gputools (version 0.26)

gpuSvmTrain: Train a support vector machine on a data set

Description

This function trains, with the aid of a GPU, a support vector machine using the input data x separated into classes y. The function is capable of both regression (the entries of y are continuous) and non-regression (each entry of y is either -1.f or 1.f). The underlying code is adapted from Austin Carpenter's cuSVM which can be found at http://patternsonascreen.net/cuSVM.html

Usage

gpuSvmTrain(y, x, C = 10, kernelWidth = 0.125, eps = 0.5,
	    stoppingCrit = 0.001, isRegression = FALSE)

Arguments

y

a vector of floating point numbers. The length of y should equal the number of rows of x. In the case of isRegression = FALSE, each entry of y is the category of the row of data x. The negative category is indicated by -1 and the positive catego

x

a matrix of floating point numbers. Each row i is a point with a category given by y[i]. This is the data set used for training the svm.

C

a single floating point number. This is the SVM regularization parameter.

kernelWidth

a single floating point number. This is the scalar Gaussian kernel parameter.

eps

a single floating point number. This is the epsilon used in regression mode.

stoppingCrit

a single floating point number. This is the optimization stopping criterion.

isRegression

a single logical value. If isRegression is set to TRUE then regression is performed and the y value may be continuously valued. If not, then we use normal svm training and each value in y must be either -1 or 1.

Value

• a list consisting of the following elements: supportVectors, svCoefficients, and svOffset. The element supportVectors is a matrix of single precision floating point numbers. These are the support vectors corresponding to the coefficients in svCoefficients. Row i of supportVectors contains ncol(x) columns and has coefficient svCoefficients[i]. The element svCoefficients is a single precision vector of the support vector coefficients. The element svOffset is a single floating point number of single precision. It is the offset for the prediction function.

References

Carpenter, Austin, cuSVM: a cuda implementation of support vector classification and regression, http://http://patternsonascreen.net/cuSVM.html

Examples

# y is discrete: -1 or 1 and we set isRegression to FALSE
y <- round(runif(100, min = 0, max = 1))
for(i in 1:5) { if(y[i] == 0) {y[i] <- -1}}

x <- matrix(runif(500), 100, 5)

# a <- gpuSvmTrain(y, x, isRegression = FALSE)
# print(a)

# b <- gpuSvmPredict(x, a$supportVectors, a$svCoefficients, a$svOffset, isRegression = FALSE)
# print(b)

# this time around, y : -1 or 1 and we set isRegression to FALSE
y <- runif(100, min = -1, max = 1)

x <- matrix(runif(500), 100, 5)

# a <- gpuSvmTrain(y, x, isRegression = TRUE)
# print(a)

# b <- gpuSvmPredict(x, a$supportVectors, a$svCoefficients, a$svOffset, isRegression = TRUE)
# print(b)