# layoutlm_CORD
# Introduction
This repo is a implementation of the Layoutlm Model, see [1], from the sourcecode (as I didn't manage to make it work with the
huggingface implementation : [HuggingFace Implementation](https://huggingface.co/transformers/model_doc/layoutlm.html) and benchmarked on the CORD Dataset,
see [2].
# Results
I compare the performance of the pre-train LayoutLM on IIT-CDIP dataset (version LARGE) with the Bert (version Large).
## Validation Set
Model | F1_Score | Precision | Recall
--- | --- | --- | --- |
LayoutLM Large| **0.9562** | **0.9577** | **0.9546** |
Bert Large | **0.9474** | **0.9466** | **0.9481** |
## Test Set
Model | F1_Score | Precision | Recall
--- | --- | --- | --- |
LayoutLM Large| **0.9843** | **0.9845** | **0.9841** |
Bert Large | **0.9859** | **0.9861** | **0.9856** |
In the validation set, Layoutlm outperformed Bert, but it is not the case in the test set. I need to do more
investigation. \
Nevertheless, it took Bert 11 minutes to finish the training (4 epochs) while Layoutlm needed only 3 minutes.
(same environment, setup ..)
# Important files
I am using the Layoutlm Large, files of the pre-trained model can be found on these links : \
[OneDrive](https://1drv.ms/u/s!ApPZx_TWwibInSy2nj7YabBsTWNa?e=p4LQo1) /
[GoogleDrive](https://drive.google.com/open?id=1tatUuWVuNUxsP02smZCbB5NspyGo7g2g) \
Other ressources can be found on the original repository : [Official Layoutlm](https://github.com/microsoft/unilm/tree/master/layoutlm)
##TODO
I will soon put a script for the training, otherwise you can always check my notebooks. \
I will also give more details about the dataset, the notebook's structure ...
## References
[1] Yiheng Xu and Minghao Li and Lei Cui and Shaohan Huang and Furu Wei and Ming Zhou (2019) ,
LayoutLM: Pre-training of Text and Layout for Document Image Understanding (https://arxiv.org/abs/1912.13318),
https://github.com/microsoft/unilm/tree/master/layoutlm
[2] Park, Seunghyun and Shin, Seung and Lee, Bado and Lee, Junyeop and Surh, Jaeheung and Seo, Minjoon and Lee, Hwalsuk (2019)
CORD: A Consolidated Receipt Dataset for Post-OCR Parsing (Document Intelligence Workshop at Neural Information Processing Systems)
https://github.com/clovaai/cord
没有合适的资源?快使用搜索试试~ 我知道了~
layoutlm_CORD:在 CORD 数据集上评估 Layoutlm 模型
共6个文件
ipynb:3个
license:1个
gitignore:1个
需积分: 44 6 下载量 92 浏览量
2021-05-29
15:57:36
上传
评论
收藏 29KB ZIP 举报
温馨提示
layoutlm_CORD 介绍 这个 repo 是 Layoutlm 模型的一个实现,参见 [1],来自源代码(因为我没有设法让它与 Huggingface 实现一起工作: 并在 CORD 数据集上进行了基准测试,参见 [2]。 结果 我将预训练 LayoutLM 在 IIT-CDIP 数据集(大版本)上的性能与 Bert(大版本)进行了比较。 验证集 模型 F1_Score 精确 记起 布局LM大 0.9562 0.9577 0.9546 伯特大 0.9474 0.9466 0.9481 测试集 模型 F1_Score 精确 记起 布局LM大 0.9843 0.9845 0.9841 伯特大 0.9859 0.9861 0.9856 在验证集中,Layoutlm 的表现优于 Bert,但在测试集中却并非如此。 我需要做更多调查。 尽管如此,Bert 花了 11
资源推荐
资源详情
资源评论
收起资源包目录
layoutlm_CORD-main.zip (6个子文件)
layoutlm_CORD-main
bertlarge.ipynb 41KB
LICENSE 1KB
layoutlm.ipynb 58KB
.gitignore 2KB
generate_dataset.ipynb 23KB
README.md 2KB
共 6 条
- 1
资源评论
佳同学
- 粉丝: 28
- 资源: 4583
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功