深度学习基础知识 学习率调度器的用法解析
1、自定义学习率调度器**：**torch.optim.lr_scheduler.LambdaLR
实验代码：
python
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'])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：
python
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