optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=30, gamma=0.1)
 
def train(...):
    for i, data in enumerate(train_loader):
        ......
        y_ = model(x)
        loss = criterion(y_,y)
        loss.backward()
        optimizer.step()
        ......
 
for epoch in range(epochs):
    scheduler.step()
    train(...)
    test(...)

torch.optim.lr_scheduler.StepLR(optimizer, step_size, gamma=0.1, last_epoch=-1)

torch.optim.lr_sheduler.MultiStepLR(optimizer, milestones, gamma=0.1, last_epoch=-1)

torch.optim.lr_sheduler.ExponentialLR(optimizer, gamma, last_epoch)

torch.optim.lr_sheduler.CosineAnnealingLR(optimizer, T_max, eta_min=0, last_epoch=-1)

torch.optim.lr_sheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=10,
 verbose=False, threshold=0.0001, threshold_mode='rel', cooldown=0, min_lr=0, eps=1e-08)

optimizer = torch.optim.SGD(model.parameters(), args.lr,
 momentum=args.momentum, weight_decay=args.weight_decay)
scheduler = ReducelROnPlateau(optimizer,'min')
for epoch in range( args.start epoch, args.epochs ):
    train(train_loader , model, criterion, optimizer, epoch )
    result_avg, loss_val = validate(val_loader, model, criterion, epoch)
    # Note that step should be called after validate()
    scheduler.step(loss_val )

lr = base_lr * lambda(self.last_epoch)

torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda, last_epoch=-1)

ignored_params = list(map(id, net.fc3.parameters()))
base_params = filter(lambda p: id(p) not in ignored_params, net.parameters())
optimizer = optim.SGD([
        {'params': base_params},
        {'params': net.fc3.parameters(), 'lr': 0.001*100}], 0.001,          momentum=0.9,weight_decay=1e-4)
 # Assuming optimizer has two groups.
lambda1 = lambda epoch: epoch // 3
lambda2 = lambda epoch: 0.95 ** epoch
scheduler = LambdaLR(optimizer, lr_lambda=[lambda1, lambda2])
for epoch in range(100):
    train(...)
    validate(...)
    scheduler.step()
    print('epoch: ', i, 'lr: ', scheduler.get_lr())
    
輸出：
epoch: 0 lr: [0.0, 0.1]
epoch: 1 lr: [0.0, 0.095]
epoch: 2 lr: [0.0, 0.09025]
epoch: 3 lr: [0.001, 0.0857375]
epoch: 4 lr: [0.001, 0.081450625]
epoch: 5 lr: [0.001, 0.07737809374999999]
epoch: 6 lr: [0.002, 0.07350918906249998]
epoch: 7 lr: [0.002, 0.06983372960937498]
epoch: 8 lr: [0.002, 0.06634204312890622]
epoch: 9 lr: [0.003, 0.0630249409724609]
為什么第一個參數(shù)組的學(xué)習(xí)率會是 0 呢？ 來看看學(xué)習(xí)率是如何計算的。
第一個參數(shù)組的初始學(xué)習(xí)率設(shè)置為 0.001, 
lambda1 = lambda epoch: epoch // 3,
第 1 個 epoch 時，由 lr = base_lr * lmbda(self.last_epoch)，
可知道 lr = 0.001 *(0//3) ，又因為 1//3 等于 0，所以導(dǎo)致學(xué)習(xí)率為 0。
第二個參數(shù)組的學(xué)習(xí)率變化，就很容易看啦，初始為 0.1， lr = 0.1 * 0.95^epoch ，當
epoch 為 0 時， lr=0.1 ， epoch 為 1 時， lr=0.1*0.95。

# -*- coding:utf-8 -*-
'''本文件用于測試pytorch學(xué)習(xí)率調(diào)整策略'''
__author__ = 'puxitong from UESTC'
 
import torch
import torch.optim as optim
from torch.optim import lr_scheduler
from torchvision.models import AlexNet
import matplotlib.pyplot as plt 
 
model = AlexNet(num_classes=2)
optimizer = optim.SGD(params=model.parameters(), lr=0.1)
 
# 等間隔調(diào)整學(xué)習(xí)率，每訓(xùn)練step_size個epoch，lr*gamma
# scheduler = lr_scheduler.StepLR(optimizer, step_size=30, gamma=0.1)
 
# 多間隔調(diào)整學(xué)習(xí)率，每訓(xùn)練至milestones中的epoch，lr*gamma
# scheduler = lr_scheduler.MultiStepLR(optimizer, milestones=[10, 30, 80], gamma=0.1)
 
# 指數(shù)學(xué)習(xí)率衰減，lr*gamma**epoch
# scheduler = lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
 
# 余弦退火學(xué)習(xí)率衰減，T_max表示半個周期，lr的初始值作為余弦函數(shù)0處的極大值逐漸開始下降，
# 在epoch=T_max時lr降至最小值，即pi/2處，然后進入后半個周期，lr增大
# scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max=100, eta_min=0)
 
plt.figure()
x = list(range(100))
y = []
for epoch in range(100):
    scheduler.step()
    y.append(scheduler.get_lr()[0])
 
plt.plot(x, y)
plt.show()

def adjust_learning_rate(optimizer, epoch):
    """Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
    lr = args.lr * (0.1 ** (epoch // 30))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

def adjust_learning_rate(epoch, lr):
    if epoch <= 81:  # 32k iterations
      return lr
    elif epoch <= 122:  # 48k iterations
      return lr/10
    else:
      return lr/100

for epoch in range(epochs):
    lr = adjust_learning_rate(optimizer, epoch)  # 調(diào)整學(xué)習(xí)率
    optimizer = optim.SGD(net.parameters(), lr=lr, momentum=0.9, weight_decay=5e-4)
    ......
    optimizer.step()  # 采用新的學(xué)習(xí)率進行參數(shù)更新

# 例1：有兩個`param_group`即,len(optim.param_groups)==2
optim.SGD([
                {'params': model.base.parameters()},
                {'params': model.classifier.parameters(), 'lr': 1e-3}
            ], lr=1e-2, momentum=0.9)
 
# 例2：一個參數(shù)組
optim.SGD(model.parameters(), lr=1e-2, momentum=.9)

param_groups[0]: {'params': model.base.parameters()},
param_groups[1]: {'params': model.classifier.parameters(), 'lr': 1e-3}

scheduler.get_lr()

def adjust_learning_rate(optimizer, epoch):
    """Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
    lr = args.lr * (0.1 ** (epoch // 30))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

# 有兩個`param_group`即,len(optim.param_groups)==2
optim.SGD([
                {'params': model.base.parameters()},
                {'params': model.classifier.parameters(), 'lr': 1e-3}
          ], lr=1e-2, momentum=0.9)
 
#一個參數(shù)組
optim.SGD(model.parameters(), lr=1e-2, momentum=.9)