trainModel: Train softmax regression and classification model

In softmaxreg: Training Multi-Layer Neural Network for Softmax Regression and Classification

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"

weights	Optimal weights parameters found by softmax model, including list of W and B for all layers.
data	Input Training Data.
K	Number of K groups fitted by softmax model.
loss	Numeric vector of the loss function values over iterations.
fitted.values	Matrix of the fitted values yFitMat for the training data. Dimensions: number of observations by K;
iteration	Number of iteration reached before stop.
convergence	Boolean variable for whether softmax model reached convergence.

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)

softmaxreg documentation built on May 2, 2019, 6:08 a.m.