Implementation of classif.dbnDNN

DESCRIPTION

@@ -71,7 +71,7 @@ Suggests:
     Cubist,
     curl,
     dbarts,
-    deepnet,
+    deepnet (>= 0.2.1),
     earth,
     evtree,
     ExhaustiveSearch,

NAMESPACE

@@ -20,6 +20,7 @@ export(LearnerClassifC50)
 export(LearnerClassifCForest)
 export(LearnerClassifCTree)
 export(LearnerClassifCatboost)
+export(LearnerClassifDbnDNN)
 export(LearnerClassifDecisionStump)
 export(LearnerClassifDecisionTable)
 export(LearnerClassifDiagLda)

NEWS.md

@@ -10,6 +10,7 @@
   - `LearnerRegrPlsr`
   - `LearnerRegrLaGP`
   - `LearnerClassifNNTrain`
+  - `LearnerClassifDbnDNN`
 
 ## Breaking Changes
 

R/bibentries.R

@@ -432,6 +432,13 @@ bibentries = c( # nolint start
     year = "2018",
     publisher = "BioMed Central"
   ),
+  rong2014deepnet = bibentry("manual",
+    title = "{deepnet}: Deep Learning Toolkit in R",
+    author = "Xiao Rong",
+    year = "2014",
+    note = "R package version 0.2.1",
+    url = "https://CRAN.R-project.org/package=deepnet"
+  ),
   zhao2019dnnsurv = bibentry("article",
     title = "Dnnsurv: Deep neural networks for survival analysis using pseudo values",
     author = "Zhao, Lili and Feng, Dai",

R/learner_deepnet_classif_dbnDNN.R

@@ -0,0 +1,105 @@
+#' @title Deep Neural Network with DBN Pretraining Learner
+#' @author awinterstetter
+#' @name mlr_learners_classif.dbnDNN
+#'
+#' @description
+#' Deep neural network with weights initialized by a deep belief network.
+#' Calls [deepnet::dbn.dnn.train()] from \CRANpkg{deepnet}.
+#' 
+#' This learner works with tasks that have at least two features.
+#'
+#' @section Initial parameter values:
+#' - `output` is set to `"softmax"` to enable class probabilities.
+#'
+#' @templateVar id classif.dbnDNN
+#' @template learner
+#'
+#' @references
+#' `r format_bib("rong2014deepnet")`
+#'
+#' @template seealso_learner
+#' @template example
+#' @export
+LearnerClassifDbnDNN = R6Class("LearnerClassifDbnDNN",
+  inherit = LearnerClassif,
+  public = list(
+    #' @description
+    #' Creates a new instance of this [R6][R6::R6Class] class.
+    initialize = function() {
+      param_set = ps(
+        hidden = p_uty(init = 10L, tags = "train", custom_check = function(x) {
+          check_integerish(x, lower = 1, any.missing = FALSE, min.len = 1)
+        }),
+        activationfun = p_fct(levels = c("sigm", "linear", "tanh"), init = "sigm", tags = "train"),
+        learningrate = p_dbl(init = 0.8, lower = 0, tags = "train"),
+        momentum = p_dbl(init = 0.5, lower = 0, tags = "train"),
+        learningrate_scale = p_dbl(init = 1, lower = 0, tags = "train"),
+        numepochs = p_int(init = 3L, lower = 1L, tags = "train"),
+        batchsize = p_int(init = 100L, lower = 1L, tags = "train"),
+        output = p_fct(levels = c("sigm", "linear", "softmax"), init = "softmax", tags = "train"),
+        hidden_dropout = p_dbl(init = 0, lower = 0, upper = 1, tags = "train"),
+        visible_dropout = p_dbl(init = 0, lower = 0, upper = 1, tags = "train"),
+        cd = p_int(init = 1L, lower = 1L, tags = "train")
+      )
+
+      super$initialize(
+        id = "classif.dbnDNN",
+        packages = "deepnet",
+        feature_types = c("integer", "numeric"),
+        predict_types = c("response", "prob"),
+        param_set = param_set,
+        properties = c("twoclass", "multiclass"),
+        man = "mlr3extralearners::mlr_learners_classif.dbnDNN",
+        label = "Deep Belief Network DNN"
+      )
+    }
+  ),
+
+  private = list(
+    .train = function(task) {
+      pars = self$param_set$get_values(tags = "train")
+
+      x = data.matrix(task$data(cols = task$feature_names))
+      y = as.numeric(task$truth())
+      dict = sort(unique(y))
+      onehot = matrix(0, length(y), length(dict))
+
+      for (i in seq_along(dict)) {
+        ind = which(y == dict[i])
+        onehot[ind, i] = 1
+      }
+
+      invoke(
+        deepnet::dbn.dnn.train,
+        x = x,
+        y = onehot,
+        .args = pars
+      )
+    },
+
+    .predict = function(task) {
+      newdata = data.matrix(ordered_features(task, self))
+
+      pred = invoke(
+        deepnet::nn.predict,
+        self$model,
+        newdata
+      )
+
+      pred = as.matrix(pred)
+      colnames(pred) = task$class_names
+
+      if (self$predict_type == "response") {
+        response = factor(
+          task$class_names[max.col(pred, ties.method = "first")],
+          levels = task$class_names
+        )
+        return(list(response = response))
+      }
+
+      list(prob = pred)
+    }
+  )
+)
+
+.extralrns_dict$add("classif.dbnDNN", LearnerClassifDbnDNN)

R/learner_deepnet_classif_nnTrain.R

@@ -5,7 +5,6 @@
 #' Calls [deepnet::nn.train()] from \CRANpkg{deepnet}.
 #'
 #' @section Initial parameter values:
-#' - `hidden` defaults to `10`.
 #' - `output` is set to `"softmax"` for probabilistic classification.
 #'
 #' @templateVar id classif.nnTrain