Why hash values of a same tensors change if we affect it to different variables?

1

Hi all,

Here is a very simple example to understand my question:

import torch

a = torch.arange(5).cuda()
b = torch.arange(5).cuda()

print(a.__hash__())
print(b.__hash__())

If you run this you will see that the hash value of a is different from the one of b which is weird to me. Is there any explanation for this? And is there a way to avoid this behavior? (I would like to update a dictionary with tensor valued keys and because of this, the keys will appears multiple times).

Thanks a lot in advance for your help!

Samuel

2

hash of Tensor is by object, not by value. It’s not reasonable to hash by value because (1) it’s expensive for large tensors and (2) every hash of a cuda tensor will be a synchronization.

1 Like
3

Thank you for your answer I understand. But on the other hand it means that I need to update my dictionary on the CPU and so make a round trip GPU-CPU with all my tensors. Anyway thanks again.

1 Like