From afa676609daba103e43d6d4654560ca4c1c9b38b Mon Sep 17 00:00:00 2001
From: Benjamin Auder <benjamin.auder@somewhere>
Date: Thu, 17 Jun 2021 03:19:40 +0200
Subject: [PATCH] Reorganize code - unfinished: some functions not exported yet

---
 DESCRIPTION     |   3 ++
 R/R6_AgghooCV.R |  86 ++++++-----------------------
 R/agghoo.R      |  43 ++++-----------
 R/checks.R      | 100 ++++++++++++++++++++++++++++++++++
 R/compareTo.R   | 141 ++++++++++++++----------------------------------
 R/utils.R       |  28 ++++++++++
 man/AgghooCV.Rd |  17 +++---
 man/agghoo.Rd   |   5 +-
 8 files changed, 207 insertions(+), 216 deletions(-)
 create mode 100644 R/checks.R
 create mode 100644 R/utils.R

diff --git a/DESCRIPTION b/DESCRIPTION
index 2689a20..c47391f 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -32,3 +32,6 @@ Collate:
     'R6_AgghooCV.R'
     'R6_Model.R'
     'A_NAMESPACE.R'
+    'checks.R'
+    'compareTo.R'
+    'utils.R'
diff --git a/R/R6_AgghooCV.R b/R/R6_AgghooCV.R
index ed9aa5c..485c678 100644
--- a/R/R6_AgghooCV.R
+++ b/R/R6_AgghooCV.R
@@ -15,13 +15,11 @@ AgghooCV <- R6::R6Class("AgghooCV",
     #' @param task "regression" or "classification"
     #' @param gmodel Generic model returning a predictive function
     #' @param loss Function assessing the error of a prediction
-    initialize = function(data, target, task, gmodel, loss = NULL) {
+    initialize = function(data, target, task, gmodel, loss) {
       private$data <- data
       private$target <- target
       private$task <- task
       private$gmodel <- gmodel
-      if (is.null(loss))
-        loss <- private$defaultLoss
       private$loss <- loss
     },
     #' @description Fit an agghoo model.
@@ -31,15 +29,10 @@ AgghooCV <- R6::R6Class("AgghooCV",
     #'          - test_size: percentage of data in the test dataset, for MC
     #'            (irrelevant for V-fold). Default: 0.2. \cr
     #'          - shuffle: wether or not to shuffle data before V-fold.
-    #'            Irrelevant for Monte-Carlo; default: TRUE
-    fit = function(
-      CV = list(type = "MC",
-                V = 10,
-                test_size = 0.2,
-                shuffle = TRUE)
-    ) {
-      if (!is.list(CV))
-        stop("CV: list of type, V, [test_size], [shuffle]")
+    #'            Irrelevant for Monte-Carlo; default: TRUE \cr
+    #'        Default (if NULL): type="MC", V=10, test_size=0.2
+    fit = function(CV = NULL) {
+      CV <- checkCV(CV)
       n <- nrow(private$data)
       shuffle_inds <- NULL
       if (CV$type == "vfold" && CV$shuffle)
@@ -49,22 +42,14 @@ AgghooCV <- R6::R6Class("AgghooCV",
       for (v in seq_len(CV$V)) {
         # Prepare train / test data and target, from full dataset.
         # dataHO: "data Hold-Out" etc.
-        test_indices <- private$get_testIndices(CV, v, n, shuffle_inds)
-        dataHO <- private$data[-test_indices,]
-        testX <- private$data[test_indices,]
-        targetHO <- private$target[-test_indices]
-        testY <- private$target[test_indices]
-        # [HACK] R will cast 1-dim matrices into vectors:
-        if (!is.matrix(dataHO) && !is.data.frame(dataHO))
-          dataHO <- as.matrix(dataHO)
-        if (!is.matrix(testX) && !is.data.frame(testX))
-          testX <- as.matrix(testX)
+        test_indices <- get_testIndices(n, CV, v, shuffle_inds)
+        d <- splitTrainTest(private$data, private$target, test_indices)
         best_model <- NULL
         best_error <- Inf
         for (p in seq_len(private$gmodel$nmodels)) {
-          model_pred <- private$gmodel$get(dataHO, targetHO, p)
-          prediction <- model_pred(testX)
-          error <- private$loss(prediction, testY)
+          model_pred <- private$gmodel$get(d$dataTrain, d$targetTrain, p)
+          prediction <- model_pred(d$dataTest)
+          error <- private$loss(prediction, d$targetTest)
           if (error <= best_error) {
             newModel <- list(model=model_pred, param=private$gmodel$getParam(p))
             if (error == best_error)
@@ -89,7 +74,10 @@ AgghooCV <- R6::R6Class("AgghooCV",
         return (invisible(NULL))
       }
       V <- length(private$pmodels)
-      oneLineX <- t(as.matrix(X[1,]))
+      oneLineX <- X[1,]
+      if (is.matrix(X))
+        # HACK: R behaves differently with data frames and matrices.
+        oneLineX <- t(as.matrix(oneLineX))
       if (length(private$pmodels[[1]]$model(oneLineX)) >= 2)
         # Soft classification:
         return (Reduce("+", lapply(private$pmodels, function(m) m$model(X))) / V)
@@ -119,50 +107,6 @@ AgghooCV <- R6::R6Class("AgghooCV",
     task = NULL,
     gmodel = NULL,
     loss = NULL,
-    pmodels = NULL,
-    get_testIndices = function(CV, v, n, shuffle_inds) {
-      if (CV$type == "vfold") {
-        # Slice indices (optionnally shuffled)
-        first_index = round((v-1) * n / CV$V) + 1
-        last_index = round(v * n / CV$V)
-        test_indices = first_index:last_index
-        if (!is.null(shuffle_inds))
-          test_indices <- shuffle_inds[test_indices]
-      }
-      else
-        # Monte-Carlo cross-validation
-        test_indices = sample(n, round(n * CV$test_size))
-      test_indices
-    },
-    defaultLoss = function(y1, y2) {
-      if (private$task == "classification") {
-        if (is.null(dim(y1)))
-          # Standard case: "hard" classification
-          mean(y1 != y2)
-        else {
-          # "Soft" classification: predict() outputs a probability matrix
-          # In this case "target" could be in matrix form.
-          if (!is.null(dim(y2)))
-            mean(rowSums(abs(y1 - y2)))
-          else {
-            # Or not: y2 is a "factor".
-            y2 <- as.character(y2)
-            # NOTE: the user should provide target in matrix form because
-            # matching y2 with columns is rather inefficient!
-            names <- colnames(y1)
-            positions <- list()
-            for (idx in seq_along(names))
-              positions[[ names[idx] ]] <- idx
-            mean(vapply(
-              seq_along(y2),
-              function(idx) sum(abs(y1[idx,] - positions[[ y2[idx] ]])),
-              0))
-          }
-        }
-      }
-      else
-        # Regression
-        mean(abs(y1 - y2))
-    }
+    pmodels = NULL
   )
 )
diff --git a/R/agghoo.R b/R/agghoo.R
index c8765fc..48ac741 100644
--- a/R/agghoo.R
+++ b/R/agghoo.R
@@ -41,43 +41,18 @@
 #' Journal of Machine Learning Research 22(20):1--55, 2021.
 #'
 #' @export
-agghoo <- function(data, target, task = NULL, gmodel = NULL, params = NULL, loss = NULL) {
+agghoo <- function(
+  data, target, task = NULL, gmodel = NULL, params = NULL, loss = NULL
+) {
 	# Args check:
-  if (!is.data.frame(data) && !is.matrix(data))
-    stop("data: data.frame or matrix")
-  if (is.data.frame(target) || is.matrix(target)) {
-    if (nrow(target) != nrow(data) || ncol(target) == 1)
-      stop("target probability matrix does not match data size")
-  }
-  else if (!is.numeric(target) && !is.factor(target) && !is.character(target))
-    stop("target: numeric, factor or character vector")
-  if (!is.null(task))
-    task = match.arg(task, c("classification", "regression"))
-  if (is.character(gmodel))
-    gmodel <- match.arg(gmodel, c("knn", "ppr", "rf", "tree"))
-  else if (!is.null(gmodel) && !is.function(gmodel))
-    # No further checks here: fingers crossed :)
-    stop("gmodel: function(dataHO, targetHO, param) --> function(X) --> y")
-  if (is.numeric(params) || is.character(params))
-    params <- as.list(params)
-  if (!is.list(params) && !is.null(params))
-    stop("params: numerical, character, or list (passed to model)")
-  if (is.function(gmodel) && !is.list(params))
-    stop("params must be provided when using a custom model")
-  if (is.list(params) && is.null(gmodel))
-    stop("model (or family) must be provided when using custom params")
-  if (!is.null(loss) && !is.function(loss))
-    # No more checks here as well... TODO:?
-    stop("loss: function(y1, y2) --> Real")
+  checkDaTa(data, target)
+  task <- checkTask(task, target)
+  modPar <- checkModPar(gmodel, params)
+  loss <- checkLoss(loss, task)
 
-  if (is.null(task)) {
-    if (is.numeric(target))
-      task = "regression"
-    else
-      task = "classification"
-  }
   # Build Model object (= list of parameterized models)
-  model <- Model$new(data, target, task, gmodel, params)
+  model <- Model$new(data, target, task, modPar$gmodel, modPar$params)
+
   # Return AgghooCV object, to run and predict
   AgghooCV$new(data, target, task, model, loss)
 }
diff --git a/R/checks.R b/R/checks.R
new file mode 100644
index 0000000..e105dfa
--- /dev/null
+++ b/R/checks.R
@@ -0,0 +1,100 @@
+defaultLoss_classif <- function(y1, y2) {
+  if (is.null(dim(y1)))
+    # Standard case: "hard" classification
+    mean(y1 != y2)
+  else {
+    # "Soft" classification: predict() outputs a probability matrix
+    # In this case "target" could be in matrix form.
+    if (!is.null(dim(y2)))
+      mean(rowSums(abs(y1 - y2)))
+    else {
+      # Or not: y2 is a "factor".
+      y2 <- as.character(y2)
+      # NOTE: the user should provide target in matrix form because
+      # matching y2 with columns is rather inefficient!
+      names <- colnames(y1)
+      positions <- list()
+      for (idx in seq_along(names))
+        positions[[ names[idx] ]] <- idx
+      mean(vapply(
+        seq_along(y2),
+        function(idx) sum(abs(y1[idx,] - positions[[ y2[idx] ]])),
+        0))
+    }
+  }
+}
+
+defaultLoss_regress <- function(y1, y2) {
+  mean(abs(y1 - y2))
+}
+
+# TODO: allow strings like "MSE", "abs" etc
+checkLoss <- function(loss, task) {
+  if (!is.null(loss) && !is.function(loss))
+    stop("loss: function(y1, y2) --> Real")
+  if (is.null(loss)) {
+    loss <- if (task == "classification") {
+      defaultLoss_classif
+    } else {
+      defaultLoss_regress
+    }
+  }
+  loss
+}
+
+checkCV <- function(CV) {
+  if (is.null(CV))
+    CV <- list(type="MC", V=10, test_size=0.2, shuffle=TRUE)
+  else {
+    if (!is.list(CV))
+      stop("CV: list of type('MC'|'vfold'), V(integer, [test_size, shuffle]")
+    if (is.null(CV$type)) {
+      warning("CV$type not provided: set to MC")
+      CV$type <- "MC"
+    }
+    if (is.null(CV$V)) {
+      warning("CV$V not provided: set to 10")
+      CV$V <- 10
+    }
+    if (CV$type == "MC" && is.null(CV$test_size))
+      CV$test_size <- 0.2
+    if (CV$type == "vfold" && is.null(CV$shuffle))
+      CV$shuffle <- TRUE
+  }
+  CV
+}
+
+checkDaTa <- function(data, target) {
+  if (!is.data.frame(data) && !is.matrix(data))
+    stop("data: data.frame or matrix")
+  if (is.data.frame(target) || is.matrix(target)) {
+    if (!is.numeric(target))
+      stop("multi-columns target must be a probability matrix")
+    if (nrow(target) != nrow(data) || ncol(target) == 1)
+      stop("target probability matrix does not match data size")
+  }
+  else if (!is.numeric(target) && !is.factor(target) && !is.character(target))
+    stop("target: numeric, factor or character vector")
+}
+
+checkTask <- function(task, target) {
+  if (!is.null(task))
+    task <- match.arg(task, c("classification", "regression"))
+  task <- ifelse(is.numeric(target), "regression", "classification")
+}
+
+checkModPar <- function(gmodel, params) {
+  if (is.character(gmodel))
+    gmodel <- match.arg(gmodel, c("knn", "ppr", "rf", "tree"))
+  else if (!is.null(gmodel) && !is.function(gmodel))
+    stop("gmodel: function(dataHO, targetHO, param) --> function(X) --> y")
+  if (is.numeric(params) || is.character(params))
+    params <- as.list(params)
+  if (!is.list(params) && !is.null(params))
+    stop("params: numerical, character, or list (passed to model)")
+  if (is.function(gmodel) && !is.list(params))
+    stop("params must be provided when using a custom model")
+  if (is.list(params) && is.null(gmodel))
+    stop("model (or family) must be provided when using custom params")
+  list(gmodel=gmodel, params=params)
+}
diff --git a/R/compareTo.R b/R/compareTo.R
index aa3a4e8..00e90a9 100644
--- a/R/compareTo.R
+++ b/R/compareTo.R
@@ -1,86 +1,25 @@
-standardCV_core <- function(data, target, task = NULL, gmodel = NULL, params = NULL,
-  loss = NULL, CV = list(type = "MC", V = 10, test_size = 0.2, shuffle = TRUE)
-) {
-  if (!is.null(task))
-    task = match.arg(task, c("classification", "regression"))
-  if (is.character(gmodel))
-    gmodel <- match.arg(gmodel, c("knn", "ppr", "rf", "tree"))
-  if (is.numeric(params) || is.character(params))
-    params <- as.list(params)
-  if (is.null(task)) {
-    if (is.numeric(target))
-      task = "regression"
-    else
-      task = "classification"
-  }
-
-  if (is.null(loss)) {
-    loss <- function(y1, y2) {
-      if (task == "classification") {
-        if (is.null(dim(y1)))
-          mean(y1 != y2)
-        else {
-          if (!is.null(dim(y2)))
-            mean(rowSums(abs(y1 - y2)))
-          else {
-            y2 <- as.character(y2)
-            names <- colnames(y1)
-            positions <- list()
-            for (idx in seq_along(names))
-              positions[[ names[idx] ]] <- idx
-            mean(vapply(
-              seq_along(y2),
-              function(idx) sum(abs(y1[idx,] - positions[[ y2[idx] ]])),
-              0))
-          }
-        }
-      }
-      else
-        mean(abs(y1 - y2))
-    }
-  }
-
+standardCV_core <- function(data, target, task, gmodel, params, loss, CV) {
   n <- nrow(data)
   shuffle_inds <- NULL
   if (CV$type == "vfold" && CV$shuffle)
     shuffle_inds <- sample(n, n)
-  get_testIndices <- function(v, shuffle_inds) {
-    if (CV$type == "vfold") {
-      first_index = round((v-1) * n / CV$V) + 1
-      last_index = round(v * n / CV$V)
-      test_indices = first_index:last_index
-      if (!is.null(shuffle_inds))
-        test_indices <- shuffle_inds[test_indices]
-    }
-    else
-      test_indices = sample(n, round(n * CV$test_size))
-    test_indices
-  }
   list_testinds <- list()
   for (v in seq_len(CV$V))
-    list_testinds[[v]] <- get_testIndices(v, shuffle_inds)
-
+    list_testinds[[v]] <- get_testIndices(n, CV, v, shuffle_inds)
   gmodel <- agghoo::Model$new(data, target, task, gmodel, params)
   best_error <- Inf
   best_model <- NULL
   for (p in seq_len(gmodel$nmodels)) {
-    error <- 0
-    for (v in seq_len(CV$V)) {
+    error <- Reduce('+', lapply(seq_len(CV$V), function(v) {
       testIdx <- list_testinds[[v]]
-      dataHO <- data[-testIdx,]
-      testX <- data[testIdx,]
-      targetHO <- target[-testIdx]
-      testY <- target[testIdx]
-      if (!is.matrix(dataHO) && !is.data.frame(dataHO))
-        dataHO <- as.matrix(dataHO)
-      if (!is.matrix(testX) && !is.data.frame(testX))
-        testX <- as.matrix(testX)
-      model_pred <- gmodel$get(dataHO, targetHO, p)
-      prediction <- model_pred(testX)
-      error <- error + loss(prediction, testY)
-    }
+      d <- splitTrainTest(data, target, testIdx)
+      model_pred <- gmodel$get(d$dataTrain, d$targetTrain, p)
+      prediction <- model_pred(d$dataTest)
+      loss(prediction, d$targetTest)
+    }) )
     if (error <= best_error) {
-      newModel <- list(model=model_pred, param=gmodel$getParam(p))
+      newModel <- list(model=gmodel$get(data, target, p),
+                       param=gmodel$getParam(p))
       if (error == best_error)
         best_model[[length(best_model)+1]] <- newModel
       else {
@@ -89,24 +28,30 @@ standardCV_core <- function(data, target, task = NULL, gmodel = NULL, params = N
       }
     }
   }
+#browser()
   best_model[[ sample(length(best_model), 1) ]]
 }
 
 standardCV_run <- function(
-  dataTrain, dataTest, targetTrain, targetTest, verbose, CV, floss, ...
+  dataTrain, dataTest, targetTrain, targetTest, CV, floss, verbose, ...
 ) {
-  s <- standardCV_core(dataTrain, targetTrain, ...)
+  args <- list(...)
+  task <- checkTask(args$task, targetTrain)
+  modPar <- checkModPar(args$gmodel, args$params)
+  loss <- checkLoss(args$loss, task)
+  s <- standardCV_core(
+    dataTrain, targetTrain, task, modPar$gmodel, modPar$params, loss, CV)
   if (verbose)
     print(paste( "Parameter:", s$param ))
-  ps <- s$model(test)
-  err_s <- floss(ps, targetTest)
+  p <- s$model(dataTest)
+  err <- floss(p, targetTest)
   if (verbose)
-    print(paste("error CV:", err_s))
-  invisible(c(errors, err_s))
+    print(paste("error CV:", err))
+  invisible(err)
 }
 
 agghoo_run <- function(
-  dataTrain, dataTest, targetTrain, targetTest, verbose, CV, floss, ...
+  dataTrain, dataTest, targetTrain, targetTest, CV, floss, verbose, ...
 ) {
   a <- agghoo(dataTrain, targetTrain, ...)
   a$fit(CV)
@@ -118,27 +63,22 @@ agghoo_run <- function(
   err <- floss(pa, targetTest)
   if (verbose)
     print(paste("error agghoo:", err))
+  invisible(err)
 }
 
-# ... arguments passed to agghoo or any other procedure
+# ... arguments passed to method_s (agghoo, standard CV or else)
 compareTo <- function(
-  data, target, rseed=-1, verbose=TRUE, floss=NULL,
-  CV = list(type = "MC",
-            V = 10,
-            test_size = 0.2,
-            shuffle = TRUE),
-  method_s=NULL, ...
+  data, target, method_s, rseed=-1, floss=NULL, verbose=TRUE, ...
 ) {
   if (rseed >= 0)
     set.seed(rseed)
   n <- nrow(data)
   test_indices <- sample( n, round(n / ifelse(n >= 500, 10, 5)) )
-  trainData <- as.matrix(data[-test_indices,])
-  trainTarget <- target[-test_indices]
-  testData <- as.matrix(data[test_indices,])
-  testTarget <- target[test_indices]
+  d <- splitTrainTest(data, target, test_indices)
+  CV <- checkCV(list(...)$CV)
 
   # Set error function to be used on model outputs (not in core method)
+  task <- checkTask(list(...)$task, target)
   if (is.null(floss)) {
     floss <- function(y1, y2) {
       ifelse(task == "classification", mean(y1 != y2), mean(abs(y1 - y2)))
@@ -147,34 +87,33 @@ compareTo <- function(
 
   # Run (and compare) all methods:
   runOne <- function(o) {
-    o(dataTrain, dataTest, targetTrain, targetTest, verbose, CV, floss, ...)
+    o(d$dataTrain, d$dataTest, d$targetTrain, d$targetTest,
+      CV, floss, verbose, ...)
   }
+  errors <- c()
   if (is.list(method_s))
     errors <- sapply(method_s, runOne)
   else if (is.function(method_s))
     errors <- runOne(method_s)
-  else
-    errors <- c()
   invisible(errors)
 }
 
 # Run compareTo N times in parallel
+# ... : additional args to be passed to method_s
 compareMulti <- function(
-  data, target, N = 100, nc = NA,
-  CV = list(type = "MC",
-            V = 10,
-            test_size = 0.2,
-            shuffle = TRUE),
-  method_s=NULL, ...
+  data, target, method_s, N=100, nc=NA, floss=NULL, ...
 ) {
+  require(parallel)
   if (is.na(nc))
     nc <- parallel::detectCores()
+
+  # "One" comparison for each method in method_s (list)
   compareOne <- function(n) {
     print(n)
-    compareTo(data, target, n, verbose=FALSE, CV, method_s, ...)
+    compareTo(data, target, method_s, n, floss, verbose=FALSE, ...)
   }
+
   errors <- if (nc >= 2) {
-    require(parallel)
     parallel::mclapply(1:N, compareOne, mc.cores = nc)
   } else {
     lapply(1:N, compareOne)
@@ -182,5 +121,3 @@ compareMulti <- function(
   print("Errors:")
   Reduce('+', errors) / N
 }
-
-# TODO: unfinished !
diff --git a/R/utils.R b/R/utils.R
new file mode 100644
index 0000000..fa3a9df
--- /dev/null
+++ b/R/utils.R
@@ -0,0 +1,28 @@
+get_testIndices <- function(n, CV, v, shuffle_inds) {
+  if (CV$type == "vfold") {
+    # Slice indices (optionnally shuffled)
+    first_index = round((v-1) * n / CV$V) + 1
+    last_index = round(v * n / CV$V)
+    test_indices = first_index:last_index
+    if (!is.null(shuffle_inds))
+      test_indices <- shuffle_inds[test_indices]
+  }
+  else
+    # Monte-Carlo cross-validation
+    test_indices = sample(n, round(n * CV$test_size))
+  test_indices
+}
+
+splitTrainTest <- function(data, target, testIdx) {
+  dataTrain <- data[-testIdx,]
+  targetTrain <- target[-testIdx]
+  dataTest <- data[testIdx,]
+  targetTest <- target[testIdx]
+  # [HACK] R will cast 1-dim matrices into vectors:
+  if (!is.matrix(dataTrain) && !is.data.frame(dataTrain))
+    dataTrain <- as.matrix(dataTrain)
+  if (!is.matrix(dataTest) && !is.data.frame(dataTest))
+    dataTest <- as.matrix(dataTest)
+  list(dataTrain=dataTrain, targetTrain=targetTrain,
+       dataTest=dataTest, targetTest=targetTest)
+}
diff --git a/man/AgghooCV.Rd b/man/AgghooCV.Rd
index 75ce9db..5122236 100644
--- a/man/AgghooCV.Rd
+++ b/man/AgghooCV.Rd
@@ -23,7 +23,7 @@ from the list of models (see 'Model' class).
 \subsection{Method \code{new()}}{
 Create a new AgghooCV object.
 \subsection{Usage}{
-\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$new(data, target, task, gmodel, loss = NULL)}\if{html}{\out{</div>}}
+\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$new(data, target, task, gmodel, loss)}\if{html}{\out{</div>}}
 }
 
 \subsection{Arguments}{
@@ -48,19 +48,20 @@ Create a new AgghooCV object.
 \subsection{Method \code{fit()}}{
 Fit an agghoo model.
 \subsection{Usage}{
-\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$fit(CV = list(type = "MC", V = 10, test_size = 0.2, shuffle = TRUE))}\if{html}{\out{</div>}}
+\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$fit(CV = NULL)}\if{html}{\out{</div>}}
 }
 
 \subsection{Arguments}{
 \if{html}{\out{<div class="arguments">}}
 \describe{
 \item{\code{CV}}{List describing cross-validation to run. Slots: \cr
-- type: 'vfold' or 'MC' for Monte-Carlo (default: MC) \cr
-- V: number of runs (default: 10) \cr
-- test_size: percentage of data in the test dataset, for MC \cr
-  (irrelevant for V-fold). Default: 0.2.
-- shuffle: wether or not to shuffle data before V-fold.
-  Irrelevant for Monte-Carlo; default: TRUE}
+  - type: 'vfold' or 'MC' for Monte-Carlo (default: MC) \cr
+  - V: number of runs (default: 10) \cr
+  - test_size: percentage of data in the test dataset, for MC
+    (irrelevant for V-fold). Default: 0.2. \cr
+  - shuffle: wether or not to shuffle data before V-fold.
+    Irrelevant for Monte-Carlo; default: TRUE \cr
+Default (if NULL): type="MC", V=10, test_size=0.2}
 }
 \if{html}{\out{</div>}}
 }
diff --git a/man/agghoo.Rd b/man/agghoo.Rd
index 179d309..38730eb 100644
--- a/man/agghoo.Rd
+++ b/man/agghoo.Rd
@@ -33,7 +33,7 @@ loss(y1, y2) --> real number (error). Default: see R6::AgghooCV.}
 An R6::AgghooCV object o. Then, call o$fit() and finally o$predict(newData)
 }
 \description{
-Run the agghoo procedure (or standard cross-validation).
+Run the (core) agghoo procedure.
 Arguments specify the list of models, their parameters and the
 cross-validation settings, among others.
 }
@@ -52,3 +52,6 @@ pc <- a_cla$predict(iris[,-5] + rnorm(600, sd=0.1))
 Guillaume Maillard, Sylvain Arlot, Matthieu Lerasle. "Aggregated hold-out".
 Journal of Machine Learning Research 22(20):1--55, 2021.
 }
+\seealso{
+Function \code{\link{compareTo}}
+}
-- 
2.44.0