Yet another LaTex OCR Project written in PyTorch, based on ConvNeXt and Transformer.

This project is the backend of CeleryMath.

Give us a star if this project helps you 🤗

Any further developments and contributions are welcome 😄

Follow the following instructions to train by yourself:

1. Create virtual environment.

poetry install
poetry shell

2. Create dataset.

You can download generated dataset from here (2.05G) or generate by yourself with the following code, the generation may be slow:

python -m src.utils.latex2png -i dataset/data/full_math.txt -w dataset/data/full_set -b 1

3. Edit config file.

Edit the src/config/config_convnext.json and replace the dataset path to yours.

4. Run training

python -m src.train

If you have your own latex formula dataset, you can add them to dataset/data/full_math.txt and regenerate tokenizers and images.

1. tokenizers from hugging face was used, if you want to change formula file and output file location, edit src/dataset.py

python -m src.dataset

2. generate dataset, TexLive or MikeTex or similar program must be installed.

python -m src.utils.latex2png -i dataset/data/full_math.txt -w dataset/data/full_set -b 1

Open an issue if you have any questions or a PR if you can fix it.

• API
• Desktop Deploy, see CeleryMath
• ONNX
• Use pytorch-lightning to manage training and evaluation

This project was inspired by the following project, and some methods or codes were also borrowed from them, THANKS A LOT!!! 🤝

1. LaTex-OCR, LICENSE: MIT, https://github.com/lukas-blecher/LaTeX-OCR
2. LaTeX_OCR_PRO, LICENSE: GPL-3.0, https://github.com/LinXueyuanStdio/LaTeX_OCR_PRO

GPL-3.0, details here