[pytorch] the difference between cuda() and to (device)

problem

PyTorch Medium Tensor If you want to be in GPU Run in , There are two ways to do it , The first is x.cuda(), The second is x.to(device). Both methods can be realized GPU Up operation , So what's the difference between them ？

Method

import torch


device = 'cuda' if torch.cuda.is_available() else 'cpu'

a = torch.randn([3, 224, 224])

# (1) cuda() Poor extensibility of mode code , If the machine does not support GPU, You need to modify the code before you can CPU Up operation ;
a.cuda()

# (2) to(device) The mode code has good scalability , Even without GPU, You can also run code , You don't have to make any changes ;
a.to(device)

Conclusion

• Recommended to(device) The way , The main reason is that this programming method is easier to expand , and cuda() The machine must have GPU, Otherwise, you need to modify all the code ,to(device) The way is not limited by this ,device It can be either CPU It can also be GPU;