unnest fails for tuning archive if x_domain contains params with length > 1

[jemus42]

Moving here from slds-lmu/paper_2023_survival_benchmark#9

The underlying issue is the (exceedingly rare?) case where a learners as a p_uty that's a numeric vector of length > 1, in this example the survival SVM in mlr3extralearners which has a param gamma.mu = c(x, y), which is tuned by creating proxy p_dbls and using a trafo to "ressamble" the param passed down to the learner.

Minimal reprex for unnesting

xdt <- data.table::data.table(
  gamma = 2,
  mu = 1,
  x_domain = list(
    list(gamma.mu = c(2, 1))
  )
)

mlr3misc::unnest(xdt, "x_domain")

Reprex for learner with problematic param

library(mlr3)
library(paradox)
library(mlr3misc)
library(mlr3tuning)

LearnerRegrDebugMulti = R6::R6Class("LearnerRegrDebugMulti", inherit = LearnerRegr,
   public = list(
     #' @description
     #' Creates a new instance of this [R6][R6::R6Class] class.
     initialize = function() {
       super$initialize(
         id = "regr.debugmulti",
         feature_types = c("logical", "integer", "numeric", "character", "factor", "ordered"),
         predict_types = c("response"),
         param_set = ps(
           x                    = p_dbl(0, 1, tags = "train"),
           # same as in surv.svm
           gamma.mu            = p_uty(tags = c("train", "required"))
         ),
         properties = "missings",
         man = "mlr3::mlr_learners_regr.debugmulti",
         label = "Debug Learner for Regression"
       )
     }
   ),
   private = list(
     .train = function(task) {
       pv = self$param_set$get_values(tags = "train")
       truth = task$truth()
       model = list(
         response = mean(truth),
         se = sd(truth),
         pid = Sys.getpid()
       )

       set_class(model, "regr.debug_model")
     },

     .predict = function(task) {
       n = task$nrow
       pv = self$param_set$get_values(tags = "predict")

       predict_types = "response"
       prediction = named_list(mlr_reflections$learner_predict_types[["regr"]][[predict_types]])

       for (pt in names(prediction)) {
         value = rep.int(self$model[[pt]], n)

         prediction[[pt]] = value
       }

       return(prediction)
     }
   )
)
mlr_learners$add("regr.debugmulti", function() LearnerRegrDebugMulti$new())

lrn_base = lrn("regr.debugmulti", gamma.mu = c(0, 0))

instance = ti(
  task = tsk("mtcars"),
  learner = lrn_base,
  search_space = ps(
    gamma = p_dbl(0, 1),
    mu = p_dbl(0, 1),
    .extra_trafo = function(x, param_set) {
      # learner has tuple param gamma.mu = c(x, y)
      # we tune separately and reassemble via trafo
      x$gamma.mu = c(x$gamma, x$mu)
      x$gamma = x$mu = NULL
      x
    }
  ),
  resampling = rsmp("holdout"),
  terminator = trm("evals", n_evals = 3)
)

archive = tnr("grid_search")$optimize(instance)
#> INFO  [13:23:15.137] [bbotk] Starting to optimize 2 parameter(s) with '<OptimizerBatchGridSearch>' and '<TerminatorEvals> [n_evals=3, k=0]'
# [...truncated]

as.data.table(instance$archive)
#> Error: Tables have different number of rows (x: 3, y: 6)

# Because of this step
mlr3misc::unnest(archive, "x_domain")
#> Error: Tables have different number of rows (x: 1, y: 2)

^{Created on 2024-09-19 with reprex v2.1.1}