I started working on this package around 6-7 months ago to solve a practical problem I kept running into: repeatedly writing the same torch::nn_module() boilerplate for similar neural network architecture components when using {torch} package.
The goal was to wrap common {torch} NN architecture patterns into reusable, higher-level abstractions using code/expression generation (generating common torch::nn_module() pattern is the first goal of this package it came to my mind). This also has its objective to cleanly integrate with titanic ML frameworks, namely {tidymodels} (and {mlr3} soon). The package focuses on reducing boilerplate while still exposing model components like activations and optimizers in a way that can be tuned rather than hidden. There’s still room for the package to grow, particularly around time series support, which is where most of my work with neural networks is focused, as well as my goal.