realmlp

This repo is transposed from the standalone, default tuned, simplified, RealMLP implementation with Python found here: https://github.com/dholzmueller/realmlp-td-s_standalone/tree/main. This is a pure R torch implementation, without reticulate or Python dependencies, only R & torch are needed.

Installation

You can install the development version of realmlp like so:

pak::pak("frankiethull/realmlp")

Example

library(realmlp)
library(torch)
library(rsample)

corn_data <- maize::corn_data

corn_splits <- initial_validation_split(corn_data)
train <- training(corn_splits)
validate <- validation(corn_splits)
test <- testing(corn_splits)

Regression

This is a basic example which shows you how to solve a common problem:

set.seed(42)
reg <- Standalone_RealMLP_TD_S_Regressor$new(device = "cpu")

reg$fit(
  X = train |> dplyr::select(-height), 
  y = train |> dplyr::pull(height), 
  X_val = validate |> dplyr::select(-height), 
  y_val = validate |> dplyr::pull(height)
)

# predictions
yhat <- reg$predict(test |> dplyr::select(-height))

Classification

set.seed(123)
clf <- Standalone_RealMLP_TD_S_Classifier$new(device = "cpu")

clf$fit(
  X = train |> dplyr::select(-type), 
  y = train |> dplyr::pull(type), 
  X_val = validate |> dplyr::select(-type), 
  y_val = validate |> dplyr::pull(type)
)

# predict classes
y_pred <- clf$predict(test |> dplyr::select(-type))

# predict probabilities
probs <- clf$predict_proba(test |> dplyr::select(-type))

To-Do’s

• create package & implement RealMLP-TD-S based on dholzmueller/realmlp-td-s_standalone
• export Mish
• add TD dictionaries from pytabkit
• implement RealMLP-TD from pytabkit
• register preprocessors as recipes
• register RealMLP-TD-S & RealMLP-TD to parsnip as engines
• add bagging option for RealMLP engine via cv
• add dials for hyperparameter optimization (RealMLP-HPO)
• create Caruana ensemble extension for stacks
• add tests for Python vs R results
• note subtle differences (such as sample vs population maths in numpy.std vs base::sd())