Any plans for scipy.optimize?

[Rolandw0w]

Description

I wonder if there are any plans/ETAs on scipy.optimize module?
Thank you!

Additional Information

No response

[leofang]

There are some discussions ongoing but no concrete plan yet. @Rolandw0w Is there a particular functionality in the scipy.optimize module that you'd like to have?

cc: @cjnolet for vis

[Rolandw0w]

The problem that brought me here is linear programming and its slowness for scipy implementation.
I thought of RAPIDS and cuml; I've successfully used them on Kaggle earlier, so I started looking for GPU implementations.
Thanks!

[cjnolet]

@Rolandw0w,

Indeed, the current discussions have been around generalizing some of the solvers from cuml and move them to the RAFT project (https://github.com/rapidsai/raft) so they can be exposed through the cupy.optimize package. Once we have the infrastructure to expose them from RAFT, we're hoping to add new solvers as the needs arise.

[Rolandw0w]

Sounds good, thank you!
Feel free to close this issue.

[kmaehashi]

Let's keep this issue open so that others can find the latest status 😃

[leofang]

xref: rapidsai/raft#374

[cjnolet]

@Rolandw0w I just want to follow up on this issue as we are now starting to make RAFT features available to cupy through pylibraft. If you had your choice, is there any particular linear programming method you'd prefer to see implemented first?

[Rolandw0w]

@cjnolet If I had to choose, I'd prefer HiGHS dual simplex. IIRC, this method is suitable for parallel implementation.
But any other method would be great too.

[leofang]

In addition to SLSQP, CG, BFGS and L-BFGS-B are also great options for starter.

[cjnolet]

Tagging @teju85 for further viz.

[domattioli]

Any update?

[ndjvisser]

Any update on this? Would love to have a gpu version of scipy.optimize.minimize using SLSQP

[taylorpatti]

Chiming in to agree with ndjvisser - I'd love a scipy.optimize.minimize implementation.

[lpsgverrilla]

Don't know if possible, but I would like very much scipy.optimize.differential_evolution on cupy!

[jakirkham]

Maybe folks asking for this could share more about their use cases? How would you use this functionality? What would it help you accomplish?

[lpsgverrilla]

Maybe folks asking for this could share more about their use cases? How would you use this functionality? What would it help you accomplish?

I don't have much knowledge about the subject at all, so I may be wrong in multiple accounts.
That said, recently I have started a deep quest into optimizing performance of a scipy.differential_evolution workflow (CPU).
One of the venues I explored was Differential Evolution in GPU, but I didn't find any python library offering it with same functionality of scipy's implementation, which is quite robust and rich.
I have found research papers citing how gpu could greatly improve performance of DE, though.
Currently I'm exploring other venues.