search

msmbuilder / osprey

🦅Hyperparameter optimization for machine learning pipelines 🦅
http://msmbuilder.org/osprey
Apache License 2.0
74 stars 26 forks source link

refactoring #246

Closed RobertArbon closed 3 years ago

RobertArbon commented 6 years ago

Hi folks,

I'm refactoring my fork of Osprey. I thought I'd give you a heads up as you might want to accept the changes but then again you may not as it's quite a change.

I'm changing the structure so there's three main strategies: Grid, Random and Bayesian. The Bayesian strategy will implement the Gaussian processes (and eventually TPE) models.

The Bayesian strategy will load a SurrogateModel class (e.g. Gaussian Process, TPE, Regression Trees, whatever) and an AcquisitionFunction class.

Once that's done I was planning on making the Gaussian Process model use PyMC3 or GPFlow.

thoughts, comments, ideas etc?

cxhernandez commented 6 years ago

That sounds like a great plan to me! I think PyMC3 might be preferable since it has a more stable development pipeline and would probably be more flexible for acquisition function development later on. I've been burned by GPy too many times in the past.... That being said looking forward to seeing the PR!

RobertArbon commented 6 years ago

Ok great! I've just merged the latest master branch...should probably finish it off by end of this week.

Yeah, GPy needs to go, if not solely because of the horrendous documentation.

RobertArbon commented 3 years ago

Hi folks,

just finished my PhD where I worked out many of the details for a PyMC3 implementation, so I'll be devoting more time to this now.

However! I've borked my fork this so I may cancel the refactoring PR and start again. Hopefully not though.

Thanks for your patience.

Rob