Technology Sharing

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

In PyTorch, you can save and load the model and then continue training by following these steps:

1. Save the model

   There are usually two ways to save a model:

   • Save the entire model (including network structure, weights, etc.):

     torch.save(model, 'model.pth')
     

   • Only save the model's state_dict (which only contains weight parameters). This method is recommended because it saves storage space and is more flexible when loading:

     torch.save(model.state_dict(), 'model_weights.pth')
     

2. Loading the model

   Correspondingly, there are two ways to load the model:

   • If you have saved the entire model before, you can load it directly in the following way:

     model = torch.load('model.pth')
     

   • If you have only saved state_dict before, you need to instantiate a model with the same structure as the original model, and thenload_state_dict()Method to load weights:

     # 实例化一个与原模型结构相同的模型
     model = YourModelClass()
      
     # 加载保存的state_dict
     model.load_state_dict(torch.load('model_weights.pth'))
      
     # 确保将模型转移到正确的设备上（例如GPU或CPU）
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     model.to(device)
     

3. Continue training

   Once the model is loaded, you can proceed with training. Make sure you have defined the loss function and optimizer, and that their states are loaded correctly (if you saved them previously). Then, proceed with the normal training process.

   # 定义损失函数和优化器
   criterion = nn.CrossEntropyLoss()
   optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
    
   # 如果之前保存了优化器状态，也可以加载
   optimizer.load_state_dict(torch.load('optimizer.pth'))
    
   # 开始训练
   for epoch in range(num_epochs):
       for inputs, labels in dataloader:
           inputs, labels = inputs.to(device), labels.to(device)
           
           optimizer.zero_grad()
           outputs = model(inputs)
           loss = criterion(outputs, labels)
           loss.backward()
           optimizer.step()
   

This way, you can continue training the model from where you last saved it.