torch中有多种余弦退火学习率调整方法，包括：OneCycleLR、CosineAnnealingLR和CosineAnnealingWarmRestarts。
CosineAnnealingWarmRestarts（带预热的余弦退火）学习率方法定义
torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0, T_mult=1, \
eta_min=0, last_epoch=- 1, verbose=False)CosineAnnealingWarmRestarts参数：
- optimizer (Optimizer) – Wrapped optimizer. 优化器
- T_0 (int) – Number of iterations for the first restart.学习率第一次回到初始值的epoch位置
- T_mult (int, optional) – A factor increases T_{i} mult应该是multiply的意思，即T_mult=2意思是周期翻倍，第一个周期是1，则第二个周期是2，第三个周期是4。
- eta_min (float, optional) – Minimum learning rate. Default: 0.
- last_epoch (int, optional) – The index of last epoch. Default: -1.
- verbose (bool) – If True, prints a message to stdout for each update. Default: False.
CosineAnnealingWarmRestarts使用
scheduler 初始化之后，在batch中使用scheduler.step()
import torch
from torch.optim.lr_scheduler import CosineAnnealingLR, CosineAnnealingWarmRestarts,StepLR, OneCycleLR
import torch.nn as nn
from torchvision.models import resnet18
import matplotlib.pyplot as pltmodel = resnet18(pretrained=False)
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)mode = 'cosineAnnWarm'
#mode = 'cosineAnn'
if mode == 'cosineAnn':scheduler = CosineAnnealingLR(optimizer, T_max=5, eta_min=0.001)
elif mode == 'cosineAnnWarm':scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=5, T_mult=2)
plt.figure()
max_epoch = 50
iters = 200
cur_lr_list = []
for epoch in range(max_epoch):for batch in range(iters):optimizer.step()scheduler.step()cur_lr = optimizer.param_groups[-1]['lr']cur_lr_list.append(cur_lr)print('Cur lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
OneCycleLR使用
pytorch版本>=1.7
mode = 'OneCycleLR'
#mode = 'cosineAnn'
max_epoch = 500
iters = 200
if mode == 'cosineAnn':scheduler = CosineAnnealingLR(optimizer, T_max=5, eta_min=0.001)
elif mode == 'cosineAnnWarm':scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=5, T_mult=2)
elif mode == 'OneCycleLR':scheduler = OneCycleLR(optimizer, max_lr=0.5, steps_per_epoch=iters, epochs=max_epoch, pct_start=o.3)
plt.figure()
cur_lr_list = []
for epoch in range(max_epoch):for batch in range(iters):optimizer.step()scheduler.step()cur_lr = optimizer.param_groups[-1]['lr']cur_lr_list.append(cur_lr)#print('Cur lr:', cur_lr)
x_list = list(range(len(cur_lr_list)))
plt.plot(x_list, cur_lr_list)
plt.show()
在目前的pytorch1.9版本中新添加了一个three_phase参数,当这个three_phase=True
scheduler=torch.optim.lr_scheduler.OneCycleLR(optimizer,max_lr=0.9,steps_per_epoch=iters, epochs=max_epoch, pct_start=o.3,three_phase=True)得到的学习率变成下图:对称+陡降
————————————————
修改自：https://blog.csdn.net/qq_30129009/article/details/121732567
