texifast

LaTeX and markdown OCR powered by texify, without bloated dependencies like torch or transformers.

Features

• Minimal dependency graph
• Compared to Optimum, texifast is faster (~20%) and has a smaller memory footprint (~20%). For details, see benchmark.
• Supports IOBinding features of ONNXRuntime and optimizes for CUDAExecutionProvider.
• Supports quantized/mixed precision models.

Installation

You must implicitly specify the required dependencies.

pip install texifast[cpu]
# or if you want to use CUDAExecutionProvider
pip install texifast[gpu]

⚠️⚠️⚠️

Do not install with pip install texifast !!!

Quickstart

This quick start use the image in test folder, you can use whatever you like.

from texifast.model import TxfModel
from texifast.pipeline import TxfPipeline

model = TxfModel(
    encoder_model_path="./encoder_model_quantized.onnx",
    decoder_model_path="./decoder_model_merged_quantized.onnx",
)
texifast = TxfPipeline(model=model, tokenizer="./tokenizer.json")
print(texifast("./latex.png"))

You can download the quantized ONNX model here and the FP16 ONNX model here.

API

The full Python API documentation can be found here.

Credits

• https://github.com/VikParuchuri/texify
• https://github.com/MosRat/MixTex-rs
• https://github.com/xenova/transformers.js
• https://onnxruntime.ai/docs/api/python/api_summary.html
• https://github.com/ml-tooling/lazydocs