直接将正则项的梯度加入反向传播的公式中,省去了手动在loss中加正则项这一步

实验

Adamw算法:

图片参考:https://www.cnblogs.com/tfknight/p/13425532.html

ps:以现在的源码为标准,图中的代码12行中:
λ = w d ∗ l r \lambda = wd * lr λ=wd∗lr
wd=weight decay, lr=learning rate

实验代码:

my_opt是手动实现的代码,主要验证梯度下降是否正确

import torch
import torch.nn as nn
import torch.optim.lr_scheduler
from torch.optim import AdamW, Adam


class M(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(3, 1, bias=False)

    def forward(self, x):
        return self.fc(x)


class my_opt():
    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8,
                 weight_decay=1e-2):
        self.params = params
        self.lr = lr
        self.b1 = betas[0]
        self.b2 = betas[1]
        self.eps = eps
        self.wd = weight_decay
        self.m = 0
        self.v = 0
        self.b1t = 1.0
        self.b2t = 1.0

    # 模拟AdamW的step
    def stepW(self):
        for name, param in self.params.named_parameters():
            if param.grad is None:
                continue
            g = param.grad
            self.m = self.b1 * self.m + (1 - self.b1) * g
            self.v = self.b2 * self.v + (1 - self.b2) * g * g
            self.b1t *= self.b1
            self.b2t *= self.b2
            m = self.m / (1 - self.b1t)
            v = self.v / (1 - self.b2t)
            n = 1.0
            param.data -= n * (self.lr * (m / (v.sqrt() + self.eps) + self.wd * param.data))

    # 模拟Adam的step
    def step(self):
        for name, param in self.params.named_parameters():
            if param.grad is None:
                continue
            g = param.grad
            self.m = self.b1 * self.m + (1 - self.b1) * g
            self.v = self.b2 * self.v + (1 - self.b2) * g * g
            self.b1t *= self.b1
            self.b2t *= self.b2
            m = self.m / (1 - self.b1t)
            v = self.v / (1 - self.b2t)
            n = 1.0
            param.data -= n * (self.lr * m / (v.sqrt() + self.eps))


adam_model = M()
adamw_model = M()
my_adam_model = M()
my_adamw_model = M()
# 使4个模型参数相同
adamw_model.load_state_dict(adam_model.state_dict())
my_adam_model.load_state_dict(adam_model.state_dict())
my_adamw_model.load_state_dict(adam_model.state_dict())

model_ls = {'adam_model': adam_model,
            'adamw_model': adamw_model,
            'my_adam_model': my_adam_model,
            'my_adamw_model': my_adamw_model}

# 检查4个模型初始参数
for m in model_ls:
    print(f"Model : {m}")
    model = model_ls[m]
    for name, parma in model.named_parameters():
        print(name)
        print(parma)

adam_opt = Adam(adam_model.parameters(), lr=0.1)
adamw_opt = AdamW(adamw_model.parameters(), lr=0.1)
my_adam_opt = my_opt(my_adam_model, lr=0.1)
my_adamw_opt = my_opt(my_adamw_model, lr=0.1)

opt_ls = {'adam_model': adam_opt,
          'adamw_model': adamw_opt,
          'my_adam_model': my_adam_opt,
          'my_adamw_model': my_adamw_opt}

for i in range(5):
    print(">>>>>>>>>>>>>>>>>>>>>>> epoch", i)
    ip = torch.rand(2, 3)
    for m in model_ls:
        print(f">>> Model : {m}")
        model = model_ls[m]
        opt = opt_ls[m]
        loss = (model(ip).sum()) ** 2
        loss.backward()

        if m != 'my_adamw_model':
            opt.step()
        else:
            opt.stepW()

        for name, parma in model.named_parameters():
            print(name)
            print(parma)

        model.zero_grad()

代码是照着图片公式写的,和源码还是有点区别的,但是可以验证每一步的梯度下降都对了