R/AutoxgboostMC.R
In ja-thomas/autoxgboostMC: Automatic Preprocessing, Fitting and Tuning for xgboost using multiple criteria.
#' @title Fit and optimize a xgboost model for multiple criteria
#'
#' @description
#' An xgboost model is optimized based on a set of measures (see [\code{\link[mlr]{Measure}}]).
#' The bounds of the parameter in which the model is optimized, are defined by \code{\link{autoxgbparset}}.
#' For the optimization itself Bayesian Optimization with \pkg{mlrMBO} is used.
#' Without any specification of the control object, the optimizer runs for for 160 iterations or 1 hour,
#' whichever happens first.
#' Both the parameter set and the control object can be set by the user.
#'
#' Arguments to `.$new()`:
#' @param measures [list of \code{\link[mlr]{Measure}}]\cr
#' Performance measure. If \code{NULL} \code{\link[mlr]{getDefaultMeasure}} is used.
#' @param early_stopping_measure [\code{\link[mlr]{Measure}}]\cr
#' Performance measure used for early stopping. Picks the first measure
#' defined in measures by default.
#' @param parset [\code{\link[ParamHelpers]{ParamSet}}]\cr
#' Parameter set to tune over. Default is \code{\link{autoxgbparset}}.
#' Can be updated using `.$set_parset()`.
#' @param nthread [integer(1)]\cr
#' Number of cores to use.
#' If \code{NULL} (default), xgboost will determine internally how many cores to use.
#' Can be set using `.$set_nthread()`.
#'
#' Arguments to `.$fit()`:
#' @param task [\code{\link[mlr]{Task}}]\cr
#' The task to be trained.
#' @param iterations [\code{integer(1L}]\cr
#' Number of MBO iterations to do. Will be ignored if custom \code{control} is used.
#' Default is \code{160}.
#' @param time.budget [\code{integer(1L}]\cr
#' Time that can be used for tuning (in seconds). Will be ignored if a custom \code{control} is used.
#' Default is \code{3600}, i.e., one hour.
#' @param build.final.model [\code{logical(1)}]\cr
#' Should the model with the best found configuration be refitted on the complete dataset?
#' Default is \code{FALSE}.
#' @param control [\code{\link[mlrMBO]{MBOControl}}]\cr
#' Control object for optimizer.
#' If not specified, the default \code{\link[mlrMBO]{makeMBOControl}}] object will be used with
#' \code{iterations} maximum iterations and a maximum runtime of \code{time.budget} seconds.
#'
#' Additional arguments that control the process:
#' @param mbo.learner [\code{\link[mlr]{Learner}}]\cr
#' Regression learner from mlr, which is used as a surrogate to model our fitness function.
#' If \code{NULL} (default), the default learner is determined as described here:
#' \link[mlrMBO]{mbo_default_learner}.
#' Can be set using `.$set_mbo_learner()`.
#' @param design.size [\code{integer(1)}]\cr
#' Size of the initial design. Default is \code{15L}.
#' Can be set via `.$set_design_size()`
#' @param max.nrounds [\code{integer(1)}]\cr
#' Maximum number of allowed boosting iterations. Default is \code{10^6}.
#' Can be set via `.$set_max_nrounds()`.
#' @param early.stopping.rounds [\code{integer(1L}]\cr
#' After how many iterations without an improvement in the boosting OOB error should be stopped?
#' Default is \code{10}.
#' Can be set via `.$set_early_stopping_rounds()`.
#' @param early.stopping.fraction [\code{numeric(1)}]\cr
#' What fraction of the data should be used for early stopping (i.e. as a validation set).
#' Default is \code{4/5}.
#' Can be set via `.$set_early_stopping_fraction()`.
#' @param impact.encoding.boundary [\code{integer(1)}]\cr
#' Defines the threshold on how factor variables are handled. Factors with more levels than the \code{"impact.encoding.boundary"} get impact encoded while factor variables with less or equal levels than the \code{"impact.encoding.boundary"} get dummy encoded.
#' For \code{impact.encoding.boundary = 0L}, all factor variables get impact encoded while for \code{impact.encoding.boundary = .Machine$integer.max}, all of them get dummy encoded.
#' Default is \code{10}.
#' Can be set via `.$set_impact_encoding_boundary()`.
#' @param tune.threshold [logical(1)]\cr
#' Should thresholds be tuned? This has only an effect for classification, see \code{\link[mlr]{tuneThreshold}}.
#' Default is \code{TRUE}.
#' Can be set via `.$set_tune_threshold()`.
#'
#' @export
#' @examples
#' \donttest{
#' iris.task = makeClassifTask(data = iris, target = "Species")
#' axgb = AutoxgboostMC$new(measure = auc)
#' axgb$fit(t, time.budget = 5L)
#' p = axgb$predict(iris.task)
#' }
AutoxgboostMC = R6::R6Class("AutoxgboostMC",
public = list(
measures = NULL,
control = NULL,
parset = NULL,
design.size = 15L,
mbo.learner = NULL,
iterations = NULL,
time.budget = NULL,
max.nrounds = 3*10^3L,
early.stopping.rounds = 20L,
early.stopping.fraction = 4/5,
impact.encoding.boundary = 10L,
tune.threshold = TRUE,
nthread = NULL,
resample_instance = NULL,
baselearner = NULL,
preproc_pipeline = NULL,
model = NULL,
obj_fun = NULL,
optim.result = NULL,
build.final.model = NULL,
initialize = function(measures = NULL, parset = NULL, nthread = NULL) {
assert_list(measures, types = "Measure", null.ok = TRUE)
assert_class(parset, "ParamSet", null.ok = TRUE)
# Set defaults
self$measures = coalesce(measures, list(getDefaultMeasure(task)))
self$parset = coalesce(parset, autoxgboostMC::autoxgbparset)
self$nthread = assert_integerish(nthread, lower = 1, len = 1L, null.ok = TRUE)
},
print = function(...) {
catf("AutoxgboostMC Learner")
catf("Trained: %s", ifelse(is.null(self$model), "no", "yes"))
if (!is.null(self$model)) print(self$model)
},
fit = function(task, iterations = 160L, time.budget = 3600L, build.final.model = TRUE, control = NULL) {
assert_class(task, "SupervisedTask")
self$iterations = assert_integerish(iterations)
self$time.budget = assert_integerish(time.budget)
self$build.final.model = assert_flag(build.final.model)
self$control = control
if (!is.null(self$model)) {
self$continue_fit()
} else {
self$baselearner = self$make_baselearner(task)
transf_tasks = self$build_transform_pipeline(task)
self$baselearner = setHyperPars(self$baselearner, early.stopping.data = transf_tasks$task.test)
self$obj_fun = self$make_objective_function(transf_tasks)
self$optim.result = self$optimize_pipeline_mbo()
}
lrn = self$build_final_learner()
mod = NULL
if(build.final.model) mod = train(lrn, task)
self$model = AutoxgbResult$new(
optim_result = self$optim.result,
final_learner = lrn,
final_model = mod,
measures = self$measures,
preproc_pipeline = self$preproc_pipeline
)
},
continue_fit = function(task, iterations = 160L, time.budget = 3600L, build.final.model = TRUE) {
},
predict = function(newdata) {
predict(self$model$final_model, newdata)
},
# AutoxgboostMC steps
make_baselearner = function(task) {
tt = getTaskType(task)
td = getTaskDesc(task)
req_prob_measure = sapply(self$measures, function(x) {
any(getMeasureProperties(x) == "req.prob")
})
pv = list()
if (!is.null(self$nthread))
pv$nthread = self$nthread
if (tt == "classif") {
predict.type = ifelse(any(req_prob_measure) | self$tune.threshold, "prob", "response")
if(length(td$class.levels) == 2) {
objective = "binary:logistic"
eval_metric = "error"
parset = c(self$parset, makeParamSet(makeNumericParam("scale_pos_weight", lower = -10, upper = 10, trafo = function(x) 2^x)))
} else {
objective = "multi:softprob"
eval_metric = "merror"
}
baselearner = makeLearner("classif.xgboost.earlystop", id = "classif.xgboost.earlystop",
predict.type = predict.type, eval_metric = eval_metric, objective = objective,
early_stopping_rounds = self$early.stopping.rounds, maximize = !self$early_stopping_measure$minimize,
max.nrounds = self$max.nrounds, par.vals = pv)
} else if (tt == "regr") {
predict.type = NULL
objective = "reg:linear"
eval_metric = "rmse"
baselearner = makeLearner("regr.xgboost.earlystop", id = "regr.xgboost.earlystop",
eval_metric = eval_metric, objective = objective, early_stopping_rounds = self$early.stopping.rounds,
maximize = !self$early_stopping_measure$minimize, max.nrounds = self$max.nrounds, par.vals = pv)
} else {
stop("Task must be regression or classification")
}
return(baselearner)
},
# Build pipeline
build_transform_pipeline = function(task) {
has.cat.feats = sum(getTaskDesc(task)$n.feat[c("factors", "ordered")]) > 0
self$preproc_pipeline = NULLCPO
#if (!is.null(task$feature.information$timestamps))
# preproc_pipeline %<>>% cpoExtractTimeStampInformation(affect.names = unlist(task$feature.information$timestamps))
if (has.cat.feats) {
self$preproc_pipeline %<>>% generateCatFeatPipeline(task, self$impact.encoding.boundary)
}
self$preproc_pipeline %<>>% cpoDropConstants()
# process data and apply pipeline
# split early stopping data
if (is.null(self$resample_instance))
self$resample_instance = makeResampleInstance(makeResampleDesc("Holdout", split = self$early.stopping.fraction), task)
task.test = subsetTask(task, self$resample_instance$test.inds[[1]])
task.train = subsetTask(task, self$resample_instance$train.inds[[1]])
task.train %<>>% self$preproc_pipeline
task.test %<>>% retrafo(task.train)
return(list(task.train = task.train, task.test = task.test))
},
make_objective_function = function(transf_tasks) {
is_thresholded_measure = sapply(self$measures, function(x) {
props = getMeasureProperties(x)
any(props == "req.truth") & !any(props == "req.prob")
})
if (!any(is_thresholded_measure) & self$tune.threshold) {
warning("Threshold tuning is active, but no measure for tuning thresholds!
Skipping threshold tuning!")
self$tune.threshold = FALSE
}
smoof::makeMultiObjectiveFunction(name = "optimizeWrapperMultiCrit",
fn = function(x) {
x = x[!vlapply(x, is.na)]
lrn = setHyperPars(self$baselearner, par.vals = x)
mod = train(lrn, transf_tasks$task.train)
pred = predict(mod, transf_tasks$task.test)
nrounds = self$get_best_iteration(mod)
# For now we tune threshold of first applicable measure.
if (self$tune.threshold && getTaskType(transf_tasks$task.train) == "classif") {
tune.res = tuneThreshold(pred = pred, measure = self$measures[is_thresholded_measure][[1]])
if (length(self$measures[-which(is_thresholded_measure)[1]]) > 0) {
res = performance(pred, self$measures[-which(is_thresholded_measure)[1]], model = mod, task = transf_tasks$task.test)
res = c(res, tune.res$perf)
} else {
res = tune.res$perf
}
attr(res, "extras") = list(nrounds = nrounds, .threshold = tune.res$th)
} else {
res = performance(pred, self$measures, model = mod, task = transf_tasks$task.test)
attr(res, "extras") = list(nrounds = nrounds)
}
return(res)
},
par.set = self$parset, noisy = FALSE, has.simple.signature = FALSE, minimize = sapply(self$measures, function(x) x$minimize),
n.objectives = length(self$measures)
)
},
optimize_pipeline_mbo = function() {
assert_class(self$control, "MBOControl", null.ok = TRUE)
# Set defaults
if (is.null(self$control)) {
measures_ids = sapply(self$measures, function(x) x$id)
ctrl = makeMBOControl(n.objectives = length(self$measures), y.name = measures_ids)
self$control = setMBOControlTermination(ctrl, iters = self$iterations,
time.budget = self$time.budget)
if (self$is_multicrit) {
self$control = setMBOControlMultiObj(self$control, method = "dib",dib.indicator = "eps")
self$control = setMBOControlInfill(self$control, crit = makeMBOInfillCritDIB(cb.lambda = 2L))
}
}
des = generateDesign(n = self$design.size, self$parset)
optim.result = mbo(fun = self$obj_fun, control = self$control, design = des, learner = self$mbo.learner)
if (self$is_multicrit) {
# Fill best.ind, x and y using "best on early stopping measure".
optim.result$best.ind = self$get_best_ind(optim.result)
pars = names(optim.result$opt.path$par.set$pars)
optim.result$x = as.list(optim.result$opt.path$env$path[optim.result$best.ind, pars])
optim.result$y = as.list(optim.result$opt.path$env$path[optim.result$best.ind, self$measure_ids])
}
return(optim.result)
},
build_final_learner = function() {
nrounds = self$get_best_from_opt("nrounds")
pars = trafoValue(self$parset, self$optim.result$x)
pars = pars[!vlapply(pars, is.na)]
lrn = if (!is.null(self$baselearner$predict.type)) {
makeLearner("classif.xgboost.custom", nrounds = nrounds,
objective = self$baselearner$par.vals$objective,
predict.type = self$baselearner$predict.type,
predict.threshold = self$get_best_from_opt(".threshold"))
} else {
makeLearner("regr.xgboost.custom", nrounds = nrounds, objective = objective)
}
lrn = setHyperPars2(lrn, par.vals = pars)
lrn = self$preproc_pipeline %>>% lrn
#FIXME mlrCPO #39
#lrn$properties = c(lrn$properties, "weights")
return(lrn)
},
# Setters for various hyperparameters
set_max_nrounds = function(value) {
self$max.nrounds = assert_integerish(value, lower = 1L, len = 1L)
},
set_early_stopping_rounds = function(value) {
self$early.stopping.rounds = assert_integerish(value, lower = 1L, len = 1L)
},
set_early_stopping_fraction = function(value) {
self$early.stopping.fraction = assert_numeric(early.stopping.fraction, lower = 0, upper = 1, len = 1L)
},
set_design_size = function(value) {
self$design.size = assert_integerish(design.size, lower = 1L, len = 1L)
},
set_tune_threshold = function(value) {
self$tune.threshold = assert_flag(value)
},
set_impact_encoding_boundary = function(value) {
self$impact_encoding_boundary = assert_integerish(value, lower = 0, len = 1L)
},
set_nthread = function(value) {
self$nthread = assert_integerish(value, lower = 1, len = 1L, null.ok = TRUE)
},
set_measures = function(value) {
self$measures = assert_list(value, types = "Measure", null.ok = TRUE)
},
set_parset = function(value) {
self$parset = assert_class(value, "ParamSet", null.ok = TRUE)
},
set_resample_instance = function(value) {
self$resample_instance = assert_class(value, "ResampleInstance", null.ok = TRUE)
},
get_best_ind = function(optim.result) {
# We can either select a best index from the outside
if (!is.null(self$selected.ind)) best.ind = selected.ind
# Or auto-choose one according to the early stopping measure
if (is.null(optim.result$best.ind)) {
if (self$early_stopping_measure$minimize) {
best.ind = which.min(optim.result$opt.path$env$path[[self$early_stopping_measure$id]])
} else {
best.ind = which.max(optim.result$opt.path$env$path[[self$early_stopping_measure$id]])
}
} else {
best.ind = optim.result$best.ind
}
return(best.ind)
},
#Get best value from optimization result
# @param what [`character(1)`]: "nrounds" or ".threshold"
get_best_from_opt = function(what) {
self$optim.result$opt.path$env$extra[[self$optim.result$best.ind]][[what]]
},
# Get the iteration parameter of a fitted xboost model with early stopping
get_best_iteration = function(mod) {
getLearnerModel(mod, more.unwrap = TRUE)$best_iteration
},
plot_pareto_front = function() {self$model$plot_pareto_front()},
plot_results = function() {self$model$plot_results()}
),
active = list(
early_stopping_measure = function(value) {
if (missing(value)) {
self$measures[[1]]
} else {
measure_ids = sapply(self$measures, function(x) x$id)
assert_list(value, types = "Measure", null.ok = TRUE)
self$measures = c(value, self$measures[-which(value$id == measure_ids)])
messagef("Setting %s as early stopping measure!", value$id)
}
},
is_multicrit = function() {
length(self$measures) > 1
},
measure_ids = function() {
sapply(self$measures, function(x) x$id)
}
)
)
ja-thomas/autoxgboostMC documentation built on May 17, 2019, 4:22 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
#' @title Fit and optimize a xgboost model for multiple criteria
#'
#' @description
#' An xgboost model is optimized based on a set of measures (see [\code{\link[mlr]{Measure}}]).
#' The bounds of the parameter in which the model is optimized, are defined by \code{\link{autoxgbparset}}.
#' For the optimization itself Bayesian Optimization with \pkg{mlrMBO} is used.
#' Without any specification of the control object, the optimizer runs for for 160 iterations or 1 hour,
#' whichever happens first.
#' Both the parameter set and the control object can be set by the user.
#'
#' Arguments to `.$new()`:
#' @param measures [list of \code{\link[mlr]{Measure}}]\cr
#' Performance measure. If \code{NULL} \code{\link[mlr]{getDefaultMeasure}} is used.
#' @param early_stopping_measure [\code{\link[mlr]{Measure}}]\cr
#' Performance measure used for early stopping. Picks the first measure
#' defined in measures by default.
#' @param parset [\code{\link[ParamHelpers]{ParamSet}}]\cr
#' Parameter set to tune over. Default is \code{\link{autoxgbparset}}.
#' Can be updated using `.$set_parset()`.
#' @param nthread [integer(1)]\cr
#' Number of cores to use.
#' If \code{NULL} (default), xgboost will determine internally how many cores to use.
#' Can be set using `.$set_nthread()`.
#'
#' Arguments to `.$fit()`:
#' @param task [\code{\link[mlr]{Task}}]\cr
#' The task to be trained.
#' @param iterations [\code{integer(1L}]\cr
#' Number of MBO iterations to do. Will be ignored if custom \code{control} is used.
#' Default is \code{160}.
#' @param time.budget [\code{integer(1L}]\cr
#' Time that can be used for tuning (in seconds). Will be ignored if a custom \code{control} is used.
#' Default is \code{3600}, i.e., one hour.
#' @param build.final.model [\code{logical(1)}]\cr
#' Should the model with the best found configuration be refitted on the complete dataset?
#' Default is \code{FALSE}.
#' @param control [\code{\link[mlrMBO]{MBOControl}}]\cr
#' Control object for optimizer.
#' If not specified, the default \code{\link[mlrMBO]{makeMBOControl}}] object will be used with
#' \code{iterations} maximum iterations and a maximum runtime of \code{time.budget} seconds.
#'
#' Additional arguments that control the process:
#' @param mbo.learner [\code{\link[mlr]{Learner}}]\cr
#' Regression learner from mlr, which is used as a surrogate to model our fitness function.
#' If \code{NULL} (default), the default learner is determined as described here:
#' \link[mlrMBO]{mbo_default_learner}.
#' Can be set using `.$set_mbo_learner()`.
#' @param design.size [\code{integer(1)}]\cr
#' Size of the initial design. Default is \code{15L}.
#' Can be set via `.$set_design_size()`
#' @param max.nrounds [\code{integer(1)}]\cr
#' Maximum number of allowed boosting iterations. Default is \code{10^6}.
#' Can be set via `.$set_max_nrounds()`.
#' @param early.stopping.rounds [\code{integer(1L}]\cr
#' After how many iterations without an improvement in the boosting OOB error should be stopped?
#' Default is \code{10}.
#' Can be set via `.$set_early_stopping_rounds()`.
#' @param early.stopping.fraction [\code{numeric(1)}]\cr
#' What fraction of the data should be used for early stopping (i.e. as a validation set).
#' Default is \code{4/5}.
#' Can be set via `.$set_early_stopping_fraction()`.
#' @param impact.encoding.boundary [\code{integer(1)}]\cr
#' Defines the threshold on how factor variables are handled. Factors with more levels than the \code{"impact.encoding.boundary"} get impact encoded while factor variables with less or equal levels than the \code{"impact.encoding.boundary"} get dummy encoded.
#' For \code{impact.encoding.boundary = 0L}, all factor variables get impact encoded while for \code{impact.encoding.boundary = .Machine$integer.max}, all of them get dummy encoded.
#' Default is \code{10}.
#' Can be set via `.$set_impact_encoding_boundary()`.
#' @param tune.threshold [logical(1)]\cr
#' Should thresholds be tuned? This has only an effect for classification, see \code{\link[mlr]{tuneThreshold}}.
#' Default is \code{TRUE}.
#' Can be set via `.$set_tune_threshold()`.
#'
#' @export
#' @examples
#' \donttest{
#' iris.task = makeClassifTask(data = iris, target = "Species")
#' axgb = AutoxgboostMC$new(measure = auc)
#' axgb$fit(t, time.budget = 5L)
#' p = axgb$predict(iris.task)
#' }
AutoxgboostMC = R6::R6Class("AutoxgboostMC",
public = list(
measures = NULL,
control = NULL,
parset = NULL,
design.size = 15L,
mbo.learner = NULL,
iterations = NULL,
time.budget = NULL,
max.nrounds = 3*10^3L,
early.stopping.rounds = 20L,
early.stopping.fraction = 4/5,
impact.encoding.boundary = 10L,
tune.threshold = TRUE,
nthread = NULL,
resample_instance = NULL,
baselearner = NULL,
preproc_pipeline = NULL,
model = NULL,
obj_fun = NULL,
optim.result = NULL,
build.final.model = NULL,
initialize = function(measures = NULL, parset = NULL, nthread = NULL) {
assert_list(measures, types = "Measure", null.ok = TRUE)
assert_class(parset, "ParamSet", null.ok = TRUE)
# Set defaults
self$measures = coalesce(measures, list(getDefaultMeasure(task)))
self$parset = coalesce(parset, autoxgboostMC::autoxgbparset)
self$nthread = assert_integerish(nthread, lower = 1, len = 1L, null.ok = TRUE)
},
print = function(...) {
catf("AutoxgboostMC Learner")
catf("Trained: %s", ifelse(is.null(self$model), "no", "yes"))
if (!is.null(self$model)) print(self$model)
},
fit = function(task, iterations = 160L, time.budget = 3600L, build.final.model = TRUE, control = NULL) {
assert_class(task, "SupervisedTask")
self$iterations = assert_integerish(iterations)
self$time.budget = assert_integerish(time.budget)
self$build.final.model = assert_flag(build.final.model)
self$control = control
if (!is.null(self$model)) {
self$continue_fit()
} else {
self$baselearner = self$make_baselearner(task)
transf_tasks = self$build_transform_pipeline(task)
self$baselearner = setHyperPars(self$baselearner, early.stopping.data = transf_tasks$task.test)
self$obj_fun = self$make_objective_function(transf_tasks)
self$optim.result = self$optimize_pipeline_mbo()
}
lrn = self$build_final_learner()
mod = NULL
if(build.final.model) mod = train(lrn, task)
self$model = AutoxgbResult$new(
optim_result = self$optim.result,
final_learner = lrn,
final_model = mod,
measures = self$measures,
preproc_pipeline = self$preproc_pipeline
)
},
continue_fit = function(task, iterations = 160L, time.budget = 3600L, build.final.model = TRUE) {
},
predict = function(newdata) {
predict(self$model$final_model, newdata)
},
# AutoxgboostMC steps
make_baselearner = function(task) {
tt = getTaskType(task)
td = getTaskDesc(task)
req_prob_measure = sapply(self$measures, function(x) {
any(getMeasureProperties(x) == "req.prob")
})
pv = list()
if (!is.null(self$nthread))
pv$nthread = self$nthread
if (tt == "classif") {
predict.type = ifelse(any(req_prob_measure) | self$tune.threshold, "prob", "response")
if(length(td$class.levels) == 2) {
objective = "binary:logistic"
eval_metric = "error"
parset = c(self$parset, makeParamSet(makeNumericParam("scale_pos_weight", lower = -10, upper = 10, trafo = function(x) 2^x)))
} else {
objective = "multi:softprob"
eval_metric = "merror"
}
baselearner = makeLearner("classif.xgboost.earlystop", id = "classif.xgboost.earlystop",
predict.type = predict.type, eval_metric = eval_metric, objective = objective,
early_stopping_rounds = self$early.stopping.rounds, maximize = !self$early_stopping_measure$minimize,
max.nrounds = self$max.nrounds, par.vals = pv)
} else if (tt == "regr") {
predict.type = NULL
objective = "reg:linear"
eval_metric = "rmse"
baselearner = makeLearner("regr.xgboost.earlystop", id = "regr.xgboost.earlystop",
eval_metric = eval_metric, objective = objective, early_stopping_rounds = self$early.stopping.rounds,
maximize = !self$early_stopping_measure$minimize, max.nrounds = self$max.nrounds, par.vals = pv)
} else {
stop("Task must be regression or classification")
}
return(baselearner)
},
# Build pipeline
build_transform_pipeline = function(task) {
has.cat.feats = sum(getTaskDesc(task)$n.feat[c("factors", "ordered")]) > 0
self$preproc_pipeline = NULLCPO
#if (!is.null(task$feature.information$timestamps))
# preproc_pipeline %<>>% cpoExtractTimeStampInformation(affect.names = unlist(task$feature.information$timestamps))
if (has.cat.feats) {
self$preproc_pipeline %<>>% generateCatFeatPipeline(task, self$impact.encoding.boundary)
}
self$preproc_pipeline %<>>% cpoDropConstants()
# process data and apply pipeline
# split early stopping data
if (is.null(self$resample_instance))
self$resample_instance = makeResampleInstance(makeResampleDesc("Holdout", split = self$early.stopping.fraction), task)
task.test = subsetTask(task, self$resample_instance$test.inds[[1]])
task.train = subsetTask(task, self$resample_instance$train.inds[[1]])
task.train %<>>% self$preproc_pipeline
task.test %<>>% retrafo(task.train)
return(list(task.train = task.train, task.test = task.test))
},
make_objective_function = function(transf_tasks) {
is_thresholded_measure = sapply(self$measures, function(x) {
props = getMeasureProperties(x)
any(props == "req.truth") & !any(props == "req.prob")
})
if (!any(is_thresholded_measure) & self$tune.threshold) {
warning("Threshold tuning is active, but no measure for tuning thresholds!
Skipping threshold tuning!")
self$tune.threshold = FALSE
}
smoof::makeMultiObjectiveFunction(name = "optimizeWrapperMultiCrit",
fn = function(x) {
x = x[!vlapply(x, is.na)]
lrn = setHyperPars(self$baselearner, par.vals = x)
mod = train(lrn, transf_tasks$task.train)
pred = predict(mod, transf_tasks$task.test)
nrounds = self$get_best_iteration(mod)
# For now we tune threshold of first applicable measure.
if (self$tune.threshold && getTaskType(transf_tasks$task.train) == "classif") {
tune.res = tuneThreshold(pred = pred, measure = self$measures[is_thresholded_measure][[1]])
if (length(self$measures[-which(is_thresholded_measure)[1]]) > 0) {
res = performance(pred, self$measures[-which(is_thresholded_measure)[1]], model = mod, task = transf_tasks$task.test)
res = c(res, tune.res$perf)
} else {
res = tune.res$perf
}
attr(res, "extras") = list(nrounds = nrounds, .threshold = tune.res$th)
} else {
res = performance(pred, self$measures, model = mod, task = transf_tasks$task.test)
attr(res, "extras") = list(nrounds = nrounds)
}
return(res)
},
par.set = self$parset, noisy = FALSE, has.simple.signature = FALSE, minimize = sapply(self$measures, function(x) x$minimize),
n.objectives = length(self$measures)
)
},
optimize_pipeline_mbo = function() {
assert_class(self$control, "MBOControl", null.ok = TRUE)
# Set defaults
if (is.null(self$control)) {
measures_ids = sapply(self$measures, function(x) x$id)
ctrl = makeMBOControl(n.objectives = length(self$measures), y.name = measures_ids)
self$control = setMBOControlTermination(ctrl, iters = self$iterations,
time.budget = self$time.budget)
if (self$is_multicrit) {
self$control = setMBOControlMultiObj(self$control, method = "dib",dib.indicator = "eps")
self$control = setMBOControlInfill(self$control, crit = makeMBOInfillCritDIB(cb.lambda = 2L))
}
}
des = generateDesign(n = self$design.size, self$parset)
optim.result = mbo(fun = self$obj_fun, control = self$control, design = des, learner = self$mbo.learner)
if (self$is_multicrit) {
# Fill best.ind, x and y using "best on early stopping measure".
optim.result$best.ind = self$get_best_ind(optim.result)
pars = names(optim.result$opt.path$par.set$pars)
optim.result$x = as.list(optim.result$opt.path$env$path[optim.result$best.ind, pars])
optim.result$y = as.list(optim.result$opt.path$env$path[optim.result$best.ind, self$measure_ids])
}
return(optim.result)
},
build_final_learner = function() {
nrounds = self$get_best_from_opt("nrounds")
pars = trafoValue(self$parset, self$optim.result$x)
pars = pars[!vlapply(pars, is.na)]
lrn = if (!is.null(self$baselearner$predict.type)) {
makeLearner("classif.xgboost.custom", nrounds = nrounds,
objective = self$baselearner$par.vals$objective,
predict.type = self$baselearner$predict.type,
predict.threshold = self$get_best_from_opt(".threshold"))
} else {
makeLearner("regr.xgboost.custom", nrounds = nrounds, objective = objective)
}
lrn = setHyperPars2(lrn, par.vals = pars)
lrn = self$preproc_pipeline %>>% lrn
#FIXME mlrCPO #39
#lrn$properties = c(lrn$properties, "weights")
return(lrn)
},
# Setters for various hyperparameters
set_max_nrounds = function(value) {
self$max.nrounds = assert_integerish(value, lower = 1L, len = 1L)
},
set_early_stopping_rounds = function(value) {
self$early.stopping.rounds = assert_integerish(value, lower = 1L, len = 1L)
},
set_early_stopping_fraction = function(value) {
self$early.stopping.fraction = assert_numeric(early.stopping.fraction, lower = 0, upper = 1, len = 1L)
},
set_design_size = function(value) {
self$design.size = assert_integerish(design.size, lower = 1L, len = 1L)
},
set_tune_threshold = function(value) {
self$tune.threshold = assert_flag(value)
},
set_impact_encoding_boundary = function(value) {
self$impact_encoding_boundary = assert_integerish(value, lower = 0, len = 1L)
},
set_nthread = function(value) {
self$nthread = assert_integerish(value, lower = 1, len = 1L, null.ok = TRUE)
},
set_measures = function(value) {
self$measures = assert_list(value, types = "Measure", null.ok = TRUE)
},
set_parset = function(value) {
self$parset = assert_class(value, "ParamSet", null.ok = TRUE)
},
set_resample_instance = function(value) {
self$resample_instance = assert_class(value, "ResampleInstance", null.ok = TRUE)
},
get_best_ind = function(optim.result) {
# We can either select a best index from the outside
if (!is.null(self$selected.ind)) best.ind = selected.ind
# Or auto-choose one according to the early stopping measure
if (is.null(optim.result$best.ind)) {
if (self$early_stopping_measure$minimize) {
best.ind = which.min(optim.result$opt.path$env$path[[self$early_stopping_measure$id]])
} else {
best.ind = which.max(optim.result$opt.path$env$path[[self$early_stopping_measure$id]])
}
} else {
best.ind = optim.result$best.ind
}
return(best.ind)
},
#Get best value from optimization result
# @param what [`character(1)`]: "nrounds" or ".threshold"
get_best_from_opt = function(what) {
self$optim.result$opt.path$env$extra[[self$optim.result$best.ind]][[what]]
},
# Get the iteration parameter of a fitted xboost model with early stopping
get_best_iteration = function(mod) {
getLearnerModel(mod, more.unwrap = TRUE)$best_iteration
},
plot_pareto_front = function() {self$model$plot_pareto_front()},
plot_results = function() {self$model$plot_results()}
),
active = list(
early_stopping_measure = function(value) {
if (missing(value)) {
self$measures[[1]]
} else {
measure_ids = sapply(self$measures, function(x) x$id)
assert_list(value, types = "Measure", null.ok = TRUE)
self$measures = c(value, self$measures[-which(value$id == measure_ids)])
messagef("Setting %s as early stopping measure!", value$id)
}
},
is_multicrit = function() {
length(self$measures) > 1
},
measure_ids = function() {
sapply(self$measures, function(x) x$id)
}
)
)
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.