pytorch 如何从checkpoints中继续训练

左1：从头开始训练时，lr的变化。 左2：从epoch100时开始训练

---

‍♂️ 张同学，[email protected] ，有问题请联系我

一、导入一些必要的包

import os
import sys
import pandas as pd
import numpy as np 
from tqdm import tqdm,trange
from matplotlib import pyplot as plt
import seaborn as sns
import json
import pathlib
from pathlib import Path
import torch
from torch import nn, einsum
import torch.nn.functional as F

from einops import rearrange, repeat
from einops.layers.torch import Rearrange

from transformers import(
    get_scheduler
)

二、定义模型

class MyModal(nn.Module):
    def __init__(self) -> None:
        super().__init__()
        self.tf = nn.TransformerEncoderLayer(d_model=512, nhead=8)
        self.linear = nn.Linear(512, 10)
        self.softmax = nn.Softmax(dim=1)
    
    def forward(self, x):
        x = self.tf(x)
        x = self.linear(x[:,0])
        x = self.softmax(x)
        return x

三、训练时代码

model = MyModal()
max_train_steps = 200

optimizer = torch.optim.AdamW(model.parameters(), lr=0.1)
lr_scheduler = get_scheduler(
    name='linear',
    optimizer=optimizer,
    num_warmup_steps=20,
    num_training_steps=max_train_steps,
)
criterion = nn.CrossEntropyLoss()

lrs = []

start_step = 0
for i in range(start_step):
    optimizer.step()
    lr_scheduler.step()   # 对优化器中的lr进行更新
    optimizer.zero_grad() # 更新模型记录的梯度为0

for i in range(max_train_steps-start_step):
    src = torch.rand(10,32,512)
    labels = torch.randint(0,10,[10]) # bs = 10
    output = model(src)
    loss = criterion(output, labels)
    loss.backward()
    optimizer.step()
    lr_scheduler.step()   # 对优化器中的lr进行更新
    optimizer.zero_grad() # 更新模型记录的梯度为0

    # https://blog.csdn.net/qq_41375318/article/details/115540896
    lrs.append(optimizer.state_dict()['param_groups'][0]['lr'])

x = np.arange(1, len(lrs)+1)
plt.plot(x, lrs)
plt.show()