Update detoxify.py

[ACMCMC]

Optimize the generation of results in predict().

This PR doesn't change any behavior other than using PyTorch directly to generate the results - no need to do a halfway conversion to NumPy. I also removed the need to do a list comprehension.

[laurahanu]

Thanks for the PR! Looks good, although I'm still in favour of returning numpy arrays as switching to pytorch tensors might break compatibility if external code expects numpy arrays

[ACMCMC]

Thanks for the comment! That's a good point.

However, if you look at the code more closely, you'll see that we're never actually returning NumPy arrays. Instead, we do: PyTorch Tensor -> NumPy Array -> Python List (that's what's returned).

So, in other words, the end user would still get a Dict[str, Union[float, List[float]]]. i.e., not a breaking change :)

If you want I can make that change but I think we could omit that as the return data type is not changed 👍🏻