RapidLatexOCR ONNX weights (re-hosted)
This repository is a re-hosted, unmodified copy of the ONNX model weights from
RapidAI/RapidLatexOCR, used to
recognize handwritten/typeset math formulas and convert them to LaTeX, fully
client-side in the browser via onnxruntime-web.
No training, fine-tuning, or modification has been done to these weights. They are mirrored here only so we can fetch them from a stable, CORS-enabled, pinned URL without committing ~179 MB of binary files to our own git repository.
Files
| File | Description |
|---|---|
image_resizer.onnx |
Predicts the input size bucket for a formula crop |
encoder.onnx |
Vision transformer encoder |
decoder.onnx |
Autoregressive LaTeX token decoder |
tokenizer.json |
Hugging Face tokenizers BPE vocabulary for the LaTeX token stream |
Attribution
All credit for training and exporting these weights belongs to the original authors:
- RapidAI/RapidLatexOCR — packaged the trained model as ONNX and published the release these weights are copied from.
- lukas-blecher/LaTeX-OCR (pix2tex) — the original model architecture and training this work is built on.
License
MIT, matching the license of both upstream projects. See LICENSE in this repo.
Usage
These weights are consumed directly by web-app I am working on. They are not intended to be a general-purpose published model — for the canonical, actively maintained version, use the upstream RapidLatexOCR repository or PyPI package instead.