survivalmodels (version 0.1.9)

build_pytorch_net: Build a Pytorch Multilayer Perceptron

Description

Utility function to build an MLP with a choice of activation function and weight initialization with optional dropout and batch normalization.

Usage

build_pytorch_net(
  n_in,
  n_out,
  nodes = c(32, 32),
  activation = "relu",
  act_pars = list(),
  dropout = 0.1,
  bias = TRUE,
  batch_norm = TRUE,
  batch_pars = list(eps = 1e-05, momentum = 0.1, affine = TRUE),
  init = "uniform",
  init_pars = list()
)

Arguments

n_in

(integer(1)) Number of input features.

n_out

(integer(1)) Number of targets.

nodes

(numeric()) Hidden nodes in network, each element in vector represents number of hidden nodes in respective layer.

activation

(character(1)|list()) Activation function, can either be a single character and the same function is used in all layers, or a list of length length(nodes). See get_pycox_activation for options.

act_pars

(list()) Passed to get_pycox_activation.

dropout

(numeric()) Optional dropout layer, if NULL then no dropout layer added otherwise either a single numeric which will be added to all layers or a vector of differing drop-out amounts.

bias

(logical(1)) If TRUE (default) then a bias parameter is added to all linear layers.

batch_norm

(logical(1)) If TRUE (default) then batch normalisation is applied to all layers.

batch_pars

(list()) Parameters for batch normalisation, see reticulate::py_help(torch$nn$BatchNorm1d).

init

(character(1)) Weight initialization method. See get_pycox_init for options.

init_pars

(list()) Passed to get_pycox_init.

Details

This function is a helper for R users with less Python experience. Currently it is limited to simple MLPs. More advanced networks will require manual creation with reticulate.

Examples

# NOT RUN {
if (requireNamespaces("reticulate")) {
  build_pytorch_net(4L, 2L, nodes = c(32, 64, 32), activation = "selu")

  # pass parameters to activation and initializer functions
  build_pytorch_net(4L, 2L, activation = "elu", act_pars = list(alpha = 0.1),
  init  = "kaiming_uniform", init_pars = list(mode = "fan_out"))
}
# }
# NOT RUN {
# }