Multimodal (text + layout/format + image) pre-training for document understanding
LayoutLM is a simple but effective pre-training method of text and layout for document image understanding and information extraction tasks, such as form understanding and receipt understanding. LayoutLM archives the SOTA results on multiple datasets. For more details, please refer to our paper:
LayoutLM: Pre-training of Text and Layout for Document Image Understanding Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou, KDD 2020
***** New May 16th, 2020: Our LayoutLM paper has been accepted to KDD 2020 as a full paper in the research track*****
***** New Feb 18th, 2020: Initial release of pre-trained models and fine-tuning code for LayoutLM v1 *****
We pre-train LayoutLM on IIT-CDIP Test Collection 1.0* dataset with two settings.
- LayoutLM-Base, Uncased (11M documents, 2 epochs): 12-layer, 768-hidden, 12-heads, 113M parameters || OneDrive | Google Drive
- LayoutLM-Large, Uncased (11M documents, 2 epochs): 24-layer, 1024-hidden, 16-heads, 343M parameters || OneDrive | Google Drive
*As some downstream datasets are the subsets of IIT-CDIP, we have carefully excluded the overlap portion from the pre-training data.
We evaluate LayoutLM on several document image understanding datasets, and it outperforms several SOTA pre-trained models and approaches.
Setup environment as follows:
conda create -n layoutlm python=3.6
conda activate layoutlm
conda install pytorch==1.4.0 cudatoolkit=10.1 -c pytorch
git clone https://github.com/NVIDIA/apex && cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
pip install .
## For development mode
# pip install -e ".[dev]"We give a fine-tuning example for sequence labeling tasks. You can run this example on FUNSD, a dataset for document understanding tasks.
First, we need to preprocess the JSON file into txt. You can run the preprocessing scripts funsd_preprocess.py in the scripts directory. For more options, please refer to the arguments.
cd examples/seq_labeling
./preprocess.shAfter preprocessing, run LayoutLM as follows:
python run_seq_labeling.py --data_dir data \
--model_type layoutlm \
--model_name_or_path path/to/pretrained/model/directory \
--do_lower_case \
--max_seq_length 512 \
--do_train \
--num_train_epochs 100.0 \
--logging_steps 10 \
--save_steps -1 \
--output_dir path/to/output/directory \
--labels data/labels.txt \
--per_gpu_train_batch_size 16 \
--per_gpu_eval_batch_size 16 \
--fp16Note: The DataParallel will be enabled automatically to utilize all GPUs. If you want to train with DistributedDataParallel, please run the script like:
# Suppose you have 4 GPUs.
python -m torch.distributed.launch --nproc_per_node=4 run_seq_labeling.py --data_dir data \
--model_type layoutlm \
--model_name_or_path path/to/pretrained/model/directory \
--do_lower_case \
--max_seq_length 512 \
--do_train \
--num_train_epochs 100.0 \
--logging_steps 10 \
--save_steps -1 \
--output_dir path/to/output/directory \
--labels data/labels.txt \
--per_gpu_train_batch_size 16 \
--per_gpu_eval_batch_size 16 \
--fp16Then you can do evaluation or inference by replacing --do_train with --do_eval or --do_predict
Also, you can run Bert and RoBERTa baseline by modifying the --model_type argument. For more options, please refer to the arguments of run.py.
We also fine-tune LayoutLM on the document image classification task. You can download the RVL-CDIP dataset from here. Because this dataset only provides the document image, you should use the OCR tool to get the texts and bounding boxes. For example, you can easily use Tesseract, an open-source OCR engine, to generate corresponding OCR data in hOCR format. For more details, please refer to the Tesseract wiki. Your processed data should look like this sample data.
With the processed OCR data, you can run LayoutLM as follows:
python run_classification.py --data_dir data \
--model_type layoutlm \
--model_name_or_path path/to/pretrained/model/directory \
--output_dir path/to/output/directory \
--do_lower_case \
--max_seq_length 512 \
--do_train \
--do_eval \
--num_train_epochs 40.0 \
--logging_steps 5000 \
--save_steps 5000 \
--per_gpu_train_batch_size 16 \
--per_gpu_eval_batch_size 16 \
--evaluate_during_training \
--fp16 Similarly, you can do evaluation by changing --do_train to --do_eval and --do_test
Like the sequence labeling task, you can run Bert and RoBERTa baseline by modifying the --model_type argument.
| Model | Hmean |
|---|---|
| BERT-Large | 90.99% |
| RoBERTa-Large | 92.80% |
| Ranking 1st in SROIE | 94.02% |
| LayoutLM | 96.04% |
| Model | Accuracy |
|---|---|
| BERT-Large | 89.92% |
| RoBERTa-Large | 90.11% |
| VGG-16 (Afzal et al., 2017) | 90.97% |
| Stacked CNN Ensemble (Das et al., 2018) | 92.21% |
| LadderNet (Sarkhel & Nandi, 2019) | 92.77% |
| Multimodal Ensemble (Dauphinee et al., 2019) | 93.07% |
| LayoutLM | 94.42% |
| Model | Precision | Recall | F1 |
|---|---|---|---|
| BERT-Large | 0.6113 | 0.7085 | 0.6563 |
| RoBERTa-Large | 0.6780 | 0.7391 | 0.7072 |
| LayoutLM | 0.7677 | 0.8195 | 0.7927 |
If you find LayoutLM useful in your research, please cite the following paper:
@misc{xu2019layoutlm,
title={LayoutLM: Pre-training of Text and Layout for Document Image Understanding},
author={Yiheng Xu and Minghao Li and Lei Cui and Shaohan Huang and Furu Wei and Ming Zhou},
year={2019},
eprint={1912.13318},
archivePrefix={arXiv},
primaryClass={cs.CL}
}This project is licensed under the license found in the LICENSE file in the root directory of this source tree. Portions of the source code are based on the transformers project. Microsoft Open Source Code of Conduct
For help or issues using LayoutLM, please submit a GitHub issue.
For other communications related to LayoutLM, please contact Lei Cui ([email protected]), Furu Wei ([email protected]).