ViTLP

This repository releases code of paper Visually Guided Generative Text-Layout Pre-training for Document Intelligence (NAACL-2024).

We provide the pre-trained checkpoint ViTLP-medium (380M). The pre-trained ViTLP model can natively perform OCR text localization and recognition, which is accessible at Huggingface. Clone (or download) ViTLP checkpoint weight to the directory ./ckpts/ViTLP-medium.

<pre><code>git clone ViTLP && cd ViTLP pip install -r requirements.txt # Clone ViTLP-medium checkpoint mkdir -p ckpts/ViTLP-medium git clone https://huggingface.co/veason/ViTLP-medium ckpts/ViTLP-medium</code></pre>

Demo

With the checkpoint and dependencies set (see requirements.txt), run the demo as

<pre><code>python ocr.py</code></pre>

Upload a document image and have a shot.

See detailed inference code at decode.py and run batch decode by

<pre><code>bash decode.sh</code></pre>

Fine-tuning ViTLP

Please refer to ./finetuning for post-training on OCR datasets and fine-tuning on VQA datasets.

We also release a tool for synthesizing documents with grounding-box metadata at ./finetuning/SynthDog-bbox.

Preset FAQ

• Why is ViTLP-medium (380M)?

  When I commenced this project, it was on the eve of LLMs (precisely speaking, ChatGPT). ViTLP-base presented in our paper, is actually a rather small pre-trained model. We know it is expected to scale up ViTLP in this LLM era. However, the pre-training scale is commonly constrained by computation resources and the pre-training dataset scale, in which context ViTLP-medium (380M) is the largest pre-training scale so far we can support.

  Besides, this scale of ViTLP also brings inference sweetness including speed and memory usage. Typically, OCR on a page of a document image can be processed within 5~10 seconds in an Nvidia 4090, which is comparable to (and faster than) most OCR engines (and LLMs).

Citation

@inproceedings{mao-etal-ViTLP,
    title = "Visually Guided Generative Text-Layout Pre-training for Document Intelligence",
    author = "Mao, Zhiming  and
              Bai, Haoli  and
              Hou, Lu  and
              Shang, Lifeng  and
              Jiang, Xin  and
              Liu, Qun  and
              Wong, Kam-Fai",
    editor = "Duh, Kevin  and
              Gomez, Helena  and
              Bethard, Steven",
    booktitle = "Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
    month = jun,
    year = "2024",
    address = "Mexico City, Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.naacl-long.264",
    pages = "4713--4730"
}

Note

ViTLP is pronounced /ˈvai·tlp/ (vital). The first version of our paper was submitted to OpenReview in June 2023.