Package 'elmNNRcpp'

Fit an extreme learning model

Description

Formula interface for elm_train, transforms a data frame and formula into the necessary input for elm_train, automatically calls onehot_encode for classification.

Usage

elm(
  formula,
  data,
  nhid,
  actfun,
  init_weights = "normal_gaussian",
  bias = FALSE,
  moorep_pseudoinv_tol = 0.01,
  leaky_relu_alpha = 0,
  seed = 1,
  verbose = FALSE
)

Arguments

formula	formula used to specify the regression or classification.
data	data.frame with the data
nhid	a numeric value specifying the hidden neurons. Must be >= 1
actfun	a character string specifying the type of activation function. It should be one of the following : 'sig' ( sigmoid ), 'sin' ( sine ), 'radbas' ( radial basis ), 'hardlim' ( hard-limit ), 'hardlims' ( symmetric hard-limit ), 'satlins' ( satlins ), 'tansig' ( tan-sigmoid ), 'tribas' ( triangular basis ), 'relu' ( rectifier linear unit ) or 'purelin' ( linear )
init_weights	a character string spcecifying the distribution from which the input-weights and the bias should be initialized. It should be one of the following : 'normal_gaussian' (normal / Gaussian distribution with zero mean and unit variance), 'uniform_positive' ( in the range [0,1] ) or 'uniform_negative' ( in the range [-1,1] )
bias	either TRUE or FALSE. If TRUE then bias weights will be added to the hidden layer
moorep_pseudoinv_tol	a numeric value. See the references web-link for more details on Moore-Penrose pseudo-inverse and specifically on the pseudo inverse tolerance value
leaky_relu_alpha	a numeric value between 0.0 and 1.0. If 0.0 then a simple relu ( f(x) = 0.0 for x < 0, f(x) = x for x >= 0 ) activation function will be used, otherwise a leaky-relu ( f(x) = alpha * x for x < 0, f(x) = x for x >= 0 ). It is applicable only if actfun equals to 'relu'
seed	a numeric value specifying the random seed. Defaults to 1
verbose	a boolean. If TRUE then information will be printed in the console

Value

elm object which can be used with predict, residuals and fitted.

Examples

elm(Species ~ ., data = iris, nhid = 20, actfun="sig")

mod_elm <- elm(Species ~ ., data = iris, nhid = 20, actfun="sig")

# predict classes
predict(mod_elm, newdata = iris[1:3,-5])

# predict probabilities
predict(mod_elm, newdata = iris[1:3,-5], type="prob")

# predict elm output
predict(mod_elm, newdata = iris[1:3,-5], type="raw")

data("Boston")
elm(medv ~ ., data = Boston, nhid = 40, actfun="relu")

data("ionosphere")
elm(class ~ ., data = ionosphere, nhid=20, actfun="relu")

---

Extreme Learning Machine predict function

Description

Extreme Learning Machine predict function

Usage

elm_predict(elm_train_object, newdata, normalize = FALSE)

Arguments

elm_train_object	it should be the output of the elm_train function
newdata	an input matrix with number of columns equal to the x parameter of the elm_train function
normalize	a boolean specifying if the output predictions in case of classification should be normalized. If TRUE then the values of each row of the output-probability-matrix that are less than 0 and greater than 1 will be pushed to the [0,1] range

Examples

library(elmNNRcpp)

#-----------
# Regression
#-----------

data(Boston, package = 'KernelKnn')

Boston = as.matrix(Boston)
dimnames(Boston) = NULL

x = Boston[, -ncol(Boston)]
y = matrix(Boston[, ncol(Boston)], nrow = length(Boston[, ncol(Boston)]), ncol = 1)

out_regr = elm_train(x, y, nhid = 20, actfun = 'purelin', init_weights = 'uniform_negative')

pr_regr = elm_predict(out_regr, x)


#---------------
# Classification
#---------------

data(ionosphere, package = 'KernelKnn')

x_class = ionosphere[, -c(2, ncol(ionosphere))]
x_class = as.matrix(x_class)
dimnames(x_class) = NULL

y_class = as.numeric(ionosphere[, ncol(ionosphere)])

y_class_onehot = onehot_encode(y_class - 1)     # class labels should begin from 0

out_class = elm_train(x_class, y_class_onehot, nhid = 20, actfun = 'relu')

pr_class = elm_predict(out_class, x_class, normalize = TRUE)

---

Extreme Learning Machine training function

Description

Extreme Learning Machine training function

Usage

elm_train(
  x,
  y,
  nhid,
  actfun,
  init_weights = "normal_gaussian",
  bias = FALSE,
  moorep_pseudoinv_tol = 0.01,
  leaky_relu_alpha = 0,
  seed = 1,
  verbose = FALSE
)

Arguments

x	a matrix. The columns of the input matrix should be of type numeric
y	a matrix. In case of regression the matrix should have n rows and 1 column. In case of classification it should consist of n rows and n columns, where n > 1 and equals to the number of the unique labels.
nhid	a numeric value specifying the hidden neurons. Must be >= 1
actfun	a character string specifying the type of activation function. It should be one of the following : 'sig' ( sigmoid ), 'sin' ( sine ), 'radbas' ( radial basis ), 'hardlim' ( hard-limit ), 'hardlims' ( symmetric hard-limit ), 'satlins' ( satlins ), 'tansig' ( tan-sigmoid ), 'tribas' ( triangular basis ), 'relu' ( rectifier linear unit ) or 'purelin' ( linear )
init_weights	a character string spcecifying the distribution from which the input-weights and the bias should be initialized. It should be one of the following : 'normal_gaussian' (normal / Gaussian distribution with zero mean and unit variance), 'uniform_positive' ( in the range [0,1] ) or 'uniform_negative' ( in the range [-1,1] )
bias	either TRUE or FALSE. If TRUE then bias weights will be added to the hidden layer
moorep_pseudoinv_tol	a numeric value. See the references web-link for more details on Moore-Penrose pseudo-inverse and specifically on the pseudo inverse tolerance value
leaky_relu_alpha	a numeric value between 0.0 and 1.0. If 0.0 then a simple relu ( f(x) = 0.0 for x < 0, f(x) = x for x >= 0 ) activation function will be used, otherwise a leaky-relu ( f(x) = alpha * x for x < 0, f(x) = x for x >= 0 ). It is applicable only if actfun equals to 'relu'
seed	a numeric value specifying the random seed. Defaults to 1
verbose	a boolean. If TRUE then information will be printed in the console

Details

The input matrix should be of type numeric. This means the user should convert any character, factor or boolean columns to numeric values before using the elm_train function

References

http://arma.sourceforge.net/docs.html

https://en.wikipedia.org/wiki/Moore

https://www.kaggle.com/robertbm/extreme-learning-machine-example

http://rt.dgyblog.com/ml/ml-elm.html

Examples

library(elmNNRcpp)

#-----------
# Regression
#-----------

data(Boston, package = 'KernelKnn')

Boston = as.matrix(Boston)
dimnames(Boston) = NULL

x = Boston[, -ncol(Boston)]
y = matrix(Boston[, ncol(Boston)], nrow = length(Boston[, ncol(Boston)]), ncol = 1)

out_regr = elm_train(x, y, nhid = 20, actfun = 'purelin', init_weights = 'uniform_negative')


#---------------
# Classification
#---------------

data(ionosphere, package = 'KernelKnn')

x_class = ionosphere[, -c(2, ncol(ionosphere))]
x_class = as.matrix(x_class)
dimnames(x_class) = NULL

y_class = as.numeric(ionosphere[, ncol(ionosphere)])

y_class_onehot = onehot_encode(y_class - 1)     # class labels should begin from 0

out_class = elm_train(x_class, y_class_onehot, nhid = 20, actfun = 'relu')

---

One-hot-encoding of the labels in case of classification

Description

One-hot-encoding of the labels in case of classification

Usage

onehot_encode(y)

Arguments

y	a numeric vector consisting of the response variable labels. The minimum value of the unique labels should begin from 0

Examples

library(elmNNRcpp)

y = sample(0:3, 100, replace = TRUE)

y_expand = onehot_encode(y)

---

Predict with elm

Description

Wrapper for elm_predict.

Usage

## S3 method for class 'elm'
predict(object, newdata, type = c("class", "prob", "raw"), ...)

Arguments

object	elm model fitted with elm.
newdata	data.frame with the new data
type	only used with classification, can be either "class", "prob", "raw", which are class (vector), probability (matrix) or the output of the elm function (matrix).
...	not used

Value

predicted values

---