对基于深度学习的模型调参是炼丹师的必备技能，而学习率的衰减策略正是调参方法之一，今天就系统地学习一下pytorch自带学习率衰减策略

1）实例化模型和优化器

import torch
from torch.optim.lr_scheduler import *
import torch.nn as nn
from torchvision.models import resnet50
import matplotlib.pyplot as plt


model = resnet50(False)
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)

2）选择不同的学习率衰减策略

1. LambdaLR

该策略可以将学习率自定义为一个与epoch有关的lambda函数，如：lr_lambda=lambda epoch: 0.95 ** epoch

scheduler1 = LambdaLR(optimizer=optimizer, lr_lambda=lambda epoch: 0.95 ** epoch)

plt.figure()
max_epoch = 30
cur_lr_list = []
for epoch in range(max_epoch):
    optimizer.step()
    scheduler1.step()
    cur_lr = optimizer.param_groups[-1]['lr']
    cur_lr_list.append(cur_lr)
    print('Current lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
plt.savefig('D:/lr_scheduler_learning/lambdalr.png')

2. StepLR

每过step size轮epoch，将当前学习率乘以gamma

scheduler2 = StepLR(optimizer, step_size=10, gamma=0.1)

plt.figure()
max_epoch = 30
cur_lr_list = []
for epoch in range(max_epoch):
    optimizer.step()
    scheduler2.step()
    cur_lr = optimizer.param_groups[-1]['lr']
    cur_lr_list.append(cur_lr)
    print('Current lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
plt.savefig('D:/lr_scheduler_learning/steplr.png')

3. MultiStepLR

当epoch为milestones里的元素时，将当前学习率乘以gamma，如：milestones=[5,10,15,20,25]，当epoch为5、10、15、20、25时，new_lr = cur_lr * gamma

scheduler3 = MultiStepLR(optimizer, milestones=[5,10,15,20,25], gamma=0.1)

plt.figure()
max_epoch = 30
cur_lr_list = []
for epoch in range(max_epoch):
    optimizer.step()
    scheduler3.step()
    cur_lr = optimizer.param_groups[-1]['lr']
    cur_lr_list.append(cur_lr)
    print('Current lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
plt.savefig('D:/lr_scheduler_learning/multisteplr.png')

4. ExponentialLR

指数型学习率衰减策略，new_lr = cur_lr * gamma ** cur_epoch

scheduler4 = ExponentialLR(optimizer, gamma=0.8) 

plt.figure()
max_epoch = 30
cur_lr_list = []
for epoch in range(max_epoch):
    optimizer.step()
    scheduler4.step()
    cur_lr = optimizer.param_groups[-1]['lr']
    cur_lr_list.append(cur_lr)
    print('Current lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
plt.savefig('D:/lr_scheduler_learning/exponentiallr.png')

5. CosineAnnealingLR

余弦退火学习率，T_max：余弦函数周期的一半，即每T_max*2个epoch循环一次余弦学习率；eta_min：最小学习率。公式如下：

scheduler5 = CosineAnnealingLR(optimizer,T_max=5,eta_min=0.05)

plt.figure()
max_epoch = 30
cur_lr_list = []
for epoch in range(max_epoch):
    optimizer.step()
    scheduler5.step()
    cur_lr = optimizer.param_groups[-1]['lr']
    cur_lr_list.append(cur_lr)
    print('Current lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
plt.savefig('D:/lr_scheduler_learning/cosineannealinglr.png')

6. CyclicLR

base_lr：lr的下限；max_lr：lr的上限；step_size_up：lr从下限到上限需要的epoch数；step_size_down：lr从上限到下限需要的epoch数

scheduler6 = CyclicLR(optimizer, base_lr=0.01, max_lr=0.2, step_size_up=10, step_size_down=5)

plt.figure()
max_epoch = 30
cur_lr_list = []
for epoch in range(max_epoch):
    optimizer.step()
    scheduler6.step()
    cur_lr = optimizer.param_groups[-1]['lr']
    cur_lr_list.append(cur_lr)
    print('Current lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
plt.savefig('D:/lr_scheduler_learning/cycliclr.png')

7. CosineAnnealingWarmRestarts

T_0：lr第一次 (周期) 下降到eta_min所需要的epoch数；T_mult：之后lr衰减需要的epoch数变成了前一周期需要的epoch数乘以T_mult

scheduler7 = CosineAnnealingWarmRestarts(optimizer, T_0=5, T_mult=2, eta_min=0.01)

plt.figure()
max_epoch = 30
cur_lr_list = []
for epoch in range(max_epoch):
    optimizer.step()
    scheduler7.step()
    cur_lr = optimizer.param_groups[-1]['lr']
    cur_lr_list.append(cur_lr)
    print('Current lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
plt.savefig('D:/lr_scheduler_learning/cosineannealingwarmrestarts.png')