深度学习基础知识 学习率调度器的用法解析_create_lr_scheduler

1、自定义学习率调度器**：**torch.optim.lr_scheduler.LambdaLR

实验代码：

import torch
import torch.nn as nn

def lr_lambda(x):
    return x*2
    
net=nn.Sequential(nn.Conv2d(3,16,3,1,1))

optimizer=torch.optim.SGD(net.parameters(),lr=0.01,momentum=0.9)

lr_scheduler=torch.optim.lr_scheduler.LambdaLR(optimizer,lr_lambda=lr_lambda)

for _ in range(10):
    optimizer.step()
    lr_scheduler.step()
    print(optimizer.param_groups[0]['lr'])
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16

打印结果：

分析数据变化如下图所示：

2、正儿八经的模型搭建流程以及学习率调度器的使用设置

代码：

import torch
import torch.nn as nn
import numpy as np

def create_lr_scheduler(optimizer,
                        num_step:int,
                        epochs:int,
                        warmup=True,
                        warmup_epochs=1,
                        warmup_factor=1e-3):
    assert num_step>0 and epochs>0
    if warmup is False:
        warmup_epochs=0
    
    def f(x):
        """
            根据step数,返回一个学习率倍率因子，
            注意在训练开始之前,pytorch会提前调用一次create_lr_scheduler.step()方法

        
        """

        if warmup is True and x <= (warmup_epochs * num_step):
            alpha=float(x) / (warmup_epochs * num_step)
            # warmup过程中，学习率因子(learning rate factor)：warmup_factor -----> 1
            return warmup_factor * (1-alpha) + alpha
        else:
            # warmup后，学习率因子(learning rate factor)：warmup_factor -----> 0
            return (1-(x - warmup_epochs * num_step) / (epochs-warmup_epochs * num_step)) ** 0.9
        
    return torch.optim.lr_scheduler.LambdaLR(optimizer,lr_lambda=f)


net=nn.Sequential(nn.Conv2d(3,16,1,1))
optimizer=torch.optim.SGD(net.parameters(),lr=0.01,momentum=0.9)

lr_scheduler=create_lr_scheduler(optimizer=optimizer,num_step=5,epochs=20,warmup=True)

image=(np.random.rand(1,3,64,64)).astype(np.float32)
image_tensor=torch.tensor(image.copy(),dtype=torch.float32)
print(image.dtype)

for epoch in range(20):
    net.train()

    predict=net(image_tensor)
    optimizer.zero_grad()
    optimizer.step()
    lr_scheduler.step()
    print(optimizer.param_groups[0]['lr'])   # 打印学习率变化情况

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51