pix2tex 0.1.2

pix2tex: Using a ViT to convert images of equations into LaTeX code.

pix2tex - LaTeX OCR

The goal of this project is to create a learning based system that takes an image of a math formula and returns corresponding LaTeX code.

Using the model

To run the model you need Python 3.7+

If you don't have PyTorch installed. Follow their instructions here.

Install the package pix2tex:

pip install pix2tex[gui]

Model checkpoints will be downloaded automatically.

There are three ways to get a prediction from an image.

1. You can use the command line tool by calling pix2tex. Here you can parse already existing images from the disk and images in your clipboard.

2. Thanks to @katie-lim, you can use a nice user interface as a quick way to get the model prediction. Just call the GUI with latexocr. From here you can take a screenshot and the predicted latex code is rendered using MathJax and copied to your clipboard.

   Under linux, it is possible to use the GUI with gnome-screenshot which comes with multiple monitor support if gnome-screenshot was installed beforehand.

   

   If the model is unsure about the what's in the image it might output a different prediction every time you click "Retry". With the temperature parameter you can control this behavior (low temperature will produce the same result).

3. You can use an API. This has additional dependencies. Install via pip install -U pix2tex[api] and run

   python -m pix2tex.api.run
   

   to start a Streamlit demo that connects to the API at port 8502. There is also a docker image available for the API: https://hub.docker.com/r/lukasblecher/pix2tex

   docker pull lukasblecher/pix2tex:api
   docker run --rm -p 8502:8502 lukasblecher/pix2tex:api
   

   To also run the streamlit demo run

   docker run --rm -it -p 8501:8501 --entrypoint python lukasblecher/pix2tex:api pix2tex/api/run.py
   

   and navigate to http://localhost:8501/

4. Use from within Python

   from PIL import Image
   from pix2tex.cli import LatexOCR
   
   img = Image.open('path/to/image.png')
   model = LatexOCR()
   print(model(img))
   

The model works best with images of smaller resolution. That's why I added a preprocessing step where another neural network predicts the optimal resolution of the input image. This model will automatically resize the custom image to best resemble the training data and thus increase performance of images found in the wild. Still it's not perfect and might not be able to handle huge images optimally, so don't zoom in all the way before taking a picture.

Always double check the result carefully. You can try to redo the prediction with an other resolution if the answer was wrong.

Want to use the package?

I'm trying to compile a documentation right now.

Visit here: https://pix2tex.readthedocs.io/

Training the model

Install a couple of dependencies pip install pix2tex[train].

1. First we need to combine the images with their ground truth labels. I wrote a dataset class (which needs further improving) that saves the relative paths to the images with the LaTeX code they were rendered with. To generate the dataset pickle file run

python -m pix2tex.dataset.dataset --equations path_to_textfile --images path_to_images --out dataset.pkl

To use your own tokenizer pass it via --tokenizer (See below).

You can find my generated training data on the Google Drive as well (formulae.zip - images, math.txt - labels). Repeat the step for the validation and test data. All use the same label text file.

2. Edit the data (and valdata) entry in the config file to the newly generated .pkl file. Change other hyperparameters if you want to. See pix2tex/model/settings/config.yaml for a template.
3. Now for the actual training run

python -m pix2tex.train --config path_to_config_file

If you want to use your own data you might be interested in creating your own tokenizer with

python -m pix2tex.dataset.dataset --equations path_to_textfile --vocab-size 8000 --out tokenizer.json

Don't forget to update the path to the tokenizer in the config file and set num_tokens to your vocabulary size.

Model

The model consist of a ViT [1] encoder with a ResNet backbone and a Transformer [2] decoder.

Performance

BLEU score	normed edit distance	token accuracy
0.88	0.10	0.60

Data

We need paired data for the network to learn. Luckily there is a lot of LaTeX code on the internet, e.g. wikipedia, arXiv. We also use the formulae from the im2latex-100k [3] dataset. All of it can be found here

Dataset Requirements

In order to render the math in many different fonts we use XeLaTeX, generate a PDF and finally convert it to a PNG. For the last step we need to use some third party tools:

• XeLaTeX
• ImageMagick with Ghostscript. (for converting pdf to png)
• Node.js to run KaTeX (for normalizing Latex code)
• Python 3.7+ & dependencies (specified in setup.py)

Fonts

Latin Modern Math, GFSNeohellenicMath.otf, Asana Math, XITS Math, Cambria Math

TODO

• add more evaluation metrics
• create a GUI
• add beam search
• support handwritten formulae (kinda done, see training colab notebook)
• reduce model size (distillation)
• find optimal hyperparameters
• tweak model structure
• fix data scraping and scrape more data
• trace the model (#2)

Contribution

Contributions of any kind are welcome.

Acknowledgment

Code taken and modified from lucidrains, rwightman, im2markup, arxiv_leaks, pkra: Mathjax, harupy: snipping tool

References

[1] An Image is Worth 16x16 Words

[2] Attention Is All You Need

[3] Image-to-Markup Generation with Coarse-to-Fine Attention