The goal of this project is to write a character classifier which outputs the command needed in Typst to generate the drawn symbol. The project is currently under active development and not functional yet.
In order to build this project, you will need to have Rust, Typst, ImageMagick and Node as well as NPM installed.
The generation of the training images is done by running make databuild which will download the
relevant Rust file from the Typst project, extract the symbol names from it, generate a JSON file
and the training images within the folder data and save the icons for the frontend in
frontend/static/symbols.
To train the neural network in the training directory, you will need to have
JAX with proper GPU support installed. Moreover, this project uses
NumPy, Matplotlib, OpenCV,
Optax and Flax which
you can install via PIP.
pip install numpy matplotlib opencv-python optax flaxThe file training/dataload.py writes the training data to the file training/numpydata.pkl, the
file training/train.py contains the training loop and training/datawrite.py writes the
model checkpoint into a binary file. All these scripts can be run in order using make training.
To generate the WASM binary as well as the glue code, you will need to install wasm-bingen-cli
using cargo install, make sure its version matches the one in inference/Cargo.toml and run
make inference. Finally, in the frontend directory, you can install the Node packages using
npm install, start the development server using npm run dev and build the project as a static
site using npm run build.
This project is licensed under the MIT License.