1% Generated by roxygen2: do not edit by hand 2% Please edit documentation in R/relu.R 3\name{step_relu} 4\alias{step_relu} 5\title{Apply (Smoothed) Rectified Linear Transformation} 6\usage{ 7step_relu( 8 recipe, 9 ..., 10 role = "predictor", 11 trained = FALSE, 12 shift = 0, 13 reverse = FALSE, 14 smooth = FALSE, 15 prefix = "right_relu_", 16 columns = NULL, 17 skip = FALSE, 18 id = rand_id("relu") 19) 20} 21\arguments{ 22\item{recipe}{A recipe object. The step will be added to the 23sequence of operations for this recipe.} 24 25\item{...}{One or more selector functions to choose variables 26for this step. See \code{\link[=selections]{selections()}} for more details.} 27 28\item{role}{For model terms created by this step, what analysis role should 29they be assigned? By default, the new columns created by this step from 30the original variables will be used as \emph{predictors} in a model.} 31 32\item{trained}{A logical to indicate if the quantities for 33preprocessing have been estimated.} 34 35\item{shift}{A numeric value dictating a translation to apply to the data.} 36 37\item{reverse}{A logical to indicate if the left hinge should be used as 38opposed to the right hinge.} 39 40\item{smooth}{A logical indicating if the softplus function, a smooth 41approximation to the rectified linear transformation, should be used.} 42 43\item{prefix}{A prefix for generated column names, defaults to "right_relu_" 44for right hinge transformation and "left_relu_" for reversed/left hinge 45transformations.} 46 47\item{columns}{A character string of variable names that will 48be populated (eventually) by the \code{terms} argument.} 49 50\item{skip}{A logical. Should the step be skipped when the 51recipe is baked by \code{\link[=bake.recipe]{bake.recipe()}}? While all operations are baked 52when \code{\link[=prep.recipe]{prep.recipe()}} is run, some operations may not be able to be 53conducted on new data (e.g. processing the outcome variable(s)). 54Care should be taken when using \code{skip = TRUE} as it may affect 55the computations for subsequent operations.} 56 57\item{id}{A character string that is unique to this step to identify it.} 58} 59\value{ 60An updated version of \code{recipe} with the new step added to the 61sequence of any existing operations. 62} 63\description{ 64\code{step_relu} creates a \emph{specification} of a recipe step that 65will apply the rectified linear or softplus transformations to numeric 66data. The transformed data is added as new columns to the data matrix. 67} 68\details{ 69The rectified linear transformation is calculated as 70\deqn{max(0, x - c)} and is also known as the ReLu or right hinge function. 71If \code{reverse} is true, then the transformation is reflected about the 72y-axis, like so: \deqn{max(0, c - x)} Setting the \code{smooth} option 73to true will instead calculate a smooth approximation to ReLu 74according to \deqn{ln(1 + e^(x - c)} The \code{reverse} argument may 75also be applied to this transformation. 76} 77\section{Connection to MARS}{ 78 79 80The rectified linear transformation is used in Multivariate Adaptive 81Regression Splines as a basis function to fit piecewise linear functions to 82data in a strategy similar to that employed in tree based models. The 83transformation is a popular choice as an activation function in many 84neural networks, which could then be seen as a stacked generalization of 85MARS when making use of ReLu activations. The hinge function also appears 86in the loss function of Support Vector Machines, where it penalizes 87residuals only if they are within a certain margin of the decision boundary. 88} 89 90\examples{ 91library(modeldata) 92data(biomass) 93 94biomass_tr <- biomass[biomass$dataset == "Training",] 95biomass_te <- biomass[biomass$dataset == "Testing",] 96 97rec <- recipe(HHV ~ carbon + hydrogen + oxygen + nitrogen + sulfur, 98 data = biomass_tr) 99 100transformed_te <- rec \%>\% 101 step_relu(carbon, shift = 40) \%>\% 102 prep(biomass_tr) \%>\% 103 bake(biomass_te) 104 105transformed_te 106} 107\seealso{ 108Other individual transformation steps: 109\code{\link{step_BoxCox}()}, 110\code{\link{step_YeoJohnson}()}, 111\code{\link{step_bs}()}, 112\code{\link{step_harmonic}()}, 113\code{\link{step_hyperbolic}()}, 114\code{\link{step_inverse}()}, 115\code{\link{step_invlogit}()}, 116\code{\link{step_logit}()}, 117\code{\link{step_log}()}, 118\code{\link{step_mutate}()}, 119\code{\link{step_ns}()}, 120\code{\link{step_poly}()}, 121\code{\link{step_sqrt}()} 122} 123\concept{individual transformation steps} 124