π© : Document Understanding Transformer
Donut
Official Implementation of Donut and SynthDoG | Paper | Slide | Poster
Introduction
Donut
Our academic paper, which describes our method in detail and provides full experimental results and analyses, can be found here:
OCR-free Document Understanding Transformer.
Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park. In ECCV 2022.
Pre-trained Models and Web Demos
Gradio web demos are available! |
---|
- You can run the demo with
./app.py
file. - Sample images are available at
./misc
and more receipt images are available at CORD dataset link. - Web demos are available from the links in the following table.
- Note: We have updated the Google Colab demo (as of June 15, 2023) to ensure its proper working.
Task | Sec/Img | Score | Trained Model | Demo |
---|---|---|---|---|
CORD (Document Parsing) | 0.7 / 0.7 / 1.2 |
91.3 / 91.1 / 90.9 |
donut-base-finetuned-cord-v2 (1280) / donut-base-finetuned-cord-v1 (1280) / donut-base-finetuned-cord-v1-2560 |
gradio space web demo, google colab demo (updated at 23.06.15) |
Train Ticket (Document Parsing) | 0.6 | 98.7 | donut-base-finetuned-zhtrainticket | google colab demo (updated at 23.06.15) |
RVL-CDIP (Document Classification) | 0.75 | 95.3 | donut-base-finetuned-rvlcdip | gradio space web demo, google colab demo (updated at 23.06.15) |
DocVQA Task1 (Document VQA) | 0.78 | 67.5 | donut-base-finetuned-docvqa | gradio space web demo, google colab demo (updated at 23.06.15) |
The links to the pre-trained backbones are here:
donut-base
: trained with 64 A100 GPUs (~2.5 days), number of layers (encoder: {2,2,14,2}, decoder: 4), input size 2560x1920, swin window size 10, IIT-CDIP (11M) and SynthDoG (English, Chinese, Japanese, Korean, 0.5M x 4).donut-proto
: (preliminary model) trained with 8 V100 GPUs (~5 days), number of layers (encoder: {2,2,18,2}, decoder: 4), input size 2048x1536, swin window size 8, and SynthDoG (English, Japanese, Korean, 0.4M x 3).
Please see our paper for more details.
SynthDoG datasets
The links to the SynthDoG-generated datasets are here:
synthdog-en
: English, 0.5M.synthdog-zh
: Chinese, 0.5M.synthdog-ja
: Japanese, 0.5M.synthdog-ko
: Korean, 0.5M.
To generate synthetic datasets with our SynthDoG, please see ./synthdog/README.md
and our paper for details.
Updates
2023-06-15 We have updated all Google Colab demos to ensure its proper working.
2022-11-14 New version 1.0.9 is released (pip install donut-python --upgrade
). See 1.0.9 Release Notes.
2022-08-12 Donut π© is also available at huggingface/transformers donut-python
loads the pre-trained weights from the official
branch of the model repositories. See 1.0.5 Release Notes.
2022-08-05 A well-executed hands-on tutorial on donut
2022-07-20 First Commit, We release our code, model weights, synthetic data and generator.
Software installation
pip install donut-python
or clone this repository and install the dependencies:
git clone https://github.com/clovaai/donut.git
cd donut/
conda create -n donut_official python=3.7
conda activate donut_official
pip install .
We tested donut-python == 1.0.1 with:
- torch == 1.11.0+cu113
- torchvision == 0.12.0+cu113
- pytorch-lightning == 1.6.4
- transformers == 4.11.3
- timm == 0.5.4
Note: From several reported issues, we have noticed increased challenges in configuring the testing environment for donut-python
due to recent updates in key dependency libraries. While we are actively working on a solution, we have updated the Google Colab demo (as of June 15, 2023) to ensure its proper working. For assistance, we encourage you to refer to the following demo links: CORD Colab Demo, Train Ticket Colab Demo, RVL-CDIP Colab Demo, DocVQA Colab Demo.
Getting Started
Data
This repository assumes the following structure of dataset:
> tree dataset_name
dataset_name
βββ test
β βββ metadata.jsonl
β βββ {image_path0}
β βββ {image_path1}
β .
β .
βββ train
β βββ metadata.jsonl
β βββ {image_path0}
β βββ {image_path1}
β .
β .
βββ validation
βββ metadata.jsonl
βββ {image_path0}
βββ {image_path1}
.
.
> cat dataset_name/test/metadata.jsonl
{"file_name": {image_path0}, "ground_truth": "{\"gt_parse\": {ground_truth_parse}, ... {other_metadata_not_used} ... }"}
{"file_name": {image_path1}, "ground_truth": "{\"gt_parse\": {ground_truth_parse}, ... {other_metadata_not_used} ... }"}
.
.
- The structure of
metadata.jsonl
file is in JSON Lines text format, i.e.,.jsonl
. Each line consists offile_name
: relative path to the image file.ground_truth
: string format (json dumped), the dictionary contains eithergt_parse
orgt_parses
. Other fields (metadata) can be added to the dictionary but will not be used.
donut
interprets all tasks as a JSON prediction problem. As a result, alldonut
model training share a same pipeline. For training and inference, the only thing to do is preparinggt_parse
orgt_parses
for the task in format described below.
For Document Classification
The gt_parse
follows the format of {"class" : {class_name}}
, for example, {"class" : "scientific_report"}
or {"class" : "presentation"}
.
For Document Information Extraction
The gt_parse
is a JSON object that contains full information of the document image, for example, the JSON object for a receipt may look like {"menu" : [{"nm": "ICE BLACKCOFFEE", "cnt": "2", ...}, ...], ...}
.
- More examples are available at CORD dataset.
- Google colab demo is available here.
- Gradio web demo is available here.
For Document Visual Question Answering
The gt_parses
follows the format of [{"question" : {question_sentence}, "answer" : {answer_candidate_1}}, {"question" : {question_sentence}, "answer" : {answer_candidate_2}}, ...]
, for example, [{"question" : "what is the model name?", "answer" : "donut"}, {"question" : "what is the model name?", "answer" : "document understanding transformer"}]
.
- DocVQA Task1 has multiple answers, hence
gt_parses
should be a list of dictionary that contains a pair of question and answer. - Google colab demo is available here.
- Gradio web demo is available here.
For (Pseudo) Text Reading Task
The gt_parse
looks like {"text_sequence" : "word1 word2 word3 ... "}
- This task is also a pre-training task of Donut model.
- You can use our SynthDoG
πΆ to generate synthetic images for the text reading task with propergt_parse
. See./synthdog/README.md
for details.
Training
This is the configuration of Donut model training on CORD dataset used in our experiment. We ran this with a single NVIDIA A100 GPU.
python train.py --config config/train_cord.yaml \
--pretrained_model_name_or_path "naver-clova-ix/donut-base" \
--dataset_name_or_paths '["naver-clova-ix/cord-v2"]' \
--exp_version "test_experiment"
.
.
Prediction: <s_menu><s_nm>Lemon Tea (L)</s_nm><s_cnt>1</s_cnt><s_price>25.000</s_price></s_menu><s_total><s_total_price>25.000</s_total_price><s_cashprice>30.000</s_cashprice><s_changeprice>5.000</s_changeprice></s_total>
Answer: <s_menu><s_nm>Lemon Tea (L)</s_nm><s_cnt>1</s_cnt><s_price>25.000</s_price></s_menu><s_total><s_total_price>25.000</s_total_price><s_cashprice>30.000</s_cashprice><s_changeprice>5.000</s_changeprice></s_total>
Normed ED: 0.0
Prediction: <s_menu><s_nm>Hulk Topper Package</s_nm><s_cnt>1</s_cnt><s_price>100.000</s_price></s_menu><s_total><s_total_price>100.000</s_total_price><s_cashprice>100.000</s_cashprice><s_changeprice>0</s_changeprice></s_total>
Answer: <s_menu><s_nm>Hulk Topper Package</s_nm><s_cnt>1</s_cnt><s_price>100.000</s_price></s_menu><s_total><s_total_price>100.000</s_total_price><s_cashprice>100.000</s_cashprice><s_changeprice>0</s_changeprice></s_total>
Normed ED: 0.0
Prediction: <s_menu><s_nm>Giant Squid</s_nm><s_cnt>x 1</s_cnt><s_price>Rp. 39.000</s_price><s_sub><s_nm>C.Finishing - Cut</s_nm><s_price>Rp. 0</s_price><sep/><s_nm>B.Spicy Level - Extreme Hot Rp. 0</s_price></s_sub><sep/><s_nm>A.Flavour - Salt & Pepper</s_nm><s_price>Rp. 0</s_price></s_sub></s_menu><s_sub_total><s_subtotal_price>Rp. 39.000</s_subtotal_price></s_sub_total><s_total><s_total_price>Rp. 39.000</s_total_price><s_cashprice>Rp. 50.000</s_cashprice><s_changeprice>Rp. 11.000</s_changeprice></s_total>
Answer: <s_menu><s_nm>Giant Squid</s_nm><s_cnt>x1</s_cnt><s_price>Rp. 39.000</s_price><s_sub><s_nm>C.Finishing - Cut</s_nm><s_price>Rp. 0</s_price><sep/><s_nm>B.Spicy Level - Extreme Hot</s_nm><s_price>Rp. 0</s_price><sep/><s_nm>A.Flavour- Salt & Pepper</s_nm><s_price>Rp. 0</s_price></s_sub></s_menu><s_sub_total><s_subtotal_price>Rp. 39.000</s_subtotal_price></s_sub_total><s_total><s_total_price>Rp. 39.000</s_total_price><s_cashprice>Rp. 50.000</s_cashprice><s_changeprice>Rp. 11.000</s_changeprice></s_total>
Normed ED: 0.039603960396039604
Epoch 29: 100%|βββββββββββββ| 200/200 [01:49<00:00, 1.82it/s, loss=0.00327, exp_name=train_cord, exp_version=test_experiment]
Some important arguments:
--config
: config file path for model training.--pretrained_model_name_or_path
: string format, model name in Hugging Face modelhub or local path.--dataset_name_or_paths
: string format (json dumped), list of dataset names in Hugging Face datasets or local paths.--result_path
: file path to save model outputs/artifacts.--exp_version
: used for experiment versioning. The output files are saved at{result_path}/{exp_version}/*
Test
With the trained model, test images and ground truth parses, you can get inference results and accuracy scores.
python test.py --dataset_name_or_path naver-clova-ix/cord-v2 --pretrained_model_name_or_path ./result/train_cord/test_experiment --save_path ./result/output.json
100%|βββββββββββββ| 100/100 [00:35<00:00, 2.80it/s]
Total number of samples: 100, Tree Edit Distance (TED) based accuracy score: 0.9129639764131697, F1 accuracy score: 0.8406020841373987
Some important arguments:
--dataset_name_or_path
: string format, the target dataset name in Hugging Face datasets or local path.--pretrained_model_name_or_path
: string format, the model name in Hugging Face modelhub or local path.--save_path
: file path to save predictions and scores.
How to Cite
If you find this work useful to you, please cite:
@inproceedings{kim2022donut,
title = {OCR-Free Document Understanding Transformer},
author = {Kim, Geewook and Hong, Teakgyu and Yim, Moonbin and Nam, JeongYeon and Park, Jinyoung and Yim, Jinyeong and Hwang, Wonseok and Yun, Sangdoo and Han, Dongyoon and Park, Seunghyun},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2022}
}
License
MIT license
Copyright (c) 2022-present NAVER Corp.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
THE SOFTWARE.