Recherche en grille et aléatoire avec mlr

Optimisation des hyperparamètres en R

Dr. Shirin Elsinghorst

Senior Data Scientist

Optimisation d’hyperparamètres avec mlr

Dans mlr, vous devez définir

l’espace de recherche pour chaque hyperparamètre
la méthode d’optimisation (p. ex. grille ou aléatoire)
la méthode de rééchantillonnage

Définir l’espace de recherche

makeParamSet(
  makeNumericParam(),
  makeIntegerParam(),
  makeDiscreteParam(),
  makeLogicalParam(),
  makeDiscreteVectorParam()
)

                                       Type  len             Def
autoencoder                         logical    -           FALSE
use_all_factor_level                logical    -            TRUE 
activation                         discrete    -       Rectifier
hidden                        integervector <NA>         200,200 
epochs                              numeric    -              10
train_samples_per_iteration         numeric    -              -2 
seed                                integer    -               -

getParamSet("classif.h2o.deeplearning")

adaptive_rate                        logical    -            TRUE
rho                                  numeric    -            0.99
epsilon                              numeric    -           1e-08
rate                                 numeric    -           0.005

                                       Type  len             Def
autoencoder                         logical    -           FALSE
use_all_factor_level                logical    -            TRUE
activation                         discrete    -       Rectifier
hidden                        integervector <NA>         200,200
epochs                              numeric    -              10
train_samples_per_iteration         numeric    -              -2
seed                                integer    -               -
adaptive_rate                       logical    -            TRUE

rho                                 numeric    -            0.99
epsilon                             numeric    -           1e-08
rate                                numeric    -           0.005

param_set <- makeParamSet(
  makeDiscreteParam("hidden", values = list(one = 10, two = c(10, 5, 10))),
  makeDiscreteParam("activation", values = c("Rectifier", "Tanh")),
  makeNumericParam("l1", lower = 0.0001, upper = 1),
  makeNumericParam("l2", lower = 0.0001, upper = 1))

Définir la méthode d’optimisation

Recherche en grille

ctrl_grid <- makeTuneControlGrid()
ctrl_grid

Tune control: TuneControlGrid
Same resampling instance: TRUE
Imputation value: <worst>
Start: <NULL>

Tune threshold: FALSE
Further arguments: resolution=10

Ne gère que les ensembles de paramètres discrets !

Recherche aléatoire

ctrl_random <- makeTuneControlRandom()
ctrl_random

Tune control: TuneControlRandom
Same resampling instance: TRUE
Imputation value: <worst>
Start: <NULL>
Budget: 100
Tune threshold: FALSE
Further arguments: maxit=100

Définir la stratégie de rééchantillonnage

cross_val <- makeResampleDesc("RepCV", 
                              predict = "both",
                              folds = 5 * 3)

param_set <- makeParamSet(...)
ctrl_grid <- makeTuneControlGrid()

task <- makeClassifTask(data = knowledge_train_data, 
                        target = "UNS")
lrn <- makeLearner("classif.h2o.deeplearning", 
                   predict.type = "prob", 
                   fix.factors.prediction = TRUE)

lrn_tune <- tuneParams(lrn,
                       task,
                       resampling = cross_val,
                       control = ctrl_grid,
                       par.set = param_set)

Ajuster les hyperparamètres

lrn_tune <- tuneParams(lrn, task, resampling = cross_val, control = ctrl_grid, par.set = param_set)

[Tune-y] 27: mmce.test.mean=0.6200000; time: 0.0 min
[Tune-x] 28: hidden=two; activation=Rectifier; l1=0.578; l2=1
[Tune-y] 28: mmce.test.mean=0.6800000; time: 0.0 min
[Tune-x] 29: hidden=one; activation=Rectifier; l1=0.156; l2=0.68
[Tune-y] 29: mmce.test.mean=0.4400000; time: 0.0 min
[Tune-x] 30: hidden=one; activation=Rectifier; l1=0.717; l2=0.427
[Tune-y] 30: mmce.test.mean=0.6600000; time: 0.0 min

[Tune] Result: hidden=two; activation=Tanh; l1=0.113; 
l2=0.0973 : mmce.test.mean=0.2000000

# tictoc
26.13 sec elapsed

Passons à la pratique !

Optimisation des hyperparamètres en R