Open simonpcouch opened 1 year ago
This may actually be an issue with serialization—would be worth comparing object-in-memory vs. on disk sizes to ensure that rsample serialization tricks carry over inside of tune objects. If not, some notes from bringing this up in team meeting:
extract_*()
function rather than a butcher method.extract_*()
function ought to be the "happy path" inside functions that take initial results—i.e. if a function takes in initial results, it ought to convert to the extract_*()
ed object and then use it internally.
The
initial
argument totune_bayes()
,tune_sim_anneal()
, etc(?) can be expressed as either 1) a previous tuning result or 2) a number indicating the number of initial grid points to evaluate. In case 1), when initial searches result in a large object—as a result of large input data, many grid points, etc—that I know I will want to pass along to another tuning function later (i.e. in another session), I've often wished there were some way to "butcher" the object so that it only contains the information that will be extracted when it's supplied asinitial
. That "butcher"ing process may even becollect_metrics()
, in which case we could just add anotherinitial
parameterization.