Spedon/texify-quantized-onnx

texify-quantized-onnx

https://huggingface.co/vikp/texify with quantized ONNX weights, shoutout to https://huggingface.co/Xenova/texify

Usage (optimum[onnxruntime])

If you haven't already, you can install the optimum with the onnxrumtime backend

pip install "optimum[onnxruntime]"

Example:

from optimum.onnxruntime import ORTModelForVision2Seq
from optimum.pipelines import pipeline

model = ORTModelForVision2Seq.from_pretrained("Spedon/texify-quantized-onnx")
texify = pipeline(
    "image-to-text",
    model,
    feature_extractor="Spedon/texify-quantized-onnx",
    image_processor="Spedon/texify-quantized-onnx",
)
image = (
    "https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/latex.png"
)
latex = texify(image, max_new_tokens=384)
print(latex)
# [{'generated_text': "The potential $V_i$ of cell $\\mathcal{C}_i$ centred at position $\\mathbf{r}_i$ is related to the surface charge densities $\\sigma_j$ of cells $\\mathcal{C}_j$ $j\\in[1,N]$ through the superposition principle as: $$V_i\\,=\\,\\sum_{j=0}^{N}\\,\\frac{\\sigma_j}{4\\pi\\varepsilon_0}\\,\\int_{\\mathcal{C}_j}\\frac{1}{\\|\\mathbf{r}_i-\\mathbf{r}'\\|}\\,\\mathrm{d}^2\\mathbf{r}'\\,=\\,\\sum_{j=0}^{N}\\,Q_{ij}\\,\\sigma_j,$$ where the integral over the surface of cell $\\mathcal{C}_j$ only depends on $\\mathcal{C}_j$ shape and on the relative position of the target point $\\mathbf{r}_i$ with respect to $\\mathcal{C}_j$ location, as $\\sigma_j$ is assumed constant over the whole surface of cell $\\mathcal{C}_j$. "}]

Input image	Visualized output