cuml4r (version 0.1.0)

cuml_svm: Train a SVM model.

Description

Train a Support Vector Machine model for classification or regression tasks.

Usage

cuml_svm(
  x,
  y = NULL,
  formula = NULL,
  mode = c("classification", "regression"),
  cost = 1,
  kernel = c("rbf", "tanh", "polynomial", "linear"),
  gamma = 1/ncol(x),
  coef0 = 0,
  degree = 3L,
  tol = 0.001,
  max_iter = 100L * nrow(x),
  nochange_steps = 1000L,
  cache_size = 1024,
  epsilon = 0.1,
  sample_weights = NULL,
  cuml_log_level = c("off", "critical", "error", "warn", "info", "debug", "trace")
)

Arguments

x

The input matrix or dataframe. Each data point should be a row and should consist of numeric values only.

y

A numeric vector of desired responses.

formula

If 'x' is a dataframe, then a R formula syntax of the form '<response col> ~ .' or '<response col> ~ <predictor 1> + <predictor 2> + ...' may be used to specify the response column and the predictor column(s).

mode

Type of task to perform. Should be either "classification" or "regression".

cost

A positive number for the cost of predicting a sample within or on the wrong side of the margin. Default: 1.

kernel

Type of the SVM kernel function (must be one of "rbf", "tanh", "polynomial", or "linear"). Default: "rbf".

gamma

The gamma coefficient (only relevant to polynomial, RBF, and tanh kernel functions, see explanations below). Default: 1 / (num features).

The following kernels are implemented: - RBF K(x_1, x_2) = exp(-gamma |x_1-x_2|^2) - TANH K(x_1, x_2) = tanh(gamma <x_1,x_2> + coef0) - POLYNOMIAL K(x_1, x_2) = (gamma <x_1,x_2> + coef0)^degree - LINEAR K(x_1,x_2) = <x_1,x_2>, where < , > denotes the dot product.

coef0

The 0th coefficient (only applicable to polynomial and tanh kernel functions, see explanations below). Default: 0.

The following kernels are implemented: - RBF K(x_1, x_2) = exp(-gamma |x_1-x_2|^2) - TANH K(x_1, x_2) = tanh(gamma <x_1,x_2> + coef0) - POLYNOMIAL K(x_1, x_2) = (gamma <x_1,x_2> + coef0)^degree - LINEAR K(x_1,x_2) = <x_1,x_2>, where < , > denotes the dot product.

degree

Degree of the polynomial kernel function (note: not applicable to other kernel types, see explanations below). Default: 3.

The following kernels are implemented: - RBF K(x_1, x_2) = exp(-gamma |x_1-x_2|^2) - TANH K(x_1, x_2) = tanh(gamma <x_1,x_2> + coef0) - POLYNOMIAL K(x_1, x_2) = (gamma <x_1,x_2> + coef0)^degree - LINEAR K(x_1,x_2) = <x_1,x_2>, where < , > denotes the dot product.

tol

Tolerance to stop fitting. Default: 1e-3.

max_iter

Maximum number of outer iterations in SmoSolver. Default: 100 * (num samples).

nochange_steps

Number of steps with no change w.r.t convergence. Default: 1000.

cache_size

Size of kernel cache (MiB) in device memory. Default: 1024.

epsilon

Espsilon parameter of the epsilon-SVR model. There is no penalty for points that are predicted within the epsilon-tube around the target values. Please note this parameter is only relevant for regression tasks. Default: 0.1.

sample_weights

Optional weight assigned to each input data point.

cuml_log_level

Log level within cuML library functions. Must be one of "off", "critical", "error", "warn", "info", "debug", "trace". Default: off.

Value

A Support Vector Machine classifier / regressor object that can be used with the 'predict' S3 generic to make predictions on new data points.

Examples

# NOT RUN {
library(cuml4r)

model <- cuml_svm(
  iris[1:100,],
  formula = Species ~ .,
  mode = "classification",
  kernel = "rbf"
)

predictions <- predict(model, iris[1:100,])

cat("Iris species predictions: ", predictions, "\n")

model <- cuml_svm(
  mtcars,
  formula = mpg ~ .,
  mode = "regression",
  kernel = "rbf"
)

predictions <- predict(model, mtcars)

cat("MPG predictions:", predictions, "\n")
# }