dnnet_block: Blocked Feedforward Deep Neural Nets

In SkadiEye/deepTL: Deep Treatment Learning

Blocked Feedforward Deep Neural Nets

Description

Fit a Blocked Feedforward Deep Neural Network Model for Regression or Classification

Usage

dnnet_block(
  train,
  validate = NULL,
  load.param = FALSE,
  initial.param = NULL,
  norm.x = TRUE,
  norm.y = ifelse(is.factor(train@y), FALSE, TRUE),
  activate = "relu",
  n.hidden = list(dim(train@x)[2], 10, 10),
  learning.rate = ifelse(learning.rate.adaptive %in% c("adam"), 0.001, 0.01),
  l1.reg = 0,
  l2.reg = 0,
  n.batch = 100,
  n.epoch = 100,
  early.stop = ifelse(is.null(validate), FALSE, TRUE),
  early.stop.det = 1000,
  plot = FALSE,
  accel = c("rcpp", "none")[1],
  learning.rate.adaptive = c("constant", "adadelta", "adagrad", "momentum", "adam")[5],
  rho = c(0.9, 0.95, 0.99, 0.999)[ifelse(learning.rate.adaptive == "momentum", 1, 3)],
  epsilon = c(10^-10, 10^-8, 10^-6, 10^-4)[2],
  beta1 = 0.9,
  beta2 = 0.999,
  loss.f = ifelse(is.factor(train@y), "logit", "mse")
)

Arguments

train	A dnnetInput object, the training set.
validate	A dnnetInput object, the validation set, optional.
load.param	Whether initial parameters are loaded into the model.
initial.param	The initial parameters to be loaded.
norm.x	A boolean variable indicating whether to normalize the input matrix.
norm.y	A boolean variable indicating whether to normalize the response (if continuous).
activate	Activation Function. One of the following, "sigmoid", "tanh", "relu", "prelu", "elu", "celu".
n.hidden	A list of numeric vectors for the blocked hidden structures, starting from the input layer.
learning.rate	Initial learning rate, 0.001 by default; If "adam" is chosen as an adaptive learning rate adjustment method, 0.1 by defalut.
l1.reg	weight for l1 regularization, optional.
l2.reg	weight for l2 regularization, optional.
n.batch	Batch size for batch gradient descent.
n.epoch	Maximum number of epochs.
early.stop	Indicate whether early stop is used (only if there exists a validation set).
early.stop.det	Number of epochs of increasing loss to determine the early stop.
plot	Indicate whether to plot the loss.
accel	"rcpp" to use the Rcpp version and "none" (default) to use the R version for back propagation.
learning.rate.adaptive	Adaptive learning rate adjustment methods, one of the following, "constant", "adadelta", "adagrad", "momentum", "adam".
rho	A parameter used in momentum.
epsilon	A parameter used in Adagrad and Adam.
beta1	A parameter used in Adam.
beta2	A parameter used in Adam.
loss.f	Loss function of choice.

Value

Returns a DnnModelObj object.

See Also

dnnet-class
dnnetInput-class
actF

SkadiEye/deepTL documentation built on Nov. 17, 2022, 1:41 p.m.