softmaxreg (version 1.2)

trainModel: Train softmax regression and classification model

Description

This function implements a feedforward neural networks with multiple hidden layers and a softmax final layer. For classification, set type as "class" and y as factor vector; for regression, set type as "raw" and y as matrix with dimension nObs * K. K denotes the group number.

Usage

trainModel(x, y, hidden, funName, maxit, rang, type, algorithm, rate, L2, penalty, threshold, batch)

Arguments

x

matrix or data frame of x input values.

y

vector of target values for 'class' type classfication and matrix or data frame of target values for 'raw' type regression.

hidden

vector of integers specifying the number of hidden nodes in each layer.

funName

activation function name of neuron, e.g. 'sigmoid', 'tanh', 'relu' etc. In default, it is set to 'sigmoid'.

maxit

maximum number of iterations.Default 3000.

rang

parameter for the range of initial random weights. Default 0.1 [-rang, rang].

type

parameter indicating the type of softmax task: "class" denotes the softmax classfication model and the fitted values are factors; "raw" denotes softmax regression model and the fitted values are the probability or percentage of each group. Default "class".

algorithm

parameter indicating which gradient descenting learning algorithm to use, including "sgd", "adagrad", "rmsprop", "adadelta", etc. Default "adagrad".

rate

parameter of learning rate. Default 0.05.

L2

Boolean variable indicating whether L2 regularization term is added to the loss function and gradient to prevent overfitting. Default FALSE.

penalty

Parameter for the penalty cost of the L2 regularization term if L2 is TRUE. Default 1e-4.

threshold

Parameter for the threshold of iteration convergence: loss value less than threshold. Default 1e-4.

batch

Parameter for mini-batch size. Default 50.

Value

object of class "softmax"

See Also

softmaxReg

Examples

## Not run: 
# library(softmaxreg)
# data(iris)
# x = iris[,1:4]
# y = iris$Species
# sofmax_model = trainModel(x, y, hidden = c(5), funName = 'sigmoid', maxit = 3000,
#     rang = 0.1, type = "class", algorithm = "adagrad", rate = 0.05, threshold = 1e-3)
# summary(sofmax_model)
# yFitMat = sofmax_model$fitted.values
# yFit = c()
# for (i in 1:length(y)) {
# 	yFit = c(yFit, which(yFitMat[i,]==max(yFitMat[i,])))
# }
# table(y, yFit)
# ## End(Not run)