import math
import matplotlib.pyplot as plt
import torch.optim as optim
from torchvision.models import resnet18

lr_rate = 0.0001
model = resnet18(num_classes=10)

# T_max = 1000

epoch_total = 25
epoch_iter = 609
warm_up = 800

lambda1 = lambda epoch: (epoch / warm_up) if epoch < warm_up else 0.5 * (math.cos((epoch - warm_up)/(epoch_total*epoch_iter - warm_up) * math.pi) + 1)
optimizer = optim.SGD(model.parameters(), lr=lr_rate, momentum=0.9, nesterov=True)
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda1)

index = 0
x = []
y = []
for epoch in range(epoch_total):
    for batch in range(609):
        x.append(index)
        y.append(optimizer.param_groups[0]['lr'])
        index += 1
        scheduler.step()

plt.figure(figsize=(10, 8), dpi=200)
plt.xlabel('batch stop')
plt.ylabel('learning rate')
plt.plot(x, y, color='r', linewidth=2.0, label='modify data')
plt.legend(loc='upper right')
plt.savefig('result.png')
plt.show()

结果：

参考链接：基于PyTorch实现cosine learning rate_l_cos learning rate