Rapid LaTeX OCR

Rapid LaTeX OCR is a deployment-optimized version of LaTeX-OCR. It uses ONNXRuntime or OpenVINO for inference with models converted to ONNX format.

Key distinctions:

• Inference only: This repository contains simplified reasoning code for faster and easier deployment. It does not contain training code.
• Model format: All models are provided in ONNX format.
• Training: If you need to train a custom model, you must use the original LaTeX-OCR repository.

Install the package using pip. Note that the model files will be automatically downloaded and placed in a models directory within the installation path. If you encounter slow download speeds, you can download the models manually via Google Drive or Baidu NetDisk.

pip install rapid_latex_ocr

To use the OCR functionality within your Python code, import the LaTeXOCR class. You can pass raw image bytes to the model instance. The model returns a tuple containing the predicted LaTeX string and the inference elapsed time.

from rapid_latex_ocr import LaTeXOCR

model = LaTeXOCR()

img_path = "tests/test_files/6.png"
with open(img_path, "rb") as f:
    data = f.read()

res, elapse = model(data)

print(res)
print(elapse)

You can run the OCR tool directly from the terminal by passing the path to an image file as an argument. The tool will output the predicted LaTeX string followed by the inference time.

$ rapid_latex_ocr tests/test_files/6.png

# Output example:
# {\frac{x^{2}}{a^{2}}}-{\frac{y^{2}}{b^{2}}}=1
# 0.47902780000000034

The LaTeXOCR class is callable. You can pass an image path or an image object directly to the instance to perform OCR.

Call Signature

__call__(self, img: InputType) -> Tuple[str, float]

• img: An InputType (typically a file path string or a compatible image object).
• Returns: A tuple containing:
  • pred (str): The predicted LaTeX string.
  • elapse (float): The time taken for the inference in seconds.

Example

from rapid_latex_ocr.main import LaTeXOCR

engine = LaTeXOCR()
latex_string, duration = engine("path/to/formula_image.png")
print(f"LaTeX: {latex_string}")
print(f"Time: {duration:.4f}s")

You can run the OCR engine directly from the terminal by providing an image path. You can also optionally specify paths to specific model files.

Usage

python -m rapid_latex_ocr.main <img_path> [options]

Arguments

• img_path (positional): Path to the image of the formula.
• -img_resizer, --image_resizer_path: Path to the image resizer model.
• -encdoer, --encoder_path: Path to the encoder model.
• -decoder, --decoder_path: Path to the decoder model.
• -tokenizer, --tokenizer_json: Path to the tokenizer JSON file.

Example

python -m rapid_latex_ocr.main my_formula.png

The LaTeXOCRInput dataclass defines the hyperparameters and constraints used by the engine. These values are typically loaded from a config.yaml file during LaTeXOCR initialization.

Fields