feat: OptimizerChain #145
feat: OptimizerChain #145
Conversation
|
I like the idea conceptually but I don't think I would introduce this into the package as it nothing existencial to the topic of optimization. Also it would need more flexibility to be of practical use (eg. Determine the part of the instance the next optimizer is allowed to see). It would be a nice gallery post, though. |
|
I think it should be added. |
|
I agree with @mb706, this is something we will very often need.
|
Co-authored-by: Jakob Richter <code@jakob-r.de>
…to optimizer_chain
| param_sets = vector(mode = "list", length = length(optimizers)) | ||
| ids_taken = character(0L) | ||
| # for each optimizer check whether the id of the param_set | ||
| # (decuded from the optimizer class) is already taken; |
| private$.ids = map_chr(param_sets, "set_id") | ||
| super$initialize( | ||
| param_set = ParamSetCollection$new(param_sets), | ||
| param_classes = Reduce(intersect, mlr3misc::map(optimizers, "param_classes")), |
There was a problem hiding this comment.
I don't think mlr3misc:: is necessary?
| .optimize = function(inst) { | ||
| terminator = inst$terminator | ||
| on.exit({inst$terminator = terminator}) | ||
| inner_inst = inst$clone(deep = TRUE) |
There was a problem hiding this comment.
would be good if this could be avoided, so when an error happens in the inner optimization then the user's inst at least has the partial optimization progress. Did I do this in my original implementation?
There was a problem hiding this comment.
We specifically do the on.exit() thing so we can modify the inst's $terminator temporarily without needing to clone, or did something get uncovered that does make this necessary?
|
Clearly distinguish between running optimizers in a chain (without sharing info) vs. chaining optimizers that can rely on the archive data of the previous optimizer runs. |
The idea of
OptimizerChainis to allow for sequentially running multipleOptimizers with potential additionalTerminators.Credit to @mb706 who implemented this initially here.
Could also be used for running the same
Optimizerwith random restarts (as shown in the example).