Re: Add softplus implementation in scipy.special

On 4/13/23, Robert Kern <[email protected]> wrote:
> +1 on adding `softplus()`.
>
> I would not worry about reading StackOverflow in this case. Our general
> rule of thumb is that the minimum copyrightable segment of code is about 10
> lines. Everything about softplus is smaller than that. And we would not be
> copying the code straight in, in any case. We'll want a ufunc, I think.
>

For anyone reading the mailing list but not following the github
issue: I added a comment
(https://github.com/scipy/scipy/issues/17905#issuecomment-1507247170)
that shows a simple option for computing `log1pexp(x)` (aka
`softmax(x)`): use `np.logaddexp(0, x)`.

Warren


> On Thu, Apr 13, 2023 at 5:17 AM Pamphile Roy <[email protected]>
> wrote:
>
>> Hi Aadya,
>>
>> Thank you for sending the email. This is in reference to the issue:
>> https://github.com/scipy/scipy/issues/17905
>>
>> Note that we cannot use code from StackOverflow due to licensing
>> incompatibilities. See here for more details
>> https://scipy.github.io/devdocs/dev/hacking.html#license-considerations
>>
>> Cheers,
>> Pamphile
>>
>>
>> On 13.04.2023, at 10:28, [email protected] wrote:
>>
>> Hello Everyone,
>> It might be nice to have a numerically stable softplus implementation, ie
>> np.log1p(np.exp(x))
>>
>> This implementation can be based on the following stackoverflow answers :
>>
>> https://cs.stackexchange.com/questions/110798/numerically-stable-log1pexp-calculation
>>
>> https://stackoverflow.com/questions/44230635/avoid-overflow-with-softplus-function-in-python
>>
>> It can have a good place in the scipy module as it has other applications
>> apart from ML/AI like it's a quite natural penalty function in
>> optimization
>> if one desires a smooth penalty in some optimization problems.
>>
>> All opinions are welcome. Let's discuss this?
>> _______________________________________________
>> SciPy-Dev mailing list -- [email protected]
>> To unsubscribe send an email to [email protected]
>> https://mail.python.org/mailman3/lists/scipy-dev.python.org/
>> Member address: [email protected]
>>
>>
>> _______________________________________________
>> SciPy-Dev mailing list -- [email protected]
>> To unsubscribe send an email to [email protected]
>> https://mail.python.org/mailman3/lists/scipy-dev.python.org/
>> Member address: [email protected]
>>
>
>
> --
> Robert Kern
>
_______________________________________________
SciPy-Dev mailing list -- [email protected]
To unsubscribe send an email to [email protected]
https://mail.python.org/mailman3/lists/scipy-dev.python.org/
Member address: [email protected]