Expression handling

DESCRIPTION

@@ -33,7 +33,7 @@ Encoding: UTF-8
 Roxygen: list(wrap = FALSE)
 Depends:
     R (>= 3.0.2),
-    ParamHelpers (>= 1.8),
+    ParamHelpers (>= 1.9)
 Imports:
     BBmisc (>= 1.10),
     ggplot2,

NAMESPACE

@@ -13,6 +13,7 @@ S3method(downsample,Task)
 S3method(estimateRelativeOverfitting,ResampleDesc)
 S3method(estimateResidualVariance,Learner)
 S3method(estimateResidualVariance,WrappedModel)
+S3method(evaluateParamExpressions,Learner)
 S3method(generateCalibrationData,BenchmarkResult)
 S3method(generateCalibrationData,Prediction)
 S3method(generateCalibrationData,ResampleResult)
@@ -86,6 +87,7 @@ S3method(getTaskTargetNames,TaskDescUnsupervised)
 S3method(getTaskTargets,CostSensTask)
 S3method(getTaskTargets,SupervisedTask)
 S3method(getTaskTargets,UnsupervisedTask)
+S3method(hasExpression,Learner)
 S3method(impute,Task)
 S3method(impute,data.frame)
 S3method(isFailureModel,BaseWrapperModel)
@@ -825,6 +827,7 @@ export(getTaskClassLevels)
 export(getTaskCosts)
 export(getTaskData)
 export(getTaskDescription)
+export(getTaskDictionary)
 export(getTaskFeatureNames)
 export(getTaskFormula)
 export(getTaskId)

R/Learner_properties.R

@@ -76,3 +76,10 @@ getSupportedLearnerProperties = function(type = NA_character_) {
     p[[type]]
 }
 
+#' @param obj [\code{\link{Learner}} | \code{character(1)}]\cr
+#' Same as \code{learner} above.
+#' @rdname LearnerProperties
+#' @export
+hasExpression.Learner = function(obj) {
+  any(hasExpression(obj$par.set)) || any(vlapply(obj$par.vals, is.expression))
+}

R/RLearner_classif_randomForest.R

@@ -5,10 +5,10 @@ makeRLearner.classif.randomForest = function() {
     package = "randomForest",
     par.set = makeParamSet(
       makeIntegerLearnerParam(id = "ntree", default = 500L, lower = 1L),
-      makeIntegerLearnerParam(id = "mtry", lower = 1L),
+      makeIntegerLearnerParam(id = "mtry", lower = 1L, default = expression(floor(sqrt(p)))),
       makeLogicalLearnerParam(id = "replace", default = TRUE),
-      makeNumericVectorLearnerParam(id = "classwt", lower = 0),
-      makeNumericVectorLearnerParam(id = "cutoff", lower = 0, upper = 1),
+      makeNumericVectorLearnerParam(id = "classwt", lower = 0, len = expression(k)),
+      makeNumericVectorLearnerParam(id = "cutoff", lower = 0, upper = 1, len = expression(k)),
       makeUntypedLearnerParam(id = "strata", tunable = FALSE),
       makeIntegerVectorLearnerParam(id = "sampsize", lower = 1L),
       makeIntegerLearnerParam(id = "nodesize", default = 1L, lower = 1L),

R/Task_operators.R

@@ -254,7 +254,7 @@ getTaskData = function(task, subset, features, target.extra = FALSE, recode.targ
   checkTask(task, "Task")
 
   if (missing(subset)) {
-    subset = NULL 
+    subset = NULL
   } else {
     assert(checkIntegerish(subset), checkLogical(subset))
     if (is.logical(subset))
@@ -266,7 +266,7 @@ getTaskData = function(task, subset, features, target.extra = FALSE, recode.targ
   assertLogical(target.extra)
 
   task.features = getTaskFeatureNames(task)
-  
+
   # if supplied check if the input is right and always convert 'features'
   # to character vec
   if (!missing(features)) {
@@ -453,3 +453,31 @@ getTaskFactorLevels = function(task) {
 getTaskWeights = function(task) {
   task$weights
 }
+
+
+#' @title Create a dictionary based on the task.
+#'
+#' @description Returns a dictionary, which contains the \link{Task} itself
+#' (\code{task}), the number of features (\code{p}), the number of
+#' observations (\code{n}), the task type (\code{type}) and in case of
+#' classification tasks, the number of class levels (\code{k}).
+#'
+#' @template arg_task
+#' @return [\code{\link[base]{list}}]. Used for evaluating the expressions
+#' within a parameter, parameter set or list of parameters.
+#' @family task
+#' @export
+#' @examples
+#' task = makeClassifTask(data = iris, target = "Species")
+#' getTaskDictionary(task)
+getTaskDictionary = function(task) {
+  dict = list(
+    task = task,
+    p = getTaskNFeats(task),
+    n = getTaskSize(task),
+    type = getTaskType(task)
+  )
+  if (dict$type == "classif")
+    dict$k = length(getTaskClassLevels(task))
+  return(dict)
+}

R/evaluateParamExpressions.R

@@ -0,0 +1,77 @@
+#' @title Evaluates expressions within a learner or parameter set.
+#'
+#' @description
+#' A \code{\link{Learner}} can contain unevaluated \code{\link[base]{expression}s}
+#' as value for a hyperparameter. E.g., these expressions are used if the default
+#' value depends on the task size or an upper limit for a parameter is given by
+#' the number of features in a task. \code{evaluateParamExpressions} allows to
+#' evaluate these expressions using a given dictionary, which holds the following
+#' information:
+#' \itemize{
+#'   \item{\code{task}:} the task itself, allowing to access any of its elements.
+#'   \item{\code{p}:} the number of features in the task
+#'   \item{\code{n}:} the number of observations in the task
+#'   \item{\code{type}:} the task type, i.e. "classif", "regr", "surv", "cluster", "costcens" or "multilabel"
+#'   \item{\code{k}:} the number of classes of the target variable (only available for classification tasks)
+#' }
+#' Usually the evaluation of the expression is performed automatically, e.g. in
+#' \code{\link{train}} or \code{\link{tuneParams}}. Therefore calling
+#' \code{evaluateParamExpressions} manually should not be necessary.
+#' It is also possible to directly evaluate the expressions of a
+#' \code{\link[ParamHelpers]{ParamSet}}, \code{\link[base]{list}} of
+#' \code{\link[ParamHelpers]{Param}s} or single \code{\link[ParamHelpers]{Param}s}.
+#' For further information on these, please refer to the documentation of the
+#' \code{ParamHelpers} package.
+#'
+#' @param obj [\code{\link{Learner}}]\cr
+#'   The learner. If you pass a string the learner will be created via
+#'   \code{\link{makeLearner}}. Expressions within \code{length}, \code{lower}
+#'   or \code{upper} boundaries, \code{default} or \code{value} will be
+#'   evaluated using the provided dictionary (\code{dict}).
+#' @param dict [\code{environment} | \code{list} | \code{NULL}]\cr
+#'   Environment or list which will be used for evaluating the variables
+#'   of expressions within a parameter, parameter set or list of parameters.
+#'   The default is \code{NULL}.
+#' @return [\code{\link{Learner}}].
+#' @export
+#' @examples
+#' ## (1) evaluation of a learner's hyperparameters
+#' task = makeClassifTask(data = iris, target = "Species")
+#' dict = getTaskDictionary(task = task)
+#' lrn1 = makeLearner("classif.rpart", minsplit = expression(k * p),
+#'   minbucket = expression(3L + 4L * task$task.desc$has.blocking))
+#' lrn2 = evaluateParamExpressions(obj = lrn1, dict = dict)
+#'
+#' getHyperPars(lrn1)
+#' getHyperPars(lrn2)
+#'
+#' ## (2) evaluation of a learner's entire parameter set
+#' task = makeClassifTask(data = iris, target = "Species")
+#' dict = getTaskDictionary(task = task)
+#' lrn1 = makeLearner("classif.randomForest")
+#' lrn2 = evaluateParamExpressions(obj = lrn1, dict = dict)
+#'
+#' ## Note the values for parameters 'mtry', 'classwt' and 'cutoff'
+#' lrn1$par.set
+#' lrn2$par.set
+#'
+#' ## (3) evaluation of a parameter set
+#' task = makeClassifTask(data = iris, target = "Species")
+#' dict = getTaskDictionary(task = task)
+#' ps1 = makeParamSet(
+#'   makeNumericParam("C", lower = expression(k), upper = expression(n), trafo = function(x) 2^x),
+#'   makeDiscreteParam("sigma", values = expression(list(k, p)))
+#' )
+#' ps2 = evaluateParamExpressions(obj = ps1, dict = dict)
+#'
+#' ps1
+#' ps2
+evaluateParamExpressions.Learner = function(obj, dict = NULL) {
+  obj = checkLearner(obj)
+  if (hasExpression(obj)) {
+    assertList(dict, null.ok = TRUE)
+    obj$par.set = evaluateParamExpressions(obj = obj$par.set, dict = dict)
+    obj$par.vals = evaluateParamExpressions(obj = obj$par.vals, dict = dict)
+  }
+  return(obj)
+}

R/makeLearner.R

@@ -43,10 +43,12 @@
 #' @return [\code{\link{Learner}}].
 #' @family learner
 #' @export
+#' @note Learners can contain task dependent expressions, see \code{\link{evaluateParamExpressions}} for more information.
 #' @aliases Learner
 #' @examples
 #' makeLearner("classif.rpart")
 #' makeLearner("classif.lda", predict.type = "prob")
+#' makeLearner("classif.rpart", minsplit = expression(k))
 #' lrn = makeLearner("classif.lda", method = "t", nu = 10)
 #' print(lrn$par.vals)
 makeLearner = function(cl, id = cl, predict.type = "response", predict.threshold = NULL,

R/setHyperPars.R

@@ -11,14 +11,16 @@
 #' @note If a named (hyper)parameter can't be found for the given learner, the 3
 #' closest (hyper)parameter names will be output in case the user mistyped.
 #' @export
+#' @note Learners can contain task dependent expressions, see \code{\link{evaluateParamExpressions}} for more information.
 #' @family learner
 #' @importFrom utils adist
 #' @examples
 #' cl1 = makeLearner("classif.ksvm", sigma = 1)
 #' cl2 = setHyperPars(cl1, sigma = 10, par.vals = list(C = 2))
+#' cl3 = setHyperPars(cl2, C = expression(round(n / p)))
 #' print(cl1)
-#' # note the now set and altered hyperparameters:
 #' print(cl2)
+#' print(cl3)
 setHyperPars = function(learner, ..., par.vals = list()) {
   args = list(...)
   assertClass(learner, classes = "Learner")
@@ -74,7 +76,7 @@ setHyperPars2.Learner = function(learner, par.vals) {
       learner$par.set$pars[[n]] = makeUntypedLearnerParam(id = n)
       learner$par.vals[[n]] = p
     } else {
-      if (on.par.out.of.bounds != "quiet" && !isFeasible(pd, p)) {
+      if (on.par.out.of.bounds != "quiet" && !isFeasible(pd, p) && !is.expression(p)) {
         msg = sprintf("%s is not feasible for parameter '%s'!", convertToShortString(p), pd$id)
         if (on.par.out.of.bounds == "stop") {
           stop(msg)

R/train.R

@@ -31,6 +31,10 @@
 train = function(learner, task, subset, weights = NULL) {
   learner = checkLearner(learner)
   assertClass(task, classes = "Task")
+  if (hasExpression(learner)) {
+    dict = getTaskDictionary(task = task)
+    learner = evaluateParamExpressions(obj = learner, dict = dict)
+  }
   if (missing(subset)) {
     subset = seq_len(getTaskSize(task))
   } else {

R/tuneParams.R

@@ -22,15 +22,18 @@
 #' @param par.set [\code{\link[ParamHelpers]{ParamSet}}]\cr
 #'   Collection of parameters and their constraints for optimization.
 #'   Dependent parameters with a \code{requires} field must use \code{quote} and not
-#'   \code{expression} to define it.
+#'   \code{expression} to define it. On the other hand, task dependent parameters
+#'   need to be defined with expressions.
 #' @param control [\code{\link{TuneControl}}]\cr
 #'   Control object for search method. Also selects the optimization algorithm for tuning.
 #' @template arg_showinfo
 #' @return [\code{\link{TuneResult}}].
 #' @family tune
 #' @note If you would like to include results from the training data set, make
 #' sure to appropriately adjust the resampling strategy and the aggregation for
-#' the measure. See example code below.
+#' the measure. See example code below.\cr
+#' Note that learners and parameter sets can contain task dependent
+#' expressions, see \code{\link{evaluateParamExpressions}} for more information.
 #' @export
 #' @examples
 #' # a grid search for an SVM (with a tiny number of points...)
@@ -50,6 +53,16 @@
 #' print(head(generateHyperParsEffectData(res)))
 #' print(head(generateHyperParsEffectData(res, trafo = TRUE)))
 #'
+#' # tuning the parameters 'C' and 'sigma' of an SVM, where the boundaries
+#' # of 'sigma' depend on the number of features
+#' ps = makeParamSet(
+#'   makeNumericLearnerParam("sigma", lower = expression(0.2 * p), upper = expression(2.5 * p)),
+#'   makeDiscreteLearnerParam("C", values = 2^c(-1, 1))
+#' )
+#' rdesc = makeResampleDesc("Subsample")
+#' ctrl = makeTuneControlRandom(maxit = 2L)
+#' res = tuneParams("classif.ksvm", iris.task, par.set = ps, control = ctrl, resampling = rdesc)
+#'
 #' \dontrun{
 #' # we optimize the SVM over 3 kernels simultanously
 #' # note how we use dependent params (requires = ...) and iterated F-racing here
@@ -79,8 +92,14 @@
 tuneParams = function(learner, task, resampling, measures, par.set, control, show.info = getMlrOption("show.info")) {
   learner = checkLearner(learner)
   assertClass(task, classes = "Task")
+  dict = getTaskDictionary(task = task)
+  if (hasExpression(learner)) {
+    learner = evaluateParamExpressions(obj = learner, dict = dict)
+  }
   measures = checkMeasures(measures, learner)
   assertClass(par.set, classes = "ParamSet")
+  if (hasExpression(par.set))
+    par.set = evaluateParamExpressions(obj = par.set, dict = dict)
   assertClass(control, classes = "TuneControl")
   if (!inherits(resampling, "ResampleDesc") &&  !inherits(resampling, "ResampleInstance"))
     stop("Argument resampling must be of class ResampleDesc or ResampleInstance!")