Error while trying to use torch.rand_like() on a one dimensional array created using torch.arange()

[VijayV28]

Hey,
I was trying to implement the torch.rand_like() on an array generated by torch.arange() and got an error:

Array:

range1 = torch.arange(start=10, end=101, step=10)

Random function:

ten_random_numbers = torch.rand_like(input=(range1))

This threw an error as shown below:
RuntimeError: "check_uniform_bounds" not implemented for 'Long'

However, I later discovered that the dtype of the tensor generated by the torch.arange function is int64.
The rand_like() works well once I change the dtype back to float32:

range1 = torch.arange(start=10, end=101, step=10, dtype=torch.float32)

I'd really like to know why doesn't rand_like() support int64 datatype while the other tensor_like() functions do.

[mrdbourke]

Hi @VijayV28,

That's an excellent question I don't have the answer to.

I'd like to know too.

Perhaps asking in the PyTorch forums could help?

https://discuss.pytorch.org

[VijayV28]

Thanks a lot for your response. Will try your suggestion 👍

[rezanematpour]

The torch.rand_like() is specifically designed to sample from a continuous uniform distribution between 0 and 1. The numbers it generates are floating-point values (e.g., 0.112, 0.875, 0.999).
Integer data types like int64 can only store whole numbers (0, 1, 2, ...). Trying to force a fractional value like 0.739 into an integer container creates a logical conflict:
Should it be truncated to 0?
Should it be rounded to 1?
Either choice would destroy the "uniform" nature of the distribution, making the function's behavior mathematically incorrect. Because this operation is ambiguous and misleading, PyTorch disallows it.