Fixes on random forests minNode args.

models/files/Rborist.R

@@ -13,7 +13,7 @@ modelInfo <- list(label = "Random Forest",
                                         minNode = ifelse(is.factor(y), 2, 3))
                     } else {
                       out <- data.frame(predFixed = sample(1:ncol(x), size = len, replace = TRUE), #removed unique
-                                        minNode = sample(1:(min(20,nrow(trainX))), size = len, replace = TRUE)) # might cause warning for very small samples < 20
+                                        minNode = sample(1:(min(20,nrow(x))), size = len, replace = TRUE)) # might cause warning for very small samples < 20
                     }
                     out
                   },
@@ -46,4 +46,4 @@ modelInfo <- list(label = "Random Forest",
                                                       e1071::classAgreement(x$validation$confusion)[["kappa"]]))
                     names(out) <- if(x$problemType == "Regression") c("RMSE", "Rsquared") else c("Accuracy", "Kappa")
                     out
-                  })
+                  })

models/files/mlpKerasDecay.R

@@ -53,6 +53,7 @@ modelInfo <- list(label = "Multilayer Perceptron Network with Weight Decay",
                         units = param$size, 
                         activation = as.character(param$activation), 
                         input_shape = ncol(x),
+                        kernel_initializer = keras::initializer_glorot_uniform(),
                         kernel_regularizer = keras::regularizer_l2(param$lambda)
                       )
                     if(is.factor(y)) {

models/files/mlpKerasDecayCost.R

@@ -56,6 +56,7 @@ modelInfo <- list(label = "Multilayer Perceptron Network with Weight Decay",
                         units = param$size, 
                         activation = as.character(param$activation), 
                         input_shape = ncol(x),
+                        kernel_initializer = keras::initializer_glorot_uniform(),
                         kernel_regularizer = keras::regularizer_l2(param$lambda)
                       )
                     y <- class2ind(y)

models/files/mlpKerasDropout.R

@@ -53,6 +53,7 @@ modelInfo <- list(label = "Multilayer Perceptron Network with Dropout",
                       keras::layer_dense(
                         units = param$size, 
                         activation = as.character(param$activation), 
+                        kernel_initializer = keras::initializer_glorot_uniform(),
                         input_shape = ncol(x)
                       ) %>%
                       keras::layer_dropout(rate = param$dropout,

models/files/mlpKerasDropoutCost.R

@@ -55,6 +55,7 @@ modelInfo <- list(label = "Multilayer Perceptron Network with Dropout",
                       keras::layer_dense(
                         units = param$size, 
                         activation = as.character(param$activation), 
+                        kernel_initializer = keras::initializer_glorot_uniform(),
                         input_shape = ncol(x)
                       ) %>%
                       keras::layer_dropout(rate = param$dropout,

models/files/mxnet.R

@@ -46,7 +46,7 @@ modelInfo <- list(label = "Neural Network",
                                            array.layout = "rowmajor",
                                            activation = rep( as.character(param$activation), length(num_units)),
                                            # Use He/MSRA when available in R 
-                                           initializer = mx.init.Xavier(factor_type = "avg", magnitude = 3, rnd_type = 'uniform'),
+                                           initializer = mxnet::mx.init.Xavier(factor_type = "avg", magnitude = 3, rnd_type = 'uniform'),
                                            ...)
                     } else {
                       y <- as.numeric(y) - 1
@@ -60,7 +60,7 @@ modelInfo <- list(label = "Neural Network",
                                            eval.metric = mx.metric.accuracy, 
                                            array.layout = "rowmajor",
                                            activation = rep( as.character(param$activation), length(num_units)),
-                                           initializer = mx.init.Xavier(factor_type = "avg", magnitude = 3, rnd_type = 'uniform'), 
+                                           initializer = mxnet::mx.init.Xavier(factor_type = "avg", magnitude = 3, rnd_type = 'uniform'), 
                                            ...)
                     }
                     if(last)

models/files/mxnetAdam.R

@@ -41,7 +41,7 @@ modelInfo <- list(label = "Neural Network",
                                            hidden_node = num_units,
                                            activation = rep( as.character(param$activation), length(num_units)),
                                            # Consider using He/MSRA paper when available in R
-                                           initializer = mx.init.Xavier(factor_type = "avg", magnitude = 3, rnd_type = 'uniform'),
+                                           initializer = mxnet::mx.init.Xavier(factor_type = "avg", magnitude = 3, rnd_type = 'uniform'),
                                            ...)
                     } else {
                       y <- as.numeric(y) - 1 
@@ -54,7 +54,7 @@ modelInfo <- list(label = "Neural Network",
                                           dropout = param$dropout, 
                                           hidden_node = num_units,
                                           activation = rep( as.character(param$activation), length(num_units)),
-                                          initializer = mx.init.Xavier(factor_type = "avg", magnitude = 3, rnd_type = 'uniform'),
+                                          initializer = mxnet::mx.init.Xavier(factor_type = "avg", magnitude = 3, rnd_type = 'uniform'),
                                           ...)
                     }
                     if(last)

models/files/ranger.R

@@ -35,7 +35,7 @@ modelInfo <- list(label = "Random Forest",
                         c("variance", "extratrees", "maxstat")
                       out <-
                         data.frame(
-                          min.node.size= sample(0:20, size = len, replace = TRUE), 
+                          min.node.size= sample(1:(min(20,nrow(x))), size = len, replace = TRUE), 
                           mtry = sample(1:ncol(x), size = len, replace = TRUE),
                           splitrule = sample(srules, size = len, replace = TRUE)
                         )

models/files/rlm.R

@@ -38,5 +38,14 @@ modelInfo <- list(label = "Robust Linear Model",
                     predict(modelFit, newdata)
                   },
                   prob = NULL,
+                  varImp = function(object, ...) {
+                    values <- summary(object)$coef
+                    varImps <-  abs(values[ !grepl( rownames(values), pattern = 'Intercept' ), 
+                                             grep("value$", colnames(values)), drop = FALSE])
+                    out <- data.frame(varImps)
+                    colnames(out) <- "Overall"
+                    if(!is.null(names(varImps))) rownames(out) <- names(varImps)
+                    out   
+                  },
                   tags = c("Linear Regression", "Robust Model", "Accepts Case Weights"),
                   sort = function(x) x)