Open ericmjonas opened 8 years ago
I misremembered things a little bit - we've never done multi-lambda for the old BlockWeightedLeastSquares
. But for the new PerClassWeightedLeastSquares
it should be pretty simple to add it by just following wherever we have a loop around lambdas
in the unweighted implementation [1].
FWIW If we want feature compatibility across the solvers, the other one should also be possible -- but my guess is that it'll not be that cheap in terms of sharing costs and need some new code.
We often perform a sweep over lambdas to figure out the right regularization -- it matters a lot.
@shivaram mentioned that it's possible for the weighted solver to potentially solve for a large number of lambdas with minimal overhead. There was evidently some code in a repo associated with a paper that is an example of this. It might be possible (per Shiv's suggestion) that @tomerk could hack on this?