pytorch将GPU上训练的model load到CPU/GPU上

假设我们只保存了模型的参数(model.state_dict())到文件名为modelparameters.pth, model = Net()

1. cpu -> cpu或者gpu -> gpu:

checkpoint = torch.load('modelparameters.pth')

model.load_state_dict(checkpoint)

2. cpu -> gpu 1

checkpoint =torch.load('modelparameters.pth', map_location=lambda storage, loc: storage.cuda(1))

model.load_state_dict(checkpoint)

3. gpu 1 -> gpu 0

checkpoint =torch.load('modelparameters.pth', map_location={'cuda:1':'cuda:0'})

model.load_state_dict(checkpoint)

4. gpu -> cpu

checkpoint =torch.load('modelparameters.pth', map_location=lambda storage, loc: storage)

model.load_state_dict(checkpoint)

5. multi-gpu - > single gpu

Sometimes the error occur:  Unexpected key(s) in state_dict: "module.features. ...".，Expected ".features....". 这种情况往往存在于用多GPU训练，但是用单