PyG's Data.to(device) function behaves differently from PyTorch

[Boltzmachine]

In PyTorch, we now when data is a tensor, the to(device) function will return a copy, for example:

>>> data = torch.tensor(1)
>>> data.device
device(type='cpu')
>>> data.to('cuda').device
device(type='cuda', index=0)
>>> data.device
device(type='cpu')

Howevew, if data is a PyG Data object, the to(device) function will inplace change the device of original data object, for example

>>> data = Data(x=torch.tensor(1))
>>> data.x.device
device(type='cpu')
>>> data.to('cuda').x.device
device(type='cuda', index=0)
>>> data.x.device
device(type='cuda', index=0)

Is there any reason for such design. And what is the correct way if I want to make a cuda copy from PyG data in CPU.

[EdisonLeeeee]

Hi @Boltzmachine

In PyG, Data is designed as a dictionary-like object, so to(device) operation is recursively applied to each key in Data. If you want to make a copy, you might use copy.copy like that:

import copy
>>> data = Data(x=torch.tensor(1))
>>> data2 = copy.copy(data)

>>> data.to('cuda').x.device
device(type='cuda', index=0)
>>> data2.x.device
device(type='cpu')

[Boltzmachine]

Thanks so much! I also found this by checking the source code of InMemoryDataset.