search

HIPS / Spearmint

Spearmint Bayesian optimization codebase
Other
1.55k stars 327 forks source link

Multi-Task support #71

Open mccajm opened 8 years ago

mccajm commented 8 years ago

Hello,

For k-fold cross validation, how much of the multi-task support is implemented? If I were to modify line 364 of main.py locally to return a random task, although not ideal, how would this affect the underlying model? What would be required for me to implement the missing components?

(Great framework btw! Sped up my training significantly and the cluster schedulers are an especially nice addition.)

Thank you, Adam

mgelbart commented 8 years ago

Thanks for the positive feedback, Adam! The short answer is that we haven't implemented the approach described in the "Multi-Task Bayesian Optimization" paper. While we have implemented the notion of multiple tasks (for constrained optimization, especially developed in the PESC branch), we have not implemented the multi-task GP that models the relationship between tasks. In other words, we model each task independently and the different only interact at the acquisition function stage. This joint modeling is the missing piece you would need for applying multi-task Spearmint to a k-fold cross validation scenario.