densemlp

densemlp provides dense multilayer perceptron models for tabular regression and classification in R. The package is built around fully connected feed-forward neural network layers, with optional extensions such as dropout, batch normalization, residual connections, gated blocks, and input projection.

These extensions enrich the basic dense MLP architecture without changing its model class: the models remain dense feed-forward neural networks, not convolutional, recurrent, transformer-based, or tree-based models.

task is optional in the main API. When it is set to "auto" or omitted, the task is inferred from the outcome. Use it only when you need to override that inference.

Why densemlp?

The name densemlp reflects the package design: dense, fully connected neural network blocks for tabular data, extended with optional modern components when needed.

Install it from GitHub with:

remotes::install_github("ielbadisy/densemlp")

For local development without reinstalling after each edit:

pkgload::load_all(".")

Do not source individual files such as R/densemlp.R; the exported functions depend on helpers that are loaded through the package namespace.

Features

• Binary classification, multiclass classification, and regression.
• Formula-based interface.
• Automatic preprocessing for missing numeric values, missing categorical values, and unseen prediction-time categories.
• Configurable hidden layers, activation functions, dropout, batch normalization, residual blocks, gated blocks, and input projection.
• Early stopping, learning-rate schedules, and common torch optimizers.
• Task-aware predictions, metrics, training-history plots, tuning, and permutation importance.

Classification Example

library(densemlp)

fit <- densemlp(
  Species ~ .,
  data = iris,
  epochs = 10,
  validation = 0.2,
  verbose = TRUE,
  seed = 1
)

## Training dense multilayer perceptron
## Task: multiclass classification
## Optimizer: Adam
## Learning rate: 0.001
## Epochs: 10
## Batch size: 32
## Epoch 01/10 | train_loss: 1.2860 | valid_loss: 1.3458 | valid_acc: 0.0333
## Epoch 02/10 | train_loss: 1.0189 | valid_loss: 1.0308 | valid_acc: 0.4667
## Epoch 03/10 | train_loss: 0.8334 | valid_loss: 0.7775 | valid_acc: 0.7333
## Epoch 04/10 | train_loss: 0.7153 | valid_loss: 0.6115 | valid_acc: 0.9000
## Epoch 05/10 | train_loss: 0.6196 | valid_loss: 0.5043 | valid_acc: 0.9000
## Epoch 06/10 | train_loss: 0.4918 | valid_loss: 0.4336 | valid_acc: 0.9000
## Epoch 07/10 | train_loss: 0.4702 | valid_loss: 0.3883 | valid_acc: 0.8667
## Epoch 08/10 | train_loss: 0.4005 | valid_loss: 0.3524 | valid_acc: 0.9000
## Epoch 09/10 | train_loss: 0.3574 | valid_loss: 0.3257 | valid_acc: 0.9000
## Epoch 10/10 | train_loss: 0.3554 | valid_loss: 0.3058 | valid_acc: 0.9000

predict(fit, iris[1:5, ], type = "class")

## [1] setosa setosa setosa setosa setosa
## Levels: setosa versicolor virginica

predict(fit, iris[1:5, ], type = "prob")

##         setosa  versicolor   virginica
## [1,] 0.9823542 0.009474045 0.008171768
## [2,] 0.9266943 0.051082314 0.022223408
## [3,] 0.9656487 0.022011787 0.012339532
## [4,] 0.9343854 0.044696244 0.020918327
## [5,] 0.9836846 0.009770629 0.006544733

pred <- predict(fit, iris, type = "class")
densemlp_metrics(iris$Species, pred)

## $accuracy
## [1] 0.9466667

Regression Example

fit_reg <- densemlp(
  mpg ~ disp + hp + wt,
  data = mtcars,
  epochs = 10,
  validation = 0.2,
  verbose = FALSE,
  seed = 2
)

pred_reg <- predict(fit_reg, mtcars, type = "response")
densemlp_metrics(mtcars$mpg, pred_reg)

## $rmse
## [1] 5.818005
## 
## $mae
## [1] 5.026398
## 
## $rsq
## [1] 0.03807414

Tuning

tuned <- tune_densemlp(
  Species ~ .,
  data = iris,
  grid = list(
    hidden_units = list(c(8), c(16, 8)),
    activation = c("relu"),
    dropout = c(0, 0.1),
    lr = c(1e-3),
    epochs = c(10)
  ),
  patience = 3,
  seed = 1,
  verbose = FALSE
)

tuned$results

##   hidden_units dropout activation batch_norm residual gated input_projection
## 1         16-8       0       relu       TRUE    FALSE FALSE               NA
## 2         16-8     0.1       relu       TRUE    FALSE FALSE               NA
## 3            8       0       relu       TRUE    FALSE FALSE               NA
## 4            8     0.1       relu       TRUE    FALSE FALSE               NA
##   epochs batch_size    lr optimizer lr_schedule weight_decay            loss
## 1     10         32 0.001      adam      cosine            0 bce_with_logits
## 2     10         32 0.001      adam      cosine            0 bce_with_logits
## 3     10         32 0.001      adam      cosine            0 bce_with_logits
## 4     10         32 0.001      adam      cosine            0 bce_with_logits
##   label_smoothing focal_gamma   metric     score   score_sd repeats
## 1               0           2 accuracy 0.4777778 0.23648663       3
## 2               0           2 accuracy 0.4555556 0.10715168       3
## 3               0           2 accuracy 0.3666667 0.43333333       3
## 4               0           2 accuracy 0.2777778 0.01924501       3

tuned$best_fit

## <densemlp_fit>
## Task: classification
## Outcome levels: setosa, versicolor, virginica
## Encoded features: 4
## Best epoch: 8

Interpretation

plot_history(fit)

ggplot2::autoplot(fit)

importance <- perm_importance(fit, iris[, -5], iris$Species)
plot(importance)

Status

Version 0.5.0 supports binary classification, multiclass classification, and regression for tabular data frames with automatic preprocessing.