How do I assign a numpy.int64 to a torch.cuda.FloatTensor?

Question

Im not able to assign int64 to torch tensor. I have got the following tensor

tempScale = torch.zeros((total, len(scale))).cuda() if useGpu else torch.zeros((nbPatchTotal, len(scale)))

In my code, when I'm using the following line, it's throwing an error message

tmpScale[:, j] = scale

The error message is

TypeError: can't assign a numpy.int64 to a torch.cuda.FloatTensor

what am I missing?

What does `scale` contain? You might try `scale.astype('float64')`. — kHarshit, Apr 02 '20 at 18:01
@kHarshit that won't work, especially considering one is a GPU tensor and one is a numpy array. — jodag, Apr 02 '20 at 22:20

jodag · Accepted Answer · 2020-04-02T22:24:28.673

5

You have to convert scale to a torch tensor of the same type and device as tmpScale before assignment.

tmpScale[:, j] = torch.from_numpy(scale).to(tmpScale)

Note that this is casting scale from an int64 to a float32 which will likely result in a loss of precision if values in scale have magnitude larger than 2²⁴ (about 16 million).

edited Apr 02 '20 at 22:24

answered Apr 02 '20 at 22:18

jodag

19,885
5
47
66

How do I assign a numpy.int64 to a torch.cuda.FloatTensor?

1 Answers1