Create MLP Learner via mlr3torch — create_mlp

Creates a Multi-Layer Perceptron (MLP) learner for classification tasks. Uses mlr3torch for native torch integration within the mlr3 ecosystem.

Usage

create_mlp_learner(
  hidden_layers = c(64, 32),
  dropout = 0.2,
  activation = c("relu", "tanh", "sigmoid"),
  epochs = 100,
  batch_size = 32,
  lr = 0.001,
  early_stopping = TRUE,
  patience = 10,
  seed = NULL,
  device = "cpu"
)

Arguments

hidden_layers: Integer vector specifying neurons per hidden layer. Default: c(64, 32) for a 2-layer network.
dropout: Dropout rate between 0 and 1. Default: 0.2.
activation: Activation function: "relu" (default), "tanh", "sigmoid".
epochs: Maximum training epochs. Default: 100.
batch_size: Training batch size. Default: 32.
lr: Learning rate. Default: 0.001.
early_stopping: Logical, use early stopping. Default: TRUE.
patience: Early stopping patience (epochs). Default: 10.
seed: Random seed for reproducibility.
device: "cpu" or "cuda" for GPU. Default: "cpu".

Value

An mlr3 Learner object (or a fallback if mlr3torch unavailable)

Details

The MLP architecture: - Input layer: matches number of features - Hidden layers: as specified, with batch normalization and dropout - Output layer: 2 neurons (binary classification) with softmax

If mlr3torch is not installed, returns a fallback learner (glmnet) with a warning.

Examples

if (FALSE) { # \dontrun{
# Basic MLP
mlp <- create_mlp_learner()

# Deeper network with more regularization
mlp_deep <- create_mlp_learner(
  hidden_layers = c(128, 64, 32),
  dropout = 0.3,
  epochs = 200
)

# Use in OmicPipeline
pipeline <- OmicPipeline$new(data, target = "outcome")
learner <- pipeline$create_graph_learner(
  filter = "anova",
  model = mlp,  # Pass the MLP learner
  n_features = 50
)
} # }