From: Benjamin Auder <benjamin.auder@somewhere>
Date: Fri, 11 Jun 2021 16:35:01 +0000 (+0200)
Subject: Refactoring: separate standard CV from agghoo + some bug fixes
X-Git-Url: https://git.auder.net/?a=commitdiff_plain;h=504afaadc783916dc126fb87ab9e067f302eb2c5;p=agghoo.git

Refactoring: separate standard CV from agghoo + some bug fixes
---

diff --git a/NAMESPACE b/NAMESPACE
index 63138fb..f0ea804 100644
--- a/NAMESPACE
+++ b/NAMESPACE
@@ -3,7 +3,6 @@
 export(AgghooCV)
 export(Model)
 export(agghoo)
-export(compareToStandard)
 importFrom(FNN,knn.reg)
 importFrom(R6,R6Class)
 importFrom(caret,var_seq)
diff --git a/R/R6_AgghooCV.R b/R/R6_AgghooCV.R
index dba42b4..9cdf19e 100644
--- a/R/R6_AgghooCV.R
+++ b/R/R6_AgghooCV.R
@@ -14,23 +14,15 @@ AgghooCV <- R6::R6Class("AgghooCV",
     #' @param target Vector of targets (generally numeric or factor)
     #' @param task "regression" or "classification"
     #' @param gmodel Generic model returning a predictive function
-    #' @param quality Function assessing the quality of a prediction;
-    #'                quality(y1, y2) --> real number
-    initialize = function(data, target, task, gmodel, quality = NULL) {
+    #' @param loss Function assessing the error of a prediction
+    initialize = function(data, target, task, gmodel, loss = NULL) {
       private$data <- data
       private$target <- target
       private$task <- task
       private$gmodel <- gmodel
-      if (is.null(quality)) {
-        quality <- function(y1, y2) {
-          # NOTE: if classif output is a probability matrix, adapt.
-          if (task == "classification")
-            mean(y1 == y2)
-          else
-            atan(1.0 / (mean(abs(y1 - y2) + 0.01))) #experimental...
-        }
-      }
-      private$quality <- quality
+      if (is.null(loss))
+        loss <- private$defaultLoss
+      private$loss <- loss
     },
     #' @description Fit an agghoo model.
     #' @param CV List describing cross-validation to run. Slots:
@@ -40,113 +32,84 @@ AgghooCV <- R6::R6Class("AgghooCV",
     #'            (irrelevant for V-fold). Default: 0.2.
     #'          - shuffle: wether or not to shuffle data before V-fold.
     #'            Irrelevant for Monte-Carlo; default: TRUE
-    #' @param mode "agghoo" or "standard" (for usual cross-validation)
     fit = function(
       CV = list(type = "MC",
                 V = 10,
                 test_size = 0.2,
-                shuffle = TRUE),
-      mode="agghoo"
+                shuffle = TRUE)
     ) {
       if (!is.list(CV))
         stop("CV: list of type, V, [test_size], [shuffle]")
       n <- nrow(private$data)
-      shuffle_inds <- NA
+      shuffle_inds <- NULL
       if (CV$type == "vfold" && CV$shuffle)
         shuffle_inds <- sample(n, n)
-      if (mode == "agghoo") {
-        vperfs <- list()
-        for (v in 1:CV$V) {
-          test_indices <- private$get_testIndices(CV, v, n, shuffle_inds)
-          vperf <- private$get_modelPerf(test_indices)
-          vperfs[[v]] <- vperf
-        }
-        private$run_res <- vperfs
-      }
-      else {
-        # Standard cross-validation
-        best_index = 0
-        best_perf <- -1
-        for (p in 1:private$gmodel$nmodels) {
-          tot_perf <- 0
-          for (v in 1:CV$V) {
-            test_indices <- private$get_testIndices(CV, v, n, shuffle_inds)
-            perf <- private$get_modelPerf(test_indices, p)
-            tot_perf <- tot_perf + perf / CV$V
-          }
-          if (tot_perf > best_perf) {
-            # TODO: if ex-aequos: models list + choose at random
-            best_index <- p
-            best_perf <- tot_perf
+      # Result: list of V predictive models (+ parameters for info)
+      private$pmodels <- list()
+      for (v in seq_len(CV$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$target[test_indices]
+        # [HACK] R will cast 1-dim matrices into vectors:
+        if (!is.matrix(dataHO) && !is.data.frame(dataHO))
+          dataHO <- as.matrix(dataHO)
+        if (!is.matrix(testX) && !is.data.frame(testX))
+          testX <- as.matrix(testX)
+        best_model <- NULL
+        best_error <- Inf
+        for (p in seq_len(private$gmodel$nmodels)) {
+          model_pred <- private$gmodel$get(dataHO, targetHO, p)
+          prediction <- model_pred(testX)
+          error <- private$loss(prediction, testY)
+          if (error <= best_error) {
+            newModel <- list(model=model_pred, param=private$gmodel$getParam(p))
+            if (error == best_error)
+              best_model[[length(best_model)+1]] <- newModel
+            else {
+              best_model <- list(newModel)
+              best_error <- error
+            }
           }
         }
-        best_model <- private$gmodel$get(private$data, private$target, best_index)
-        private$run_res <- list( list(model=best_model, perf=best_perf) )
+        # Choose a model at random in case of ex-aequos
+        private$pmodels[[v]] <- best_model[[ sample(length(best_model),1) ]]
       }
     },
     #' @description Predict an agghoo model (after calling fit())
     #' @param X Matrix or data.frame to predict
-    #' @param weight "uniform" (default) or "quality" to weight votes or
-    #'               average models performances (TODO: bad idea?!)
-    predict = function(X, weight="uniform") {
+    predict = function(X) {
       if (!is.matrix(X) && !is.data.frame(X))
         stop("X: matrix or data.frame")
-      if (!is.list(private$run_res)) {
+      if (!is.list(private$pmodels)) {
         print("Please call $fit() method first")
-        return
-      }
-      V <- length(private$run_res)
-      if (V == 1)
-        # Standard CV:
-        return (private$run_res[[1]]$model(X))
-      # Agghoo:
-      if (weight == "uniform")
-        weights <- rep(1 / V, V)
-      else {
-        perfs <- sapply(private$run_res, function(item) item$perf)
-        perfs[perfs < 0] <- 0 #TODO: show a warning (with count of < 0...)
-        total_weight <- sum(perfs) #TODO: error if total_weight == 0
-        weights <- perfs / total_weight
+        return (invisible(NULL))
       }
+      V <- length(private$pmodels)
+      if (length(private$pmodels[[1]]$model(X[1,])) >= 2)
+        # Soft classification:
+        return (Reduce("+", lapply(private$pmodels, function(m) m$model(X))) / V)
       n <- nrow(X)
-      # TODO: detect if output = probs matrix for classif (in this case, adapt?)
-      # prediction agghoo "probabiliste" pour un nouveau x :
-      # argMax({ predict(m_v, x), v in 1..V }) ...
-      if (private$task == "classification") {
-        votes <- as.list(rep(NA, n))
-        parse_numeric <- FALSE
-      }
-      else
-        preds <- matrix(0, nrow=n, ncol=V)
-      for (v in 1:V) {
-        predictions <- private$run_res[[v]]$model(X)
-        if (private$task == "regression")
-          preds <- cbind(preds, weights[v] * predictions)
-        else {
-          if (!parse_numeric && is.numeric(predictions))
-            parse_numeric <- TRUE
-          for (i in 1:n) {
-            if (!is.list(votes[[i]]))
-              votes[[i]] <- list()
-            index <- as.character(predictions[i])
-            if (is.null(votes[[i]][[index]]))
-              votes[[i]][[index]] <- 0
-            votes[[i]][[index]] <- votes[[i]][[index]] + weights[v]
-          }
-        }
-      }
+      all_predictions <- as.data.frame(matrix(nrow=n, ncol=V))
+      for (v in 1:V)
+        all_predictions[,v] <- private$pmodels[[v]]$model(X)
       if (private$task == "regression")
-        return (rowSums(preds))
-      res <- c()
-      for (i in 1:n) {
-        # TODO: if ex-aequos, random choice...
-        ind_max <- which.max(unlist(votes[[i]]))
-        pred_class <- names(votes[[i]])[ind_max]
-        if (parse_numeric)
-          pred_class <- as.numeric(pred_class)
-        res <- c(res, pred_class)
-      }
-      res
+        # Easy case: just average each row
+        rowSums(all_predictions)
+      # "Hard" classification:
+      apply(all_predictions, 1, function(row) {
+        t <- table(row)
+        # Next lines in case of ties (broken at random)
+        tmax <- max(t)
+        sample( names(t)[which(t == tmax)], 1 )
+      })
+    },
+    #' @description Return the list of V best parameters (after calling fit())
+    getParams = function() {
+      lapply(private$pmodels, function(m) m$param)
     }
   ),
   private = list(
@@ -154,51 +117,51 @@ AgghooCV <- R6::R6Class("AgghooCV",
     target = NULL,
     task = NULL,
     gmodel = NULL,
-    quality = NULL,
-    run_res = 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$V)
         test_indices = first_index:last_index
-        if (CV$shuffle)
+        if (!is.null(shuffle_inds))
           test_indices <- shuffle_inds[test_indices]
       }
       else
+        # Monte-Carlo cross-validation
         test_indices = sample(n, round(n * CV$test_size))
       test_indices
     },
-    get_modelPerf = function(test_indices, p=0) {
-      getOnePerf <- function(p) {
-        model_pred <- private$gmodel$get(dataHO, targetHO, p)
-        prediction <- model_pred(testX)
-        perf <- private$quality(prediction, testY)
-        list(model=model_pred, perf=perf)
-      }
-      dataHO <- private$data[-test_indices,]
-      testX <- private$data[test_indices,]
-      targetHO <- private$target[-test_indices]
-      testY <- private$target[test_indices]
-      # 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)
-      if (p >= 1)
-        # Standard CV: one model at a time
-        return (getOnePerf(p)$perf)
-      # Agghoo: loop on all models
-      best_model = NULL
-      best_perf <- -1
-      for (p in 1:private$gmodel$nmodels) {
-        model_perf <- getOnePerf(p)
-        if (model_perf$perf > best_perf) {
-          # TODO: if ex-aequos: models list + choose at random
-          best_model <- model_perf$model
-          best_perf <- model_perf$perf
+    defaultLoss = function(y1, y2) {
+      if (private$task == "classification") {
+        if (is.null(dim(y1)))
+          # Standard case: "hard" classification
+          mean(y1 != y2)
+        else {
+          # "Soft" classification: predict() outputs a probability matrix
+          # In this case "target" could be in matrix form.
+          if (!is.null(dim(y2)))
+            mean(rowSums(abs(y1 - y2)))
+          else {
+            # Or not: y2 is a "factor".
+            y2 <- as.character(y2)
+            # NOTE: the user should provide target in matrix form because
+            # matching y2 with columns is rather inefficient!
+            names <- colnames(y1)
+            positions <- list()
+            for (idx in seq_along(names))
+              positions[[ names[idx] ]] <- idx
+            mean(vapply(
+              seq_along(y2),
+              function(idx) sum(abs(y1[idx,] - positions[[ y2[idx] ]])),
+              0))
+          }
         }
       }
-      list(model=best_model, perf=best_perf)
+      else
+        # Regression
+        mean(abs(y1 - y2))
     }
   )
 )
diff --git a/R/R6_Model.R b/R/R6_Model.R
index 8912cdb..8fc2324 100644
--- a/R/R6_Model.R
+++ b/R/R6_Model.R
@@ -49,12 +49,17 @@ Model <- R6::R6Class("Model",
     },
     #' @description
     #' Returns the model at index "index", trained on dataHO/targetHO.
-    #' index is between 1 and self$nmodels.
     #' @param dataHO Matrix or data.frame
     #' @param targetHO Vector of targets (generally numeric or factor)
     #' @param index Index of the model in 1...nmodels
     get = function(dataHO, targetHO, index) {
       private$gmodel(dataHO, targetHO, private$params[[index]])
+    },
+    #' @description
+    #' Returns the parameter at index "index".
+    #' @param index Index of the model in 1...nmodels
+    getParam = function(index) {
+      private$params[[index]]
     }
   ),
   private = list(
diff --git a/R/agghoo.R b/R/agghoo.R
index f3bc740..cac2cf1 100644
--- a/R/agghoo.R
+++ b/R/agghoo.R
@@ -1,9 +1,12 @@
 #' agghoo
 #'
-#' Run the agghoo procedure. (...)
+#' Run the agghoo procedure (or standard cross-validation).
+#' Arguments specify the list of models, their parameters and the
+#' cross-validation settings, among others.
 #'
 #' @param data Data frame or matrix containing the data in lines.
-#' @param target The target values to predict. Generally a vector.
+#' @param target The target values to predict. Generally a vector,
+#'        but possibly a matrix in the case of "soft classification".
 #' @param task "classification" or "regression". Default:
 #'        regression if target is numerical, classification otherwise.
 #' @param gmodel A "generic model", which is a function returning a predict
@@ -14,11 +17,12 @@
 #' @param params A list of parameters. Often, one list cell is just a
 #'        numerical value, but in general it could be of any type.
 #'        Default: see R6::Model.
-#' @param quality A function assessing the quality of a prediction.
+#' @param loss A function assessing the error of a prediction.
 #'        Arguments are y1 and y2 (comparing a prediction to known values).
-#'        Default: see R6::AgghooCV.
+#'        loss(y1, y2) --> real number (error). Default: see R6::AgghooCV.
 #'
-#' @return An R6::AgghooCV object.
+#' @return
+#' An R6::AgghooCV object o. Then, call o$fit() and finally o$predict(newData)
 #'
 #' @examples
 #' # Regression:
@@ -27,15 +31,23 @@
 #' pr <- a_reg$predict(iris[,-c(2,5)] + rnorm(450, sd=0.1))
 #' # Classification
 #' a_cla <- agghoo(iris[,-5], iris[,5])
-#' a_cla$fit(mode="standard")
+#' a_cla$fit()
 #' pc <- a_cla$predict(iris[,-5] + rnorm(600, sd=0.1))
 #'
+#' @references
+#' Guillaume Maillard, Sylvain Arlot, Matthieu Lerasle. "Aggregated hold-out".
+#' Journal of Machine Learning Research 22(20):1--55, 2021.
+#'
 #' @export
-agghoo <- function(data, target, task = NULL, gmodel = NULL, params = NULL, quality = 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.numeric(target) && !is.factor(target) && !is.character(target))
+  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"))
@@ -52,9 +64,9 @@ agghoo <- function(data, target, task = NULL, gmodel = NULL, params = NULL, qual
     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(quality) && !is.function(quality))
+  if (!is.null(loss) && !is.function(loss))
     # No more checks here as well... TODO:?
-    stop("quality: function(y1, y2) --> Real")
+    stop("loss: function(y1, y2) --> Real")
 
   if (is.null(task)) {
     if (is.numeric(target))
@@ -65,34 +77,5 @@ agghoo <- function(data, target, task = NULL, gmodel = NULL, params = NULL, qual
   # Build Model object (= list of parameterized models)
   model <- Model$new(data, target, task, gmodel, params)
   # Return AgghooCV object, to run and predict
-  AgghooCV$new(data, target, task, model, quality)
-}
-
-#' compareToStandard
-#'
-#' Temporary function to compare agghoo to CV
-#' (TODO: extended, in another file, more tests - when faster code).
-#'
-#' @export
-compareToStandard <- function(df, t_idx, task = NULL, rseed = -1) {
-  if (rseed >= 0)
-    set.seed(rseed)
-  if (is.null(task))
-    task <- ifelse(is.numeric(df[,t_idx]), "regression", "classification")
-  n <- nrow(df)
-  test_indices <- sample( n, round(n / ifelse(n >= 500, 10, 5)) )
-  a <- agghoo(df[-test_indices,-t_idx], df[-test_indices,t_idx], task)
-  a$fit(mode="agghoo") #default mode
-  pa <- a$predict(df[test_indices,-t_idx])
-  print(paste("error agghoo",
-              ifelse(task == "classification",
-                     mean(p != df[test_indices,t_idx]),
-                     mean(abs(pa - df[test_indices,t_idx])))))
-  # Compare with standard cross-validation:
-  a$fit(mode="standard")
-  ps <- a$predict(df[test_indices,-t_idx])
-  print(paste("error CV",
-              ifelse(task == "classification",
-                     mean(ps != df[test_indices,t_idx]),
-                     mean(abs(ps - df[test_indices,t_idx])))))
+  AgghooCV$new(data, target, task, model, loss)
 }
diff --git a/man/AgghooCV.Rd b/man/AgghooCV.Rd
index 75d1ab6..e37224b 100644
--- a/man/AgghooCV.Rd
+++ b/man/AgghooCV.Rd
@@ -7,12 +7,20 @@
 Class encapsulating the methods to run to obtain the best predictor
 from the list of models (see 'Model' class).
 }
+\section{Public fields}{
+\if{html}{\out{<div class="r6-fields">}}
+\describe{
+\item{\code{params}}{List of parameters of the V selected models}
+}
+\if{html}{\out{</div>}}
+}
 \section{Methods}{
 \subsection{Public methods}{
 \itemize{
 \item \href{#method-new}{\code{AgghooCV$new()}}
 \item \href{#method-fit}{\code{AgghooCV$fit()}}
 \item \href{#method-predict}{\code{AgghooCV$predict()}}
+\item \href{#method-getParams}{\code{AgghooCV$getParams()}}
 \item \href{#method-clone}{\code{AgghooCV$clone()}}
 }
 }
@@ -22,7 +30,7 @@ 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$new(data, target, task, gmodel, quality = NULL)}\if{html}{\out{</div>}}
+\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$new(data, target, task, gmodel, loss = NULL)}\if{html}{\out{</div>}}
 }
 
 \subsection{Arguments}{
@@ -36,8 +44,7 @@ Create a new AgghooCV object.
 
 \item{\code{gmodel}}{Generic model returning a predictive function}
 
-\item{\code{quality}}{Function assessing the quality of a prediction;
-quality(y1, y2) --> real number}
+\item{\code{loss}}{Function assessing the error of a prediction}
 }
 \if{html}{\out{</div>}}
 }
@@ -48,10 +55,7 @@ quality(y1, y2) --> real number}
 \subsection{Method \code{fit()}}{
 Fit an agghoo model.
 \subsection{Usage}{
-\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$fit(
-  CV = list(type = "MC", V = 10, test_size = 0.2, shuffle = TRUE),
-  mode = "agghoo"
-)}\if{html}{\out{</div>}}
+\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$fit(CV = list(type = "MC", V = 10, test_size = 0.2, shuffle = TRUE))}\if{html}{\out{</div>}}
 }
 
 \subsection{Arguments}{
@@ -64,8 +68,6 @@ Fit an agghoo model.
   (irrelevant for V-fold). Default: 0.2.
 - shuffle: wether or not to shuffle data before V-fold.
   Irrelevant for Monte-Carlo; default: TRUE}
-
-\item{\code{mode}}{"agghoo" or "standard" (for usual cross-validation)}
 }
 \if{html}{\out{</div>}}
 }
@@ -76,19 +78,26 @@ Fit an agghoo model.
 \subsection{Method \code{predict()}}{
 Predict an agghoo model (after calling fit())
 \subsection{Usage}{
-\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$predict(X, weight = "uniform")}\if{html}{\out{</div>}}
+\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$predict(X)}\if{html}{\out{</div>}}
 }
 
 \subsection{Arguments}{
 \if{html}{\out{<div class="arguments">}}
 \describe{
 \item{\code{X}}{Matrix or data.frame to predict}
-
-\item{\code{weight}}{"uniform" (default) or "quality" to weight votes or
-average models performances (TODO: bad idea?!)}
 }
 \if{html}{\out{</div>}}
 }
+}
+\if{html}{\out{<hr>}}
+\if{html}{\out{<a id="method-getParams"></a>}}
+\if{latex}{\out{\hypertarget{method-getParams}{}}}
+\subsection{Method \code{getParams()}}{
+Return the list of V best parameters (after calling fit())
+\subsection{Usage}{
+\if{html}{\out{<div class="r">}}\preformatted{AgghooCV$getParams()}\if{html}{\out{</div>}}
+}
+
 }
 \if{html}{\out{<hr>}}
 \if{html}{\out{<a id="method-clone"></a>}}
diff --git a/man/Model.Rd b/man/Model.Rd
index 32f3ca2..12fc373 100644
--- a/man/Model.Rd
+++ b/man/Model.Rd
@@ -21,6 +21,7 @@ Model family can be chosen among "rf", "tree", "ppr" and "knn" for now.
 \itemize{
 \item \href{#method-new}{\code{Model$new()}}
 \item \href{#method-get}{\code{Model$get()}}
+\item \href{#method-getParam}{\code{Model$getParam()}}
 \item \href{#method-clone}{\code{Model$clone()}}
 }
 }
@@ -55,7 +56,6 @@ automatically given data and target nature if not provided.}
 \if{latex}{\out{\hypertarget{method-get}{}}}
 \subsection{Method \code{get()}}{
 Returns the model at index "index", trained on dataHO/targetHO.
-index is between 1 and self$nmodels.
 \subsection{Usage}{
 \if{html}{\out{<div class="r">}}\preformatted{Model$get(dataHO, targetHO, index)}\if{html}{\out{</div>}}
 }
@@ -67,6 +67,23 @@ index is between 1 and self$nmodels.
 
 \item{\code{targetHO}}{Vector of targets (generally numeric or factor)}
 
+\item{\code{index}}{Index of the model in 1...nmodels}
+}
+\if{html}{\out{</div>}}
+}
+}
+\if{html}{\out{<hr>}}
+\if{html}{\out{<a id="method-getParam"></a>}}
+\if{latex}{\out{\hypertarget{method-getParam}{}}}
+\subsection{Method \code{getParam()}}{
+Returns the parameter at index "index".
+\subsection{Usage}{
+\if{html}{\out{<div class="r">}}\preformatted{Model$getParam(index)}\if{html}{\out{</div>}}
+}
+
+\subsection{Arguments}{
+\if{html}{\out{<div class="arguments">}}
+\describe{
 \item{\code{index}}{Index of the model in 1...nmodels}
 }
 \if{html}{\out{</div>}}
diff --git a/man/agghoo.Rd b/man/agghoo.Rd
index 69dafed..179d309 100644
--- a/man/agghoo.Rd
+++ b/man/agghoo.Rd
@@ -4,12 +4,13 @@
 \alias{agghoo}
 \title{agghoo}
 \usage{
-agghoo(data, target, task = NULL, gmodel = NULL, params = NULL, quality = NULL)
+agghoo(data, target, task = NULL, gmodel = NULL, params = NULL, loss = NULL)
 }
 \arguments{
 \item{data}{Data frame or matrix containing the data in lines.}
 
-\item{target}{The target values to predict. Generally a vector.}
+\item{target}{The target values to predict. Generally a vector,
+but possibly a matrix in the case of "soft classification".}
 
 \item{task}{"classification" or "regression". Default:
 regression if target is numerical, classification otherwise.}
@@ -24,15 +25,17 @@ of 'param's. See params argument. Default: see R6::Model.}
 numerical value, but in general it could be of any type.
 Default: see R6::Model.}
 
-\item{quality}{A function assessing the quality of a prediction.
+\item{loss}{A function assessing the error of a prediction.
 Arguments are y1 and y2 (comparing a prediction to known values).
-Default: see R6::AgghooCV.}
+loss(y1, y2) --> real number (error). Default: see R6::AgghooCV.}
 }
 \value{
-An R6::AgghooCV object.
+An R6::AgghooCV object o. Then, call o$fit() and finally o$predict(newData)
 }
 \description{
-Run the agghoo procedure. (...)
+Run the agghoo procedure (or standard cross-validation).
+Arguments specify the list of models, their parameters and the
+cross-validation settings, among others.
 }
 \examples{
 # Regression:
@@ -41,7 +44,11 @@ a_reg$fit()
 pr <- a_reg$predict(iris[,-c(2,5)] + rnorm(450, sd=0.1))
 # Classification
 a_cla <- agghoo(iris[,-5], iris[,5])
-a_cla$fit(mode="standard")
+a_cla$fit()
 pc <- a_cla$predict(iris[,-5] + rnorm(600, sd=0.1))
 
 }
+\references{
+Guillaume Maillard, Sylvain Arlot, Matthieu Lerasle. "Aggregated hold-out".
+Journal of Machine Learning Research 22(20):1--55, 2021.
+}
diff --git a/man/compareToStandard.Rd b/man/compareToStandard.Rd
deleted file mode 100644
index 742ecaa..0000000
--- a/man/compareToStandard.Rd
+++ /dev/null
@@ -1,12 +0,0 @@
-% Generated by roxygen2: do not edit by hand
-% Please edit documentation in R/agghoo.R
-\name{compareToStandard}
-\alias{compareToStandard}
-\title{compareToStandard}
-\usage{
-compareToStandard(df, t_idx, task = NULL, rseed = -1)
-}
-\description{
-Temporary function to compare agghoo to CV
-(TODO: extended, in another file, more tests - when faster code).
-}
diff --git a/test/README b/test/README
new file mode 100644
index 0000000..722eed7
--- /dev/null
+++ b/test/README
@@ -0,0 +1,7 @@
+# Usage
+#######
+
+source("compareToCV.R")
+
+# rseed: >= 0 for reproducibility.
+compareToCV(data, target_column_index, rseed = -1)
diff --git a/test/compareToCV.R b/test/compareToCV.R
new file mode 100644
index 0000000..255c70c
--- /dev/null
+++ b/test/compareToCV.R
@@ -0,0 +1,137 @@
+library(agghoo)
+
+standardCV <- 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))
+          }
+        }
+      }
+    }
+  }
+
+  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$V))
+    list_testinds[[v]] <- get_testIndices(v, shuffle_inds)
+
+  gmodel <- agghoo::Model$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)) {
+      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$get(dataHO, targetHO, p)
+      prediction <- model_pred(testX)
+      error <- error + loss(prediction, testY)
+    }
+    if (error <= best_error) {
+      newModel <- list(model=model_pred, param=gmodel$getParam(p))
+      if (error == best_error)
+        best_model[[length(best_model)+1]] <- newModel
+      else {
+        best_model <- list(newModel)
+        best_error <- error
+      }
+    }
+  }
+  best_model[[ sample(length(best_model), 1) ]]
+}
+
+compareToCV <- function(df, t_idx, task=NULL, rseed=-1, verbose=TRUE) {
+  if (rseed >= 0)
+    set.seed(rseed)
+  if (is.null(task))
+    task <- ifelse(is.numeric(df[,t_idx]), "regression", "classification")
+  n <- nrow(df)
+  test_indices <- sample( n, round(n / ifelse(n >= 500, 10, 5)) )
+  a <- agghoo(df[-test_indices,-t_idx], df[-test_indices,t_idx], task)
+  a$fit()
+  if (verbose) {
+    print("Parameters:")
+    print(unlist(a$getParams()))
+  }
+  pa <- a$predict(df[test_indices,-t_idx])
+  err_a <- ifelse(task == "classification",
+                  mean(pa != df[test_indices,t_idx]),
+                  mean(abs(pa - df[test_indices,t_idx])))
+  if (verbose)
+    print(paste("error agghoo:", err_a))
+  # Compare with standard cross-validation:
+  s <- standardCV(df[-test_indices,-t_idx], df[-test_indices,t_idx], task)
+  if (verbose)
+    print(paste( "Parameter:", s$param ))
+  ps <- s$model(df[test_indices,-t_idx])
+  err_s <- ifelse(task == "classification",
+                  mean(ps != df[test_indices,t_idx]),
+                  mean(abs(ps - df[test_indices,t_idx])))
+  if (verbose)
+    print(paste("error CV:", err_s))
+  c(err_a, err_s)
+}
+
+library(parallel)
+compareMulti <- function(df, t_idx, task = NULL, N = 100, nc = NA) {
+  if (is.na(nc))
+    nc <- detectCores()
+  errors <- mclapply(1:N,
+                     function(n) {
+                       compareToCV(df, t_idx, task, n, verbose=FALSE) },
+                     mc.cores = nc)
+  print("error agghoo vs. cross-validation:")
+  Reduce('+', errors) / N
+}