Nougat: Neural Optical Understanding for Academic Documents

This is the official repository for Nougat, the academic document PDF parser that understands LaTeX math and tables.

Project page: https://facebookresearch.github.io/nougat/

Install

From pip:

pip install nougat-ocr

From repository:

pip install git+https://github.com/facebookresearch/nougat

Note, on Windows: If you want to utilize a GPU, make sure you first install the correct PyTorch version. Follow instructions here

There are extra dependencies if you want to call the model from an API or generate a dataset. Install via

pip install "nougat-ocr[api]" or pip install "nougat-ocr[dataset]"

Get prediction for a PDF

CLI

To get predictions for a PDF run

$ nougat path/to/file.pdf -o output_directory

A path to a directory or to a file where each line is a path to a PDF can also be passed as a positional argument

$ nougat path/to/directory -o output_directory

usage: nougat [-h] [--batchsize BATCHSIZE] [--checkpoint CHECKPOINT] [--model MODEL] [--out OUT]
              [--recompute] [--markdown] [--no-skipping] pdf [pdf ...]

positional arguments:
  pdf                   PDF(s) to process.

options:
  -h, --help            show this help message and exit
  --batchsize BATCHSIZE, -b BATCHSIZE
                        Batch size to use.
  --checkpoint CHECKPOINT, -c CHECKPOINT
                        Path to checkpoint directory.
  --model MODEL_TAG, -m MODEL_TAG
                        Model tag to use.
  --out OUT, -o OUT     Output directory.
  --recompute           Recompute already computed PDF, discarding previous predictions.
  --full-precision      Use float32 instead of bfloat16. Can speed up CPU conversion for some setups.
  --no-markdown         Do not add postprocessing step for markdown compatibility.
  --markdown            Add postprocessing step for markdown compatibility (default).
  --no-skipping         Don't apply failure detection heuristic.
  --pages PAGES, -p PAGES
                        Provide page numbers like '1-4,7' for pages 1 through 4 and page 7. Only works for single PDFs.

The default model tag is 0.1.0-small. If you want to use the base model, use 0.1.0-base.

$ nougat path/to/file.pdf -o output_directory -m 0.1.0-base

In the output directory every PDF will be saved as a .mmd file, the lightweight markup language, mostly compatible with Mathpix Markdown (we make use of the LaTeX tables).

Note: On some devices the failure detection heuristic is not working properly. If you experience a lot of [MISSING_PAGE] responses, try to run with the --no-skipping flag. Related: #11, #67

API

With the extra dependencies you use app.py to start an API. Call

$ nougat_api

To get a prediction of a PDF file by making a POST request to http://127.0.0.1:8503/predict/. It also accepts parameters start and stop to limit the computation to select page numbers (boundaries are included).

The response is a string with the markdown text of the document.

curl -X 'POST' \
  'http://127.0.0.1:8503/predict/' \
  -H 'accept: application/json' \
  -H 'Content-Type: multipart/form-data' \
  -F 'file=@<PDFFILE.pdf>;type=application/pdf'

To use the limit the conversion to pages 1 to 5, use the start/stop parameters in the request URL: http://127.0.0.1:8503/predict/?start=1&stop=5

Dataset

Generate dataset

To generate a dataset you need

1. A directory containing the PDFs
2. A directory containing the .html files (processed .tex files by LaTeXML) with the same folder structure
3. A binary file of pdffigures2 and a corresponding environment variable export PDFFIGURES_PATH="/path/to/binary.jar"

Next run

python -m nougat.dataset.split_htmls_to_pages --html path/html/root --pdfs path/pdf/root --out path/paired/output --figure path/pdffigures/outputs

Additional arguments include

Finally create a jsonl file that contains all the image paths, markdown text and meta information.

python -m nougat.dataset.create_index --dir path/paired/output --out index.jsonl

For each jsonl file you also need to generate a seek map for faster data loading:

python -m nougat.dataset.gen_seek file.jsonl

The resulting directory structure can look as follows:

root/
├── images
├── train.jsonl
├── train.seek.map
├── test.jsonl
├── test.seek.map
├── validation.jsonl
└── validation.seek.map

Note that the .mmd and .json files in the path/paired/output (here images) are no longer required. This can be useful for pushing to a S3 bucket by halving the amount of files.

Training

To train or fine tune a Nougat model, run

python train.py --config config/train_nougat.yaml

Evaluation

Run

python test.py --checkpoint path/to/checkpoint --dataset path/to/test.jsonl --save_path path/to/results.json

To get the results for the different text modalities, run

python -m nougat.metrics path/to/results.json

FAQ

• Why am I only getting [MISSING_PAGE]?

  Nougat was trained on scientific papers found on arXiv and PMC. Is the document you're processing similar to that? What language is the document in? Nougat works best with English papers, other Latin-based languages might work. Chinese, Russian, Japanese etc. will not work. If these requirements are fulfilled it might be because of false positives in the failure detection, when computing on CPU or older GPUs (#11). Try passing the --no-skipping flag for now.

• Where can I download the model checkpoint from.

  They are uploaded here on GitHub in the release section. You can also download them during the first execution of the program. Choose the preferred preferred model by passing --model 0.1.0-{base,small}

Citation

@misc{blecher2023nougat,
      title={Nougat: Neural Optical Understanding for Academic Documents}, 
      author={Lukas Blecher and Guillem Cucurull and Thomas Scialom and Robert Stojnic},
      year={2023},
      eprint={2308.13418},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Acknowledgments

This repository builds on top of the Donut repository.

License

Nougat codebase is licensed under MIT.

Nougat model weights are licensed under CC-BY-NC.