Difference in Computation Speed and Results Between MLR and MLR3 - mlr

Question

I don't get similar results when I use the same data and models using mlr and mlr3. Also I find mlr runs at least 20-fold faster. I used lung data from survival and I was able to replicate the difference in computation speed and results since I can't share my data.
mlr was completed in 1 min with C-index generally low compared to mlr3 that took 21 min to complete with C-index being much higher despite using same data, same preprocessing, same model and setting seed.
library(tidyverse)
library(tidymodels)
library(PKPDmisc)
library(mlr)
library(parallelMap)
library(survival)
# Data and Data Splitting
data = as_tibble(lung) %>%
mutate(status = if_else(status==1, 0, 1),
sex = factor(sex, levels = c(1:2), labels = c("male", "female")),
ph.ecog = factor(ph.ecog))
na <- sample(1:228, 228*0.1)
data$sex[na] <- NA
data$ph.ecog[na]<- NA
set.seed(123)
split <- data %>% initial_split(prop = 0.8, strata = status)
train <- split %>% training()
test <- split %>% testing()
# Task
task = makeSurvTask(id = "Survival", data = train, target = c("time", "status"))
# Resample
# For model assessment before external validation on test data
set.seed(123)
outer_cv = makeResampleDesc("CV", iter=4, stratify.cols = c("status")) %>%
makeResampleInstance(task)
# For feature selection and parameter tuning
set.seed(123)
inner_cv = makeResampleDesc("CV", iter=4, stratify.cols = c("status"))
# Learners
cox1 = makeLearner(id = "COX1", "surv.coxph") %>%
makeImputeWrapper(classes = list(factor = imputeMode(), numeric = imputeMedian()),
# Create dummy variable for factor features
dummy.classes = "factor") %>%
makePreprocWrapperCaret(ppc.center = TRUE, ppc.scale = TRUE) %>%
makeFeatSelWrapper(resampling = inner_cv, show.info = TRUE,
control = makeFeatSelControlSequential(method = "sfs"))
cox_lasso = makeLearner(id = "COX LASSO", "surv.glmnet") %>%
makeImputeWrapper(classes = list(factor = imputeMode(), numeric = imputeMedian()),
# Create dummy variable for factor features
dummy.classes = "factor") %>%
# Normalize numeric features
makePreprocWrapperCaret(ppc.center = TRUE, ppc.scale = TRUE) %>%
makeTuneWrapper(resampling = inner_cv, show.info = TRUE,
par.set = makeParamSet(makeNumericParam("lambda",lower = -3, upper = 0,
trafo = function(x) 10^x)),
control = makeTuneControlGrid(resolution = 10L))
cox_net = makeLearner(id = "COX NET", "surv.glmnet") %>%
makeImputeWrapper(classes = list(factor = imputeMode(), numeric = imputeMedian()),
# Create dummy variable for factor features
dummy.classes = "factor") %>%
# Normalize numeric features
makePreprocWrapperCaret(ppc.center = TRUE, ppc.scale = TRUE) %>%
makeTuneWrapper(resampling = inner_cv, show.info = TRUE,
par.set = makeParamSet(makeNumericParam("alpha", lower = 0, upper = 1,
trafo = function(x) round(x,2)),
makeNumericParam("lambda",lower = -3, upper = 1,
trafo = function(x) 10^x)),
control = makeTuneControlGrid(resolution = 10L))
# Benchmark
# parallelStartSocket(4)
start_time <- Sys.time()
set.seed(123)
mlr_bmr = benchmark(learners = list(cox1, cox_lasso, cox_net),
tasks = task,
resamplings = outer_cv,
keep.extract= TRUE,
models = TRUE)
end_time <- Sys.time()
mlr_time = end_time - start_time
# parallelStop()
mlr_res <- getBMRPerformances(mlr_bmr, as.df = TRUE) %>%
select(Learner = learner.id, Task = task.id, Cindex = cindex) %>%
mutate(Color_Package = "mlr",
Learner = word(str_replace(Learner, "\\.", " "), 1, -2))
##################################################################
library(mlr3verse)
# Task
task2 = TaskSurv$new(id = "Survival2", backend = train, time = "time", event = "status")
task2$col_roles$stratum = c("status")
# Resmaple
set.seed(123)
outer_cv2 = rsmp("cv", folds = 4)$instantiate(task2)
# For feature selection and parameter tuning
set.seed(123)
inner_cv2 = rsmp("cv", folds = 4)
# Learners
preproc = po("imputemedian", affect_columns = selector_type("numeric")) %>>%
po("imputemode", affect_columns = selector_type("factor")) %>>%
po("scale") %>>%
po("encode")
cox2 = AutoFSelector$new(learner = as_learner(preproc %>>%
lrn("surv.coxph")),
resampling = inner_cv2,
measure = msr("surv.cindex"),
terminator = trm("none"), # need to increase later
fselector = fs("sequential", strategy = "sfs")) # sfs is the default
cox2$id = "COX1"
cox_lasso2 = AutoTuner$new(learner = as_learner(preproc %>>%
lrn("surv.glmnet",
lambda = to_tune(p_dbl(lower = -3, upper = 0,
trafo = function(x) 10^x)))),
resampling = inner_cv2,
measure = msr("surv.cindex"),
terminator = trm("none"),
tuner = tnr("grid_search", resolution = 10))
cox_lasso2$id = "COX LASSO"
cox_net2 = AutoTuner$new(learner = as_learner(preproc %>>%
lrn("surv.glmnet",
alpha = to_tune(p_dbl(lower = 0, upper = 1)),
lambda = to_tune(p_dbl(lower = -3, upper = 1,
trafo = function(x) 10^x)))),
resampling = inner_cv2,
measure = msr("surv.cindex"),
terminator = trm("none"),
tuner = tnr("grid_search", resolution = 10))
cox_net2$id = "COX NET"
# Benchmark
desgin = benchmark_grid(tasks = task2,
learners = c(cox2, cox_lasso2, cox_net2),
resamplings = outer_cv2)
# future::plan("multisession")
# Error: Output type of PipeOp select during training (Task) incompatible with input type of PipeOp surv.coxph (TaskSurv)
start_time <- Sys.time()
set.seed(123)
mlr3_bmr = mlr3::benchmark(desgin)
end_time <- Sys.time()
mlr3_time = end_time - start_time
mlr3_res <- as.data.table(mlr3_bmr$score()) %>%
select(Task=task_id, Learner=learner_id, Cindex=surv.harrell_c) %>%
mutate(Color_Package = "mlr3")
mlr_res %>%
bind_rows(mlr3_res) %>%
ggplot(aes(Learner, Cindex, fill= Color_Package )) +
geom_boxplot(position=position_dodge(.8)) +
stat_summary(fun= mean, geom = "point", aes(group = Color_Package ),
position=position_dodge(.8), size = 3) +
labs(x="", y = " C-Index") +
theme_bw() + base_theme() + theme(legend.position = "top")