The loop functions don't apply the parameter transformations

[sebffischer]

The default loopfunctions in e.g. bayesopt ego neglect parameter tranformations

library(paradox)
param_set = ps(
  batch_size = p_dbl(lower = log(16L, 2L), upper = log(512L, 2L), trafo = function(x) round(2^x)),
  max_unit = p_int(lower = log(64L, 2L), upper = log(1024L, 2L), trafo = function(x) round(2^x)),
  # TODO: In some datasets 1 or 5 does not appear (?) --> "True" bug
  num_layers = p_int(lower = 1L, upper = 5L),
  learning_rate = p_dbl(lower = 0, upper = 0.01),
  momentum = p_dbl(lower = 0.1, upper = 1),
  weight_decay = p_dbl(lower = 0, upper = 0.1)
)

generate_design_lhs(param_set, 10)$data
#>     batch_size max_unit num_layers learning_rate  momentum weight_decay
#>  1:   6.540366        8          3  0.0073469584 0.2580694  0.036643739
#>  2:   7.537337        7          2  0.0039075324 0.8925294  0.087852658
#>  3:   6.236520        7          1  0.0057498131 0.1502718  0.010984017
#>  4:   5.152692        9          3  0.0092623739 0.9222579  0.059232363
#>  5:   8.808069       10          4  0.0026422184 0.6750566  0.069379779
#>  6:   7.283262        6          4  0.0019102086 0.8052755  0.025566310
#>  7:   4.330742       10          5  0.0005144738 0.3196749  0.040949245
#>  8:   8.342514        8          5  0.0083950503 0.4488279  0.075199702
#>  9:   4.624166        6          1  0.0066763976 0.5141848  0.098863804
#> 10:   5.970330        9          2  0.0043219303 0.6260611  0.002390107

^{Created on 2022-01-05 by the reprex package (v2.0.1)}

[sebffischer]

sorry my misunderstanding ...