Open PGijsbers opened 3 years ago
Hi @PGijsbers, this would certainly be a nice feature and definitely something to have in once asklearn 2 becomes non-experimental. I had a quick look to see how possible this is and I don't imagine it's too bad depending on the definition of configurable for policy selector
and portifolio
.
Simply changing the portifolio
file location and enabling disabling it should be quite straightforward and likewise with white listing or black listing certain policys. However if there is a need for injecting a custom policy selector
object it might be a bit more involved ... but also manageable.
Do you need this for the automlbenchmark
and do you have a time urgency for this?
Yes it's something we'd ultimately like to use in the AutoML benchmark. Though realistically I don't think we can wait for this to be all done and set up properly before doing our next set of experiments, which have been delayed for months already.
No problem, seeing as this will probably not be done in time for the benchmarks, I will talk to @mfeurer about how we do this. Perhaps autosklearn2 will become the default before the next benchmark run.
Hi, We have been using autosklearn and automlbenchmark in R&D at SAS Institiute for autoML pipeline. Huge applauses to the great work of you all first! @eddiebergman @PGijsbers
I have a few questions regarding portfolio.
Thanks, Yonglin
As per the title, it will be useful for ASKL2 to have a configurable
portfolio
andpolicy selector
. It's beneficial for research (avoiding 'cheating' through meta-learning in a benchmark) or for customization.Issue opened on the request of @mfeurer