参考：https://pytorch.org/docs/stable/notes/cuda.html
    torch.cuda.device()修改默认cuda设备，修改之后不管是.cuda()还是.to(device=cuda)都会在默认cuda设备上生成张量

    1. cuda = torch.device('cuda')     # Default CUDA device
    2. cuda0 = torch.device('cuda:0')
    3. cuda2 = torch.device('cuda:2')  # GPU 2 (these are 0-indexed)
    4. x = torch.tensor([1., 2.], device=cuda0)
    5. # x.device is device(type='cuda', index=0)
    6. y = torch.tensor([1., 2.]).cuda()
    7. # y.device is device(type='cuda', index=0)
    8. with torch.cuda.device(1):
    9.     # allocates a tensor on GPU 1
    10.     a = torch.tensor([1., 2.], device=cuda)
    11.     # transfers a tensor from CPU to GPU 1
    12.     b = torch.tensor([1., 2.]).cuda()
    13.     # a.device and b.device are device(type='cuda', index=1)
    14.     # You can also use ``Tensor.to`` to transfer a tensor:
    15.     b2 = torch.tensor([1., 2.]).to(device=cuda)
    16.     # b.device and b2.device are device(type='cuda', index=1)
    17.     c = a + b
    18.     # c.device is device(type='cuda', index=1)
    19.     z = x + y
    20.     # z.device is device(type='cuda', index=0)
    21.     # even within a context, you can specify the device
    22.     # (or give a GPU index to the .cuda call)
    23.     d = torch.randn(2, device=cuda2)
    24.     e = torch.randn(2).to(cuda2)
    25.     f = torch.randn(2).cuda(cuda2)
    26.     # d.device, e.device, and f.device are all device(type='cuda', index=2)