提交 a8ba8963 编写于 作者: xuyang2233's avatar xuyang2233

merge cpnflict 0815

Global:
use_gpu: true
epoch_num: 500
log_smooth_window: 20
print_batch_step: 10
save_model_dir: ./output/sr/sr_tsrn_transformer_strock/
save_epoch_step: 3
# evaluation is run every 2000 iterations
eval_batch_step: [0, 1000]
cal_metric_during_train: False
pretrained_model:
checkpoints:
save_inference_dir: sr_output
use_visualdl: False
infer_img: doc/imgs_words_en/word_52.png
# for data or label process
character_dict_path: ./train_data/srdata/english_decomposition.txt
max_text_length: 100
infer_mode: False
use_space_char: False
save_res_path: ./output/sr/predicts_gestalt.txt
Optimizer:
name: Adam
beta1: 0.5
beta2: 0.999
clip_norm: 0.25
lr:
learning_rate: 0.0001
Architecture:
model_type: sr
algorithm: Gestalt
Transform:
name: TSRN
STN: True
infer_mode: False
Loss:
name: StrokeFocusLoss
character_dict_path: ./train_data/srdata/english_decomposition.txt
PostProcess:
name: None
Metric:
name: SRMetric
main_indicator: all
Train:
dataset:
name: LMDBDataSetSR
data_dir: ./train_data/srdata/train
transforms:
- SRResize:
imgH: 32
imgW: 128
down_sample_scale: 2
- SRLabelEncode: # Class handling label
- KeepKeys:
keep_keys: ['img_lr', 'img_hr', 'length', 'input_tensor', 'label'] # dataloader will return list in this order
loader:
shuffle: False
batch_size_per_card: 16
drop_last: True
num_workers: 4
Eval:
dataset:
name: LMDBDataSetSR
data_dir: ./train_data/srdata/test
transforms:
- SRResize:
imgH: 32
imgW: 128
down_sample_scale: 2
- SRLabelEncode: # Class handling label
- KeepKeys:
keep_keys: ['img_lr', 'img_hr','length', 'input_tensor', 'label'] # dataloader will return list in this order
loader:
shuffle: False
drop_last: False
batch_size_per_card: 16
num_workers: 4
# 关键信息抽取算法-LayoutXLM
- [1. 算法简介](#1-算法简介)
- [2. 环境配置](#2-环境配置)
- [3. 模型训练、评估、预测](#3-模型训练评估预测)
- [4. 推理部署](#4-推理部署)
- [4.1 Python推理](#41-python推理)
- [4.2 C++推理部署](#42-推理部署)
- [4.3 Serving服务化部署](#43-serving服务化部署)
- [4.4 更多推理部署](#44-更多推理部署)
- [5. FAQ](#5-faq)
- [引用](#引用)
<a name="1"></a>
## 1. 算法简介
论文信息:
> [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836)
>
> Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei
>
> 2021
在XFUND_zh数据集上,算法复现效果如下:
|模型|骨干网络|任务|配置文件|hmean|下载链接|
| --- | --- |--|--- | --- | --- |
|LayoutXLM|LayoutXLM-base|SER |[ser_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml)|90.38%|[训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)/[推理模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh_infer.tar)|
|LayoutXLM|LayoutXLM-base|RE | [re_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml)|74.83%|[训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar)/[推理模型(coming soon)]()|
<a name="2"></a>
## 2. 环境配置
请先参考[《运行环境准备》](./environment.md)配置PaddleOCR运行环境,参考[《项目克隆》](./clone.md)克隆项目代码。
<a name="3"></a>
## 3. 模型训练、评估、预测
请参考[关键信息抽取教程](./kie.md)。PaddleOCR对代码进行了模块化,训练不同的关键信息抽取模型只需要**更换配置文件**即可。
<a name="4"></a>
## 4. 推理部署
<a name="4-1"></a>
### 4.1 Python推理
**注:** 目前RE任务推理过程仍在适配中,下面以SER任务为例,介绍基于LayoutXLM模型的关键信息抽取过程。
首先将训练得到的模型转换成inference model。LayoutXLM模型在XFUND_zh数据集上训练的模型为例([模型下载地址](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)),可以使用下面的命令进行转换。
``` bash
wget https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar
tar -xf ser_LayoutXLM_xfun_zh.tar
python3 tools/export_model.py -c configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./ser_LayoutXLM_xfun_zh/best_accuracy Global.save_inference_dir=./inference/ser_layoutxlm
```
LayoutXLM模型基于SER任务进行推理,可以执行如下命令:
```bash
cd ppstructure
python3 vqa/predict_vqa_token_ser.py \
--vqa_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_layoutxlm_infer \
--image_dir=./docs/vqa/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf
```
SER可视化结果默认保存到`./output`文件夹里面,结果示例如下:
<div align="center">
<img src="../../ppstructure/docs/vqa/result_ser/zh_val_42_ser.jpg" width="800">
</div>
<a name="4-2"></a>
### 4.2 C++推理部署
暂不支持
<a name="4-3"></a>
### 4.3 Serving服务化部署
暂不支持
<a name="4-4"></a>
### 4.4 更多推理部署
暂不支持
<a name="5"></a>
## 5. FAQ
## 引用
```bibtex
@article{DBLP:journals/corr/abs-2104-08836,
author = {Yiheng Xu and
Tengchao Lv and
Lei Cui and
Guoxin Wang and
Yijuan Lu and
Dinei Flor{\^{e}}ncio and
Cha Zhang and
Furu Wei},
title = {LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich
Document Understanding},
journal = {CoRR},
volume = {abs/2104.08836},
year = {2021},
url = {https://arxiv.org/abs/2104.08836},
eprinttype = {arXiv},
eprint = {2104.08836},
timestamp = {Thu, 14 Oct 2021 09:17:23 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2104-08836.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@article{DBLP:journals/corr/abs-1912-13318,
author = {Yiheng Xu and
Minghao Li and
Lei Cui and
Shaohan Huang and
Furu Wei and
Ming Zhou},
title = {LayoutLM: Pre-training of Text and Layout for Document Image Understanding},
journal = {CoRR},
volume = {abs/1912.13318},
year = {2019},
url = {http://arxiv.org/abs/1912.13318},
eprinttype = {arXiv},
eprint = {1912.13318},
timestamp = {Mon, 01 Jun 2020 16:20:46 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-1912-13318.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@article{DBLP:journals/corr/abs-2012-14740,
author = {Yang Xu and
Yiheng Xu and
Tengchao Lv and
Lei Cui and
Furu Wei and
Guoxin Wang and
Yijuan Lu and
Dinei A. F. Flor{\^{e}}ncio and
Cha Zhang and
Wanxiang Che and
Min Zhang and
Lidong Zhou},
title = {LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding},
journal = {CoRR},
volume = {abs/2012.14740},
year = {2020},
url = {https://arxiv.org/abs/2012.14740},
eprinttype = {arXiv},
eprint = {2012.14740},
timestamp = {Tue, 27 Jul 2021 09:53:52 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2012-14740.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
```
# 关键信息抽取算法-SDMGR
- [1. 算法简介](#1-算法简介)
- [2. 环境配置](#2-环境配置)
- [3. 模型训练、评估、预测](#3-模型训练评估预测)
- [3.1 模型训练](#31-模型训练)
- [3.2 模型评估](#32-模型评估)
- [3.3 模型预测](#33-模型预测)
- [4. 推理部署](#4-推理部署)
- [4.1 Python推理](#41-python推理)
- [4.2 C++推理部署](#42-c推理部署)
- [4.3 Serving服务化部署](#43-serving服务化部署)
- [4.4 更多推理部署](#44-更多推理部署)
- [5. FAQ](#5-faq)
- [引用](#引用)
<a name="1"></a>
## 1. 算法简介
论文信息:
> [Spatial Dual-Modality Graph Reasoning for Key Information Extraction](https://arxiv.org/abs/2103.14470)
>
> Hongbin Sun and Zhanghui Kuang and Xiaoyu Yue and Chenhao Lin and Wayne Zhang
>
> 2021
在wildreceipt发票公开数据集上,算法复现效果如下:
|模型|骨干网络|配置文件|hmean|下载链接|
| --- | --- | --- | --- | --- |
|SDMGR|VGG6|[configs/kie/sdmgr/kie_unet_sdmgr.yml](../../configs/kie/sdmgr/kie_unet_sdmgr.yml)|86.7%|[训练模型]( https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)/[推理模型(coming soon)]()|
<a name="2"></a>
## 2. 环境配置
请先参考[《运行环境准备》](./environment.md)配置PaddleOCR运行环境,参考[《项目克隆》](./clone.md)克隆项目代码。
<a name="3"></a>
## 3. 模型训练、评估、预测
SDMGR是一个关键信息提取算法,将每个检测到的文本区域分类为预定义的类别,如订单ID、发票号码,金额等。
训练和测试的数据采用wildreceipt数据集,通过如下指令下载数据集:
```bash
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/wildreceipt.tar && tar xf wildreceipt.tar
```
创建数据集软链到PaddleOCR/train_data目录下:
```
cd PaddleOCR/ && mkdir train_data && cd train_data
ln -s ../../wildreceipt ./
```
### 3.1 模型训练
训练采用的配置文件是`configs/kie/sdmgr/kie_unet_sdmgr.yml`,配置文件中默认训练数据路径是`train_data/wildreceipt`,准备好数据后,可以通过如下指令执行训练:
```
python3 tools/train.py -c configs/kie/sdmgr/kie_unet_sdmgr.yml -o Global.save_model_dir=./output/kie/
```
### 3.2 模型评估
执行下面的命令进行模型评估
```bash
python3 tools/eval.py -c configs/kie/sdmgr/kie_unet_sdmgr.yml -o Global.checkpoints=./output/kie/best_accuracy
```
输出信息示例如下所示。
```py
[2022/08/10 05:22:23] ppocr INFO: metric eval ***************
[2022/08/10 05:22:23] ppocr INFO: hmean:0.8670120239257812
[2022/08/10 05:22:23] ppocr INFO: fps:10.18816520530961
```
### 3.3 模型预测
执行下面的命令进行模型预测,预测的时候需要预先加载存储图片路径以及OCR信息的文本文件,使用`Global.infer_img`进行指定。
```bash
python3 tools/infer_kie.py -c configs/kie/kie_unet_sdmgr.yml -o Global.checkpoints=kie_vgg16/best_accuracy Global.infer_img=./train_data/wildreceipt/1.txt
```
执行预测后的结果保存在`./output/sdmgr_kie/predicts_kie.txt`文件中,可视化结果保存在`/output/sdmgr_kie/kie_results/`目录下。
可视化结果如下图所示:
<div align="center">
<img src="../../ppstructure/docs/imgs/sdmgr_result.png" width="800">
</div>
<a name="4"></a>
## 4. 推理部署
<a name="4-1"></a>
### 4.1 Python推理
暂不支持
<a name="4-2"></a>
### 4.2 C++推理部署
暂不支持
<a name="4-3"></a>
### 4.3 Serving服务化部署
暂不支持
<a name="4-4"></a>
### 4.4 更多推理部署
暂不支持
<a name="5"></a>
## 5. FAQ
## 引用
```bibtex
@misc{sun2021spatial,
title={Spatial Dual-Modality Graph Reasoning for Key Information Extraction},
author={Hongbin Sun and Zhanghui Kuang and Xiaoyu Yue and Chenhao Lin and Wayne Zhang},
year={2021},
eprint={2103.14470},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
# 关键信息抽取算法-VI-LayoutXLM
- [1. 算法简介](#1-算法简介)
- [2. 环境配置](#2-环境配置)
- [3. 模型训练、评估、预测](#3-模型训练评估预测)
- [4. 推理部署](#4-推理部署)
- [4.1 Python推理](#41-python推理)
- [4.2 C++推理部署](#42-c推理部署)
- [4.3 Serving服务化部署](#43-serving服务化部署)
- [4.4 更多推理部署](#44-更多推理部署)
- [5. FAQ](#5-faq)
- [引用](#引用)
<a name="1"></a>
## 1. 算法简介
VI-LayoutXLM基于LayoutXLM进行改进,在下游任务训练过程中,去除视觉骨干网络模块,最终精度基本无损的情况下,模型推理速度进一步提升。
在XFUND_zh数据集上,算法复现效果如下:
|模型|骨干网络|任务|配置文件|hmean|下载链接|
| --- | --- |---| --- | --- | --- |
|VI-LayoutXLM |VI-LayoutXLM-base | SER |[ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)|93.19%|[训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)/[推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar)|
|VI-LayoutXLM |VI-LayoutXLM-base |RE | [re_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)|83.92%|[训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar)/[推理模型(coming soon)]()|
<a name="2"></a>
## 2. 环境配置
请先参考[《运行环境准备》](./environment.md)配置PaddleOCR运行环境,参考[《项目克隆》](./clone.md)克隆项目代码。
<a name="3"></a>
## 3. 模型训练、评估、预测
请参考[关键信息抽取教程](./kie.md)。PaddleOCR对代码进行了模块化,训练不同的关键信息抽取模型只需要**更换配置文件**即可。
<a name="4"></a>
## 4. 推理部署
<a name="4-1"></a>
### 4.1 Python推理
**注:** 目前RE任务推理过程仍在适配中,下面以SER任务为例,介绍基于VI-LayoutXLM模型的关键信息抽取过程。
首先将训练得到的模型转换成inference model。以VI-LayoutXLM模型在XFUND_zh数据集上训练的模型为例([模型下载地址](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)),可以使用下面的命令进行转换。
``` bash
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar
tar -xf ser_vi_layoutxlm_xfund_pretrained.tar
python3 tools/export_model.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./ser_vi_layoutxlm_xfund_pretrained/best_accuracy Global.save_inference_dir=./inference/ser_vi_layoutxlm_infer
```
VI-LayoutXLM模型基于SER任务进行推理,可以执行如下命令:
```bash
cd ppstructure
python3 vqa/predict_vqa_token_ser.py \
--vqa_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_vi_layoutxlm_infer \
--image_dir=./docs/vqa/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx"
```
SER可视化结果默认保存到`./output`文件夹里面,结果示例如下:
<div align="center">
<img src="../../ppstructure/docs/vqa/result_ser/zh_val_42_ser.jpg" width="800">
</div>
<a name="4-2"></a>
### 4.2 C++推理部署
暂不支持
<a name="4-3"></a>
### 4.3 Serving服务化部署
暂不支持
<a name="4-4"></a>
### 4.4 更多推理部署
暂不支持
<a name="5"></a>
## 5. FAQ
## 引用
```bibtex
@article{DBLP:journals/corr/abs-2104-08836,
author = {Yiheng Xu and
Tengchao Lv and
Lei Cui and
Guoxin Wang and
Yijuan Lu and
Dinei Flor{\^{e}}ncio and
Cha Zhang and
Furu Wei},
title = {LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich
Document Understanding},
journal = {CoRR},
volume = {abs/2104.08836},
year = {2021},
url = {https://arxiv.org/abs/2104.08836},
eprinttype = {arXiv},
eprint = {2104.08836},
timestamp = {Thu, 14 Oct 2021 09:17:23 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2104-08836.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@article{DBLP:journals/corr/abs-1912-13318,
author = {Yiheng Xu and
Minghao Li and
Lei Cui and
Shaohan Huang and
Furu Wei and
Ming Zhou},
title = {LayoutLM: Pre-training of Text and Layout for Document Image Understanding},
journal = {CoRR},
volume = {abs/1912.13318},
year = {2019},
url = {http://arxiv.org/abs/1912.13318},
eprinttype = {arXiv},
eprint = {1912.13318},
timestamp = {Mon, 01 Jun 2020 16:20:46 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-1912-13318.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@article{DBLP:journals/corr/abs-2012-14740,
author = {Yang Xu and
Yiheng Xu and
Tengchao Lv and
Lei Cui and
Furu Wei and
Guoxin Wang and
Yijuan Lu and
Dinei A. F. Flor{\^{e}}ncio and
Cha Zhang and
Wanxiang Che and
Min Zhang and
Lidong Zhou},
title = {LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding},
journal = {CoRR},
volume = {abs/2012.14740},
year = {2020},
url = {https://arxiv.org/abs/2012.14740},
eprinttype = {arXiv},
eprint = {2012.14740},
timestamp = {Tue, 27 Jul 2021 09:53:52 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2012-14740.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
```
# OCR算法
# 算法汇总
- [1. 两阶段算法](#1)
- [1. 两阶段OCR算法](#1)
- [1.1 文本检测算法](#11)
- [1.2 文本识别算法](#12)
- [2. 端到端算法](#2)
- [2. 端到端OCR算法](#2)
- [3. 表格识别算法](#3)
- [4. 关键信息抽取算法](#4)
本文给出了PaddleOCR已支持的OCR算法列表,以及每个算法在**英文公开数据集**上的模型和指标,主要用于算法简介和算法性能对比,更多包括中文在内的其他数据集上的模型请参考[PP-OCR v2.0 系列模型下载](./models_list.md)
......@@ -116,3 +117,34 @@
|模型|骨干网络|配置文件|acc|下载链接|
|---|---|---|---|---|
|TableMaster|TableResNetExtra|[configs/table/table_master.yml](../../configs/table/table_master.yml)|77.47%|[训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/tablemaster/table_structure_tablemaster_train.tar) / [推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/tablemaster/table_structure_tablemaster_infer.tar)|
## 4. 关键信息抽取算法
已支持的关键信息抽取算法列表(戳链接获取使用教程):
- [x] [VI-LayoutXLM](./algorithm_kie_vi_laoutxlm.md)
- [x] [LayoutLM](./algorithm_kie_laoutxlm.md)
- [x] [LayoutLMv2](./algorithm_kie_laoutxlm.md)
- [x] [LayoutXLM](./algorithm_kie_laoutxlm.md)
- [x] [SDMGR](././algorithm_kie_sdmgr.md)
在wildreceipt发票公开数据集上,算法复现效果如下:
|模型|骨干网络|配置文件|hmean|下载链接|
| --- | --- | --- | --- | --- |
|SDMGR|VGG6|[configs/kie/sdmgr/kie_unet_sdmgr.yml](../../configs/kie/sdmgr/kie_unet_sdmgr.yml)|86.7%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)|
在XFUND_zh公开数据集上,算法效果如下:
|模型|骨干网络|任务|配置文件|hmean|下载链接|
| --- | --- | --- | --- | --- | --- |
|VI-LayoutXLM| VI-LayoutXLM-base | SER | [ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)|**93.19%**|[训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)|
|LayoutXLM| LayoutXLM-base | SER | [ser_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml)|90.38%|[训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)|
|LayoutLM| LayoutLM-base | SER | [ser_layoutlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutlm_xfund_zh.yml)|77.31%|[训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar)|
|LayoutLMv2| LayoutLMv2-base | SER | [ser_layoutlmv2_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutlmv2_xfund_zh.yml)|85.44%|[训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar)|
|VI-LayoutXLM| VI-LayoutXLM-base | RE | [re_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)|**83.92%**|[训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar)|
|LayoutXLM| LayoutXLM-base | RE | [re_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml)|74.83%|[训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar)|
|LayoutLMv2| LayoutLMv2-base | RE | [re_layoutlmv2_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutlmv2_xfund_zh.yml)|67.77%|[训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar)|
......@@ -101,7 +101,7 @@ python3 tools/export_model.py -c configs/rec/rec_r45_visionlan.yml -o Global.pre
执行如下命令进行模型推理:
```shell
python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words/en/word_2.png' --rec_model_dir='./inference/rec_r45_visionlan/' --rec_algorithm='VisionLAN' --rec_image_shape='3,64,256' --rec_char_dict_path='./ppocr/utils/dict36.txt'
python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words/en/word_2.png' --rec_model_dir='./inference/rec_r45_visionlan/' --rec_algorithm='VisionLAN' --rec_image_shape='3,64,256' --rec_char_dict_path='./ppocr/utils/ic15_dict.txt' --use_space_char=False
# 预测文件夹下所有图像时,可修改image_dir为文件夹,如 --image_dir='./doc/imgs_words_en/'。
```
......@@ -110,7 +110,7 @@ python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words/en/word_2.png'
执行命令后,上面图像的预测结果(识别的文本和得分)会打印到屏幕上,示例如下:
结果如下:
```shell
Predicts of ./doc/imgs_words/en/word_2.png:('yourself', 0.97076982)
Predicts of ./doc/imgs_words/en/word_2.png:('yourself', 0.9999493)
```
**注意**
......
# Text Gestalt
- [1. 算法简介](#1)
- [2. 环境配置](#2)
- [3. 模型训练、评估、预测](#3)
- [3.1 训练](#3-1)
- [3.2 评估](#3-2)
- [3.3 预测](#3-3)
- [4. 推理部署](#4)
- [4.1 Python推理](#4-1)
- [4.2 C++推理](#4-2)
- [4.3 Serving服务化部署](#4-3)
- [4.4 更多推理部署](#4-4)
- [5. FAQ](#5)
<a name="1"></a>
## 1. 算法简介
论文信息:
> [Text Gestalt: Stroke-Aware Scene Text Image Super-Resolution](https://arxiv.org/pdf/2112.08171.pdf)
> Chen, Jingye and Yu, Haiyang and Ma, Jianqi and Li, Bin and Xue, Xiangyang
> AAAI, 2022
参考[FudanOCR](https://github.com/FudanVI/FudanOCR/tree/main/text-gestalt) 数据下载说明,在TextZoom测试集合上超分算法效果如下:
|模型|骨干网络|PSNR_Avg|SSIM_Avg|配置文件|下载链接|
|---|---|---|---|---|---|
|Text Gestalt|tsrn|19.28|0.6560| [configs/sr/sr_tsrn_transformer_strock.yml](../../configs/sr/sr_tsrn_transformer_strock.yml)|[训练模型](https://paddleocr.bj.bcebos.com/sr_tsrn_transformer_strock_train.tar)|
<a name="2"></a>
## 2. 环境配置
请先参考[《运行环境准备》](./environment.md)配置PaddleOCR运行环境,参考[《项目克隆》](./clone.md)克隆项目代码。
<a name="3"></a>
## 3. 模型训练、评估、预测
请参考[文本识别训练教程](./recognition.md)。PaddleOCR对代码进行了模块化,训练不同的识别模型只需要**更换配置文件**即可。
- 训练
在完成数据准备后,便可以启动训练,训练命令如下:
```
#单卡训练(训练周期长,不建议)
python3 tools/train.py -c configs/sr/sr_tsrn_transformer_strock.yml
#多卡训练,通过--gpus参数指定卡号
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/sr/sr_tsrn_transformer_strock.yml
```
- 评估
```
# GPU 评估, Global.pretrained_model 为待测权重
python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/sr/sr_tsrn_transformer_strock.yml -o Global.pretrained_model={path/to/weights}/best_accuracy
```
- 预测:
```
# 预测使用的配置文件必须与训练一致
python3 tools/infer_sr.py -c configs/sr/sr_tsrn_transformer_strock.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.infer_img=doc/imgs_words_en/word_52.png
```
![](../imgs_words_en/word_52.png)
执行命令后,上面图像的超分结果如下:
![](../imgs_results/sr_word_52.png)
<a name="4"></a>
## 4. 推理部署
<a name="4-1"></a>
### 4.1 Python推理
首先将文本超分训练过程中保存的模型,转换成inference model。以 Text-Gestalt 训练的[模型](https://paddleocr.bj.bcebos.com/sr_tsrn_transformer_strock_train.tar) 为例,可以使用如下命令进行转换:
```shell
python3 tools/export_model.py -c configs/sr/sr_tsrn_transformer_strock.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.save_inference_dir=./inference/sr_out
```
Text-Gestalt 文本超分模型推理,可以执行如下命令:
```
python3 tools/infer/predict_sr.py --sr_model_dir=./inference/sr_out --image_dir=doc/imgs_words_en/word_52.png --sr_image_shape=3,32,128
```
执行命令后,图像的超分结果如下:
![](../imgs_results/sr_word_52.png)
<a name="4-2"></a>
### 4.2 C++推理
暂未支持
<a name="4-3"></a>
### 4.3 Serving服务化部署
暂未支持
<a name="4-4"></a>
### 4.4 更多推理部署
暂未支持
<a name="5"></a>
## 5. FAQ
## 引用
```bibtex
@inproceedings{chen2022text,
title={Text gestalt: Stroke-aware scene text image super-resolution},
author={Chen, Jingye and Yu, Haiyang and Ma, Jianqi and Li, Bin and Xue, Xiangyang},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={36},
number={1},
pages={285--293},
year={2022}
}
```
## DocVQA数据集
# 信息抽取数据集
这里整理了常见的DocVQA数据集,持续更新中,欢迎各位小伙伴贡献数据集~
- [FUNSD数据集](#funsd)
- [XFUND数据集](#xfund)
- [wildreceipt数据集](#wildreceipt)
<a name="funsd"></a>
#### 1、FUNSD数据集
## 1. FUNSD数据集
- **数据来源**:https://guillaumejaume.github.io/FUNSD/
- **数据简介**:FUNSD数据集是一个用于表单理解的数据集,它包含199张真实的、完全标注的扫描版图片,类型包括市场报告、广告以及学术报告等,并分为149张训练集以及50张测试集。FUNSD数据集适用于多种类型的DocVQA任务,如字段级实体分类、字段级实体连接等。部分图像以及标注框可视化如下所示:
<div align="center">
......@@ -16,12 +20,33 @@
- **下载地址**:https://guillaumejaume.github.io/FUNSD/download/
<a name="xfund"></a>
#### 2、XFUND数据集
## 2. XFUND数据集
- **数据来源**:https://github.com/doc-analysis/XFUND
- **数据简介**:XFUND是一个多语种表单理解数据集,它包含7种不同语种的表单数据,并且全部用人工进行了键-值对形式的标注。其中每个语种的数据都包含了199张表单数据,并分为149张训练集以及50张测试集。部分图像以及标注框可视化如下所示:
<div align="center">
<img src="../../datasets/xfund_demo/gt_zh_train_0.jpg" width="500">
<img src="../../datasets/xfund_demo/gt_zh_train_1.jpg" width="500">
</div>
- **下载地址**:https://github.com/doc-analysis/XFUND/releases/tag/v1.0
<a name="wildreceipt"></a>
## 3. wildreceipt数据集
- **数据来源**:https://arxiv.org/abs/2103.14470
- **数据简介**:wildreceipt数据集是英文发票数据集,包含26个类别(此处类别体系包含`Ignore`类别),共标注了50000个文本框。其中训练集包含1267张图片,测试集包含472张图片。部分图像以及标注框可视化如下所示:
<div align="center">
<img src="../../datasets/wildreceipt_demo/2769.jpeg" width="500">
<img src="../../datasets/wildreceipt_demo/1bbe854b8817dedb8585e0732089fd1f752d2cec.jpeg" width="500">
</div>
**注:** 这里对于类别为`Ignore`或者`Others`的文本,没有进行可视化。
- **下载地址**
- 原始数据下载地址:[链接](https://download.openmmlab.com/mmocr/data/wildreceipt.tar)
- 数据格式转换后适配于PaddleOCR训练的数据下载地址:[链接](https://paddleocr.bj.bcebos.com/ppstructure/dataset/wildreceipt.tar)
# 关键信息抽取
本文提供了PaddleOCR关键信息抽取的全流程指南,包括语义实体识别 (Semantic Entity Recognition) 以及关系抽取 (Relation Extraction, RE) 任务的数据准备、模型训练、调优、评估、预测,各个阶段的详细说明。
- [1. 数据准备](#1-数据准备)
- [1.1. 准备数据集](#11-准备数据集)
- [1.2. 自定义数据集](#12-自定义数据集)
- [1.3. 数据下载](#13-数据下载)
- [2. 开始训练](#2-开始训练)
- [2.1. 启动训练](#21-启动训练)
- [2.2. 断点训练](#22-断点训练)
- [2.3. 混合精度训练](#24-混合精度训练)
- [2.4. 分布式训练](#25-分布式训练)
- [2.5. 知识蒸馏训练](#26-知识蒸馏训练)
- [2.6. 其他训练环境](#27-其他训练环境)
- [3. 模型评估与预测](#3-模型评估与预测)
- [3.1. 指标评估](#31-指标评估)
- [3.2. 测试信息抽取效果](#32-测试识别效果)
- [4. 模型导出与预测](#4-模型导出与预测)
- [5. FAQ](#5-faq)
# 1. 数据准备
## 1.1. 准备数据集
在训练信息抽取相关模型时,PaddleOCR支持以下数据格式。
- `通用数据` 用于训练以文本文件存储的数据集(SimpleDataSet);
训练数据的默认存储路径是 `PaddleOCR/train_data`,如果您的磁盘上已有数据集,只需创建软链接至数据集目录:
```
# linux and mac os
ln -sf <path/to/dataset> <path/to/paddle_ocr>/train_data/dataset
# windows
mklink /d <path/to/paddle_ocr>/train_data/dataset <path/to/dataset>
```
## 1.2. 自定义数据集
训练过程中一般包含训练集与验证集,二者数据格式相同,下面介绍如何自定义数据集。
**(1)训练集**
建议将训练图片放入同一个文件夹,并用一个文本文件记录图片路径和标签,文本文件里的内容如下:
```py
" 图像文件名 图像标注信息 "
zh_train_0.jpg [{"transcription": "汇丰晋信", "label": "other", "points": [[104, 114], [530, 114], [530, 175], [104, 175]], "id": 1, "linking": []}, {"transcription": "受理时间:", "label": "question", "points": [[126, 267], [266, 267], [266, 305], [126, 305]], "id": 7, "linking": [[7, 13]]}, {"transcription": "2020.6.15", "label": "answer", "points": [[321, 239], [537, 239], [537, 285], [321, 285]], "id": 13, "linking": [[7, 13]]}]
zh_train_1.jpg [{"transcription": "中国人体器官捐献", "label": "other", "points": [[544, 459], [954, 459], [954, 517], [544, 517]], "id": 1, "linking": []}, {"transcription": ">编号:MC545715483585", "label": "other", "points": [[1462, 470], [2054, 470], [2054, 543], [1462, 543]], "id": 10, "linking": []}, {"transcription": "CHINAORGANDONATION", "label": "other", "points": [[543, 516], [958, 516], [958, 551], [543, 551]], "id": 14, "linking": []}, {"transcription": "中国人体器官捐献志愿登记表", "label": "header", "points": [[635, 793], [1892, 793], [1892, 904], [635, 904]], "id": 18, "linking": []}]
...
```
**注意:** 文本文件中默认请将图片路径和图片标签用 `\t` 分割,如用其他方式分割将造成训练报错。
其中图像标注信息字符串经过json解析之后可以得到一个列表信息,列表中每个元素是一个字典,存储了每个文本行的需要信息,各个字段的含义如下。
- transcription: 存储了文本行的文字内容
- label: 该文本行内容所属的类别
- points: 存储文本行的四点位置信息
- id: 存储文本行的id信息,用于RE任务的训练
- linking: 存储文本行的之间的连接信息,用于RE任务的训练
**(2)验证集**
验证集构建方式与训练集相同。
* 字典文件
训练集与验证集中的文本行包含标签信息,所有标签的列表存在字典文件中(如`class_list.txt`),字典文件中的每一行表示为一个类别名称。
以XFUND_zh数据为例,共包含4个类别,字典文件内容如下所示。
```
OTHER
QUESTION
ANSWER
HEADER
```
在标注文件中,每个标注的文本行内容的`label`字段标注信息需要属于字典内容。
最终数据集应有如下文件结构:
```
|-train_data
|-data_name
|- train.json
|- train
|- zh_train_0.png
|- zh_train_1.jpg
| ...
|- val.json
|- val
|- zh_val_0.png
|- zh_val_1.jpg
| ...
```
**注:**
- 标注文件中的类别信息不区分大小写,如`HEADER``header`会被解析为相同的类别id,因此在标注的时候,不能使用小写处理后相同的字符串表示不同的类别。
- 在整理标注文件的时候,建议将other这个类别(其他,无需关注的文本行可以标注为other)放在第一行,在解析的时候,会将`other`类别的类别id解析为0,后续不会对该类进行可视化。
## 1.3. 数据下载
如果你没有本地数据集,可以从[XFUND](https://github.com/doc-analysis/XFUND)或者[FUNSD](https://guillaumejaume.github.io/FUNSD/)官网下载数据,然后使用XFUND与FUNSD的处理脚本([XFUND](../../ppstructure/vqa/tools/trans_xfun_data.py), [FUNSD](../../ppstructure/vqa/tools/trans_funsd_label.py)),生成用于PaddleOCR训练的数据格式,并使用公开数据集快速体验关键信息抽取的流程。
更多关于公开数据集的介绍,请参考[关键信息抽取数据集说明文档](./dataset/kie_datasets.md)
PaddleOCR也支持了关键信息抽取模型的标注,具体使用方法请参考:[PPOCRLabel使用文档](../../PPOCRLabel/README_ch.md)
# 2. 开始训练
PaddleOCR提供了训练脚本、评估脚本和预测脚本,本节将以 VI-LayoutXLM 多模态预训练模型为例进行讲解。
> 如果希望使用基于SDMGR的关键信息抽取算法,请参考:[SDMGR使用](./algorithm_kie_sdmgr.md)。
## 2.1. 启动训练
如果你没有使用自定义数据集,可以使用PaddleOCR中已经处理好的XFUND_zh数据集进行快速体验。
```bash
mkdir train_data
cd train_data
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar && tar -xf XFUND.tar
cd ..
```
如果不希望训练,直接体验后面的模型评估、预测、动转静、推理的流程,可以下载PaddleOCR中提供的预训练模型,并跳过2.1部分。
使用下面的方法,下载基于XFUND数据的SER与RE任务预训练模型。
```bash
mkdir pretrained_model
cd pretrained_model
# 下载并解压SER预训练模型
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar & tar -xf ser_vi_layoutxlm_xfund_pretrained.tar
# 下载并解压RE预训练模型
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar & tar -xf re_vi_layoutxlm_xfund_pretrained.tar
```
开始训练:
- 如果您安装的是cpu版本,请将配置文件中的 `use_gpu` 字段修改为false
- PaddleOCR在训练时,会默认下载VI-LayoutXLM预训练模型,这里无需预先下载。
```bash
# GPU训练 支持单卡,多卡训练
# 训练日志会自动保存到 配置文件中"{Global.save_model_dir}" 下的train.log文件中
# SER单卡训练
python3 tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml
# SER多卡训练,通过--gpus参数指定卡号
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml
# RE任务单卡训练
python3 tools/train.py -c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml
```
以SER任务为例,正常启动训练后,会看到以下log输出:
```
[2022/08/08 16:28:28] ppocr INFO: epoch: [1/200], global_step: 10, lr: 0.000006, loss: 1.871535, avg_reader_cost: 0.28200 s, avg_batch_cost: 0.82318 s, avg_samples: 8.0, ips: 9.71838 samples/s, eta: 0:51:59
[2022/08/08 16:28:33] ppocr INFO: epoch: [1/200], global_step: 19, lr: 0.000018, loss: 1.461939, avg_reader_cost: 0.00042 s, avg_batch_cost: 0.32037 s, avg_samples: 6.9, ips: 21.53773 samples/s, eta: 0:37:55
[2022/08/08 16:28:39] ppocr INFO: cur metric, precision: 0.11526348939743859, recall: 0.19776657060518732, hmean: 0.14564265817747712, fps: 34.008392345050055
[2022/08/08 16:28:45] ppocr INFO: save best model is to ./output/ser_vi_layoutxlm_xfund_zh/best_accuracy
[2022/08/08 16:28:45] ppocr INFO: best metric, hmean: 0.14564265817747712, precision: 0.11526348939743859, recall: 0.19776657060518732, fps: 34.008392345050055, best_epoch: 1
[2022/08/08 16:28:51] ppocr INFO: save model in ./output/ser_vi_layoutxlm_xfund_zh/latest
```
log 中自动打印如下信息:
| 字段 | 含义 |
| :----: | :------: |
| epoch | 当前迭代轮次 |
| iter | 当前迭代次数 |
| lr | 当前学习率 |
| loss | 当前损失函数 |
| reader_cost | 当前 batch 数据处理耗时 |
| batch_cost | 当前 batch 总耗时 |
| samples | 当前 batch 内的样本数 |
| ips | 每秒处理图片的数量 |
PaddleOCR支持训练和评估交替进行, 可以在 `configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml` 中修改 `eval_batch_step` 设置评估频率,默认每19个iter评估一次。评估过程中默认将最佳hmean模型,保存为 `output/ser_vi_layoutxlm_xfund_zh/best_accuracy/`
如果验证集很大,测试将会比较耗时,建议减少评估次数,或训练完再进行评估。
**提示:** 可通过 -c 参数选择 `configs/kie/` 路径下的多种模型配置进行训练,PaddleOCR支持的信息抽取算法可以参考[前沿算法列表](./algorithm_overview.md)
如果你希望训练自己的数据集,需要修改配置文件中的数据配置、字典文件以及类别数。
`configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml` 为例,修改的内容如下所示。
```yaml
Architecture:
# ...
Backbone:
name: LayoutXLMForSer
pretrained: True
mode: vi
# 假设字典中包含n个字段(包含other),由于采用BIO标注,则类别数为2n-1
num_classes: &num_classes 7
PostProcess:
name: VQASerTokenLayoutLMPostProcess
# 修改字典文件的路径为你自定义的数据集的字典路径
class_path: &class_path train_data/XFUND/class_list_xfun.txt
Train:
dataset:
name: SimpleDataSet
# 修改为你自己的训练数据目录
data_dir: train_data/XFUND/zh_train/image
# 修改为你自己的训练数据标签文件
label_file_list:
- train_data/XFUND/zh_train/train.json
...
loader:
# 训练时的单卡batch_size
batch_size_per_card: 8
...
Eval:
dataset:
name: SimpleDataSet
# 修改为你自己的验证数据目录
data_dir: train_data/XFUND/zh_val/image
# 修改为你自己的验证数据标签文件
label_file_list:
- train_data/XFUND/zh_val/val.json
...
loader:
# 验证时的单卡batch_size
batch_size_per_card: 8
```
**注意,预测/评估时的配置文件请务必与训练一致。**
## 2.2. 断点训练
如果训练程序中断,如果希望加载训练中断的模型从而恢复训练,可以通过指定` Architecture.Backbone.checkpoints`指定要加载的模型路径:
```bash
python3 tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy
```
**注意**
- `Architecture.Backbone.checkpoints`的优先级高于`Architecture.Backbone.pretrained`,需要加载之前训练好的训练模型进行模型微调、恢复训练、模型评估时,需要使用`Architecture.Backbone.checkpoints`指定模型参数路径;如果需要使用默认提供的通用预训练模型进行训练,则需要指定`Architecture.Backbone.pretrained``True`,同时指定`Architecture.Backbone.checkpoints`为空(`null`)。
- LayoutXLM系列模型均是调用了PaddleNLP中的预训练模型,模型加载与保存的逻辑与PaddleNLP基本一致,因此在这里不需要指定`Global.pretrained_model`或者`Global.checkpoints`参数;此外,LayoutXLM系列模型的蒸馏训练目前不支持断点训练。
## 2.3. 混合精度训练
coming soon!
## 2.4. 分布式训练
多机多卡训练时,通过 `--ips` 参数设置使用的机器IP地址,通过 `--gpus` 参数设置使用的GPU ID:
```bash
python3 -m paddle.distributed.launch --ips="xx.xx.xx.xx,xx.xx.xx.xx" --gpus '0,1,2,3' tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml
```
**注意:** (1)采用多机多卡训练时,需要替换上面命令中的ips值为您机器的地址,机器之间需要能够相互ping通;(2)训练时需要在多个机器上分别启动命令。查看机器ip地址的命令为`ifconfig`;(3)更多关于分布式训练的性能优势等信息,请参考:[分布式训练教程](./distributed_training.md)
## 2.5. 知识蒸馏训练
PaddleOCR支持了基于U-DML知识蒸馏的关键信息抽取模型训练过程,配置文件请参考:[ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml),更多关于知识蒸馏的说明文档请参考:[知识蒸馏说明文档](./knowledge_distillation.md)
**注意**: PaddleOCR中LayoutXLM系列关键信息抽取模型的保存与加载逻辑与PaddleNLP保持一致,因此在蒸馏的过程中仅保存了学生模型的参数,如果希望使用保存的模型进行评估,需要使用学生模型的配置(上面的蒸馏文件对应的学生模型为[ser_vi_layoutxlm_xfund_zh.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml)
## 2.6. 其他训练环境
- Windows GPU/CPU
在Windows平台上与Linux平台略有不同:
Windows平台只支持`单卡`的训练与预测,指定GPU进行训练`set CUDA_VISIBLE_DEVICES=0`
在Windows平台,DataLoader只支持单进程模式,因此需要设置 `num_workers` 为0;
- macOS
不支持GPU模式,需要在配置文件中设置`use_gpu`为False,其余训练评估预测命令与Linux GPU完全相同。
- Linux DCU
DCU设备上运行需要设置环境变量 `export HIP_VISIBLE_DEVICES=0,1,2,3`,其余训练评估预测命令与Linux GPU完全相同。
# 3. 模型评估与预测
## 3.1. 指标评估
训练中模型参数默认保存在`Global.save_model_dir`目录下。在评估指标时,需要设置`Architecture.Backbone.checkpoints`指向保存的参数文件。评估数据集可以通过 `configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml` 修改Eval中的 `label_file_path` 设置。
```bash
# GPU 评估, Global.checkpoints 为待测权重
python3 tools/eval.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy
```
会输出以下信息,打印出precision、recall、hmean等信息。
```py
[2022/08/09 07:59:28] ppocr INFO: metric eval ***************
[2022/08/09 07:59:28] ppocr INFO: precision:0.697476609016161
[2022/08/09 07:59:28] ppocr INFO: recall:0.8861671469740634
[2022/08/09 07:59:28] ppocr INFO: hmean:0.7805806758686339
[2022/08/09 07:59:28] ppocr INFO: fps:17.367364606899105
```
## 3.2. 测试信息抽取结果
使用 PaddleOCR 训练好的模型,可以通过以下脚本进行快速预测。
默认预测的图片存储在 `infer_img` 里,通过 `-o Architecture.Backbone.checkpoints` 加载训练好的参数文件:
根据配置文件中设置的 `save_model_dir``save_epoch_step` 字段,会有以下几种参数被保存下来:
```
output/ser_vi_layoutxlm_xfund_zh/
├── best_accuracy
├── metric.states
├── model_config.json
├── model_state.pdparams
├── best_accuracy.pdopt
├── config.yml
├── train.log
├── latest
├── metric.states
├── model_config.json
├── model_state.pdparams
├── latest.pdopt
```
其中 best_accuracy.* 是评估集上的最优模型;latest.* 是最新保存的一个模型。
预测使用的配置文件必须与训练一致,如您通过 `python3 tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml` 完成了模型的训练过程。
您可以使用如下命令进行中文模型预测。
```bash
python3 tools/infer_vqa_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.infer_img=./ppstructure/docs/vqa/input/zh_val_42.jpg
```
预测图片如下所示,图片会存储在`Global.save_res_path`路径中。
<div align="center">
<img src="../../ppstructure/docs/vqa/result_ser/zh_val_42_ser.jpg" width="800">
</div>
预测过程中,默认会加载PP-OCRv3的检测识别模型,用于OCR的信息抽取,如果希望加载预先获取的OCR结果,可以使用下面的方式进行预测,指定`Global.infer_img`为标注文件,其中包含图片路径以及OCR信息,同时指定`Global.infer_mode`为False,表示此时不使用OCR预测引擎。
```bash
python3 tools/infer_vqa_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.infer_img=./train_data/XFUND/zh_val/val.json Global.infer_mode=False
```
对于上述图片,如果使用标注的OCR结果进行信息抽取,预测结果如下。
<div align="center">
<img src="../../ppstructure/docs/vqa/result_ser_with_gt_ocr/zh_val_42_ser.jpg" width="800">
</div>
可以看出,部分检测框信息更加准确,但是整体信息抽取识别结果基本一致。
在RE任务模型预测时,需要先给出模型SER结果,因此需要同时加载SER的配置文件与模型权重,示例如下。
```bash
python3 ./tools/infer_vqa_token_ser_re.py -c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_udml_xfund_zh/re_layoutxlm_xfund_zh_v4_udml/best_accuracy/ Global.infer_img=./train_data/XFUND/zh_val/image/ -c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o_ser Architecture.Backbone.checkpoints=pretrain_models/ser_vi_layoutxlm_udml_xfund_zh/best_accuracy/
```
预测结果如下所示。
<div align="center">
<img src="../../ppstructure/docs/vqa/result_re/zh_val_42_re.jpg" width="800">
</div>
如果希望使用标注或者预先获取的OCR信息进行关键信息抽取,同上,可以指定`Global.infer_mode`为False,指定`Global.infer_img`为标注文件。
```bash
python3 ./tools/infer_vqa_token_ser_re.py -c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_udml_xfund_zh/re_layoutxlm_xfund_zh_v4_udml/best_accuracy/ Global.infer_img=./train_data/XFUND/zh_val/val.json Global.infer_mode=False -c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o_ser Architecture.Backbone.checkpoints=pretrain_models/ser_vi_layoutxlm_udml_xfund_zh/best_accuracy/
```
其中`c_ser`表示SER的配置文件,`o_ser` 后面需要加上待修改的SER模型与配置文件,如预训练权重等。
预测结果如下所示。
<div align="center">
<img src="../../ppstructure/docs/vqa/result_re_with_gt_ocr/zh_val_42_re.jpg" width="800">
</div>
可以看出,直接使用标注的OCR结果的RE预测结果要更加准确一些。
# 4. 模型导出与预测
## 4.1 模型导出
inference 模型(`paddle.jit.save`保存的模型)
一般是模型训练,把模型结构和模型参数保存在文件中的固化模型,多用于预测部署场景。
训练过程中保存的模型是checkpoints模型,保存的只有模型的参数,多用于恢复训练等。
与checkpoints模型相比,inference 模型会额外保存模型的结构信息,在预测部署、加速推理上性能优越,灵活方便,适合于实际系统集成。
信息抽取模型中的SER任务转inference模型步骤如下:
```bash
# -c 后面设置训练算法的yml配置文件
# -o 配置可选参数
# Global.pretrained_model 参数设置待转换的训练模型地址。
# Global.save_inference_dir参数设置转换的模型将保存的地址。
python3 tools/export_model.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.save_inference_dir=./inference/ser_vi_layoutxlm
```
转换成功后,在目录下有三个文件:
```
inference/ser_vi_layoutxlm/
├── inference.pdiparams # inference模型的参数文件
├── inference.pdiparams.info # inference模型的参数信息,可忽略
└── inference.pdmodel # inference模型的模型结构文件
```
RE任务的动转静过程适配中,敬请期待。
## 4.2 模型推理
VI-LayoutXLM模型基于SER任务进行推理,可以执行如下命令:
```bash
cd ppstructure
python3 vqa/predict_vqa_token_ser.py \
--vqa_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_vi_layoutxlm \
--image_dir=./docs/vqa/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx"
```
可视化SER结果结果默认保存到`./output`文件夹里面。结果示例如下:
<div align="center">
<img src="../../ppstructure/docs/vqa/result_ser/zh_val_42_ser.jpg" width="800">
</div>
# 5. FAQ
Q1: 训练模型转inference 模型之后预测效果不一致?
**A**:该问题多是trained model预测时候的预处理、后处理参数和inference model预测的时候的预处理、后处理参数不一致导致的。可以对比训练使用的配置文件中的预处理、后处理和预测时是否存在差异。
......@@ -90,7 +90,7 @@ After the conversion is successful, there are three files in the directory:
For VisionLAN text recognition model inference, the following commands can be executed:
```
python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words/en/word_2.png' --rec_model_dir='./inference/rec_r45_visionlan/' --rec_algorithm='VisionLAN' --rec_image_shape='3,64,256' --rec_char_dict_path='./ppocr/utils/dict36.txt'
python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words/en/word_2.png' --rec_model_dir='./inference/rec_r45_visionlan/' --rec_algorithm='VisionLAN' --rec_image_shape='3,64,256' --rec_char_dict_path='./ppocr/utils/ic15_dict.txt' --use_space_char=False
```
![](../imgs_words/en/word_2.png)
......@@ -98,7 +98,7 @@ python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words/en/word_2.png'
After executing the command, the prediction result (recognized text and score) of the image above is printed to the screen, an example is as follows:
The result is as follows:
```shell
Predicts of ./doc/imgs_words/en/word_2.png:('yourself', 0.97076982)
Predicts of ./doc/imgs_words/en/word_2.png:('yourself', 0.9999493)
```
<a name="4-2"></a>
......
# Text Gestalt
- [1. Introduction](#1)
- [2. Environment](#2)
- [3. Model Training / Evaluation / Prediction](#3)
- [3.1 Training](#3-1)
- [3.2 Evaluation](#3-2)
- [3.3 Prediction](#3-3)
- [4. Inference and Deployment](#4)
- [4.1 Python Inference](#4-1)
- [4.2 C++ Inference](#4-2)
- [4.3 Serving](#4-3)
- [4.4 More](#4-4)
- [5. FAQ](#5)
<a name="1"></a>
## 1. Introduction
Paper:
> [Text Gestalt: Stroke-Aware Scene Text Image Super-Resolution](https://arxiv.org/pdf/2112.08171.pdf)
> Chen, Jingye and Yu, Haiyang and Ma, Jianqi and Li, Bin and Xue, Xiangyang
> AAAI, 2022
Referring to the [FudanOCR](https://github.com/FudanVI/FudanOCR/tree/main/text-gestalt) data download instructions, the effect of the super-score algorithm on the TextZoom test set is as follows:
|Model|Backbone|config|Acc|Download link|
|---|---|---|---|---|---|
|Text Gestalt|tsrn|19.28|0.6560| [configs/sr/sr_tsrn_transformer_strock.yml](../../configs/sr/sr_tsrn_transformer_strock.yml)|[train model](https://paddleocr.bj.bcebos.com/sr_tsrn_transformer_strock_train.tar)|
<a name="2"></a>
## 2. Environment
Please refer to ["Environment Preparation"](./environment_en.md) to configure the PaddleOCR environment, and refer to ["Project Clone"](./clone_en.md) to clone the project code.
<a name="3"></a>
## 3. Model Training / Evaluation / Prediction
Please refer to [Text Recognition Tutorial](./recognition_en.md). PaddleOCR modularizes the code, and training different models only requires **changing the configuration file**.
Training:
Specifically, after the data preparation is completed, the training can be started. The training command is as follows:
```
#Single GPU training (long training period, not recommended)
python3 tools/train.py -c configs/sr/sr_tsrn_transformer_strock.yml
#Multi GPU training, specify the gpu number through the --gpus parameter
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/sr/sr_tsrn_transformer_strock.yml
```
Evaluation:
```
# GPU evaluation
python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/sr/sr_tsrn_transformer_strock.yml -o Global.pretrained_model={path/to/weights}/best_accuracy
```
Prediction:
```
# The configuration file used for prediction must match the training
python3 tools/infer_sr.py -c configs/sr/sr_tsrn_transformer_strock.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.infer_img=doc/imgs_words_en/word_52.png
```
![](../imgs_words_en/word_52.png)
After executing the command, the super-resolution result of the above image is as follows:
![](../imgs_results/sr_word_52.png)
<a name="4"></a>
## 4. Inference and Deployment
<a name="4-1"></a>
### 4.1 Python Inference
First, the model saved during the training process is converted into an inference model. ( [Model download link](https://paddleocr.bj.bcebos.com/sr_tsrn_transformer_strock_train.tar) ), you can use the following command to convert:
```shell
python3 tools/export_model.py -c configs/sr/sr_tsrn_transformer_strock.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.save_inference_dir=./inference/sr_out
```
For Text-Gestalt super-resolution model inference, the following commands can be executed:
```
python3 tools/infer/predict_sr.py --sr_model_dir=./inference/sr_out --image_dir=doc/imgs_words_en/word_52.png --sr_image_shape=3,32,128
```
After executing the command, the super-resolution result of the above image is as follows:
![](../imgs_results/sr_word_52.png)
<a name="4-2"></a>
### 4.2 C++ Inference
Not supported
<a name="4-3"></a>
### 4.3 Serving
Not supported
<a name="4-4"></a>
### 4.4 More
Not supported
<a name="5"></a>
## 5. FAQ
## Citation
```bibtex
@inproceedings{chen2022text,
title={Text gestalt: Stroke-aware scene text image super-resolution},
author={Chen, Jingye and Yu, Haiyang and Ma, Jianqi and Li, Bin and Xue, Xiangyang},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={36},
number={1},
pages={285--293},
year={2022}
}
```
......@@ -34,7 +34,7 @@ import paddle.distributed as dist
from ppocr.data.imaug import transform, create_operators
from ppocr.data.simple_dataset import SimpleDataSet
from ppocr.data.lmdb_dataset import LMDBDataSet
from ppocr.data.lmdb_dataset import LMDBDataSet, LMDBDataSetSR
from ppocr.data.pgnet_dataset import PGDataSet
from ppocr.data.pubtab_dataset import PubTabDataSet
......@@ -54,7 +54,8 @@ def build_dataloader(config, mode, device, logger, seed=None):
config = copy.deepcopy(config)
support_dict = [
'SimpleDataSet', 'LMDBDataSet', 'PGDataSet', 'PubTabDataSet'
'SimpleDataSet', 'LMDBDataSet', 'PGDataSet', 'PubTabDataSet',
'LMDBDataSetSR'
]
module_name = config[mode]['dataset']['name']
assert module_name in support_dict, Exception(
......
......@@ -1236,6 +1236,54 @@ class ABINetLabelEncode(BaseRecLabelEncode):
return dict_character
class SRLabelEncode(BaseRecLabelEncode):
def __init__(self,
max_text_length,
character_dict_path=None,
use_space_char=False,
**kwargs):
super(SRLabelEncode, self).__init__(max_text_length,
character_dict_path, use_space_char)
self.dic = {}
with open(character_dict_path, 'r') as fin:
for line in fin.readlines():
line = line.strip()
character, sequence = line.split()
self.dic[character] = sequence
english_stroke_alphabet = '0123456789'
self.english_stroke_dict = {}
for index in range(len(english_stroke_alphabet)):
self.english_stroke_dict[english_stroke_alphabet[index]] = index
def encode(self, label):
stroke_sequence = ''
for character in label:
if character not in self.dic:
continue
else:
stroke_sequence += self.dic[character]
stroke_sequence += '0'
label = stroke_sequence
length = len(label)
input_tensor = np.zeros(self.max_text_len).astype("int64")
for j in range(length - 1):
input_tensor[j + 1] = self.english_stroke_dict[label[j]]
return length, input_tensor
def __call__(self, data):
text = data['label']
length, input_tensor = self.encode(text)
data["length"] = length
data["input_tensor"] = input_tensor
if text is None:
return None
return data
class SPINLabelEncode(AttnLabelEncode):
""" Convert between text-label and text-index """
......
......@@ -24,6 +24,7 @@ import six
import cv2
import numpy as np
import math
from PIL import Image
class DecodeImage(object):
......@@ -440,3 +441,52 @@ class KieResize(object):
points[:, 0::2] = np.clip(points[:, 0::2], 0, img_shape[1])
points[:, 1::2] = np.clip(points[:, 1::2], 0, img_shape[0])
return points
class SRResize(object):
def __init__(self,
imgH=32,
imgW=128,
down_sample_scale=4,
keep_ratio=False,
min_ratio=1,
mask=False,
infer_mode=False,
**kwargs):
self.imgH = imgH
self.imgW = imgW
self.keep_ratio = keep_ratio
self.min_ratio = min_ratio
self.down_sample_scale = down_sample_scale
self.mask = mask
self.infer_mode = infer_mode
def __call__(self, data):
imgH = self.imgH
imgW = self.imgW
images_lr = data["image_lr"]
transform2 = ResizeNormalize(
(imgW // self.down_sample_scale, imgH // self.down_sample_scale))
images_lr = transform2(images_lr)
data["img_lr"] = images_lr
if self.infer_mode:
return data
images_HR = data["image_hr"]
label_strs = data["label"]
transform = ResizeNormalize((imgW, imgH))
images_HR = transform(images_HR)
data["img_hr"] = images_HR
return data
class ResizeNormalize(object):
def __init__(self, size, interpolation=Image.BICUBIC):
self.size = size
self.interpolation = interpolation
def __call__(self, img):
img = img.resize(self.size, self.interpolation)
img_numpy = np.array(img).astype("float32")
img_numpy = img_numpy.transpose((2, 0, 1)) / 255
return img_numpy
......@@ -16,6 +16,9 @@ import os
from paddle.io import Dataset
import lmdb
import cv2
import string
import six
from PIL import Image
from .imaug import transform, create_operators
......@@ -116,3 +119,58 @@ class LMDBDataSet(Dataset):
def __len__(self):
return self.data_idx_order_list.shape[0]
class LMDBDataSetSR(LMDBDataSet):
def buf2PIL(self, txn, key, type='RGB'):
imgbuf = txn.get(key)
buf = six.BytesIO()
buf.write(imgbuf)
buf.seek(0)
im = Image.open(buf).convert(type)
return im
def str_filt(self, str_, voc_type):
alpha_dict = {
'digit': string.digits,
'lower': string.digits + string.ascii_lowercase,
'upper': string.digits + string.ascii_letters,
'all': string.digits + string.ascii_letters + string.punctuation
}
if voc_type == 'lower':
str_ = str_.lower()
for char in str_:
if char not in alpha_dict[voc_type]:
str_ = str_.replace(char, '')
return str_
def get_lmdb_sample_info(self, txn, index):
self.voc_type = 'upper'
self.max_len = 100
self.test = False
label_key = b'label-%09d' % index
word = str(txn.get(label_key).decode())
img_HR_key = b'image_hr-%09d' % index # 128*32
img_lr_key = b'image_lr-%09d' % index # 64*16
try:
img_HR = self.buf2PIL(txn, img_HR_key, 'RGB')
img_lr = self.buf2PIL(txn, img_lr_key, 'RGB')
except IOError or len(word) > self.max_len:
return self[index + 1]
label_str = self.str_filt(word, self.voc_type)
return img_HR, img_lr, label_str
def __getitem__(self, idx):
lmdb_idx, file_idx = self.data_idx_order_list[idx]
lmdb_idx = int(lmdb_idx)
file_idx = int(file_idx)
sample_info = self.get_lmdb_sample_info(self.lmdb_sets[lmdb_idx]['txn'],
file_idx)
if sample_info is None:
return self.__getitem__(np.random.randint(self.__len__()))
img_HR, img_lr, label_str = sample_info
data = {'image_hr': img_HR, 'image_lr': img_lr, 'label': label_str}
outs = transform(data, self.ops)
if outs is None:
return self.__getitem__(np.random.randint(self.__len__()))
return outs
......@@ -57,6 +57,9 @@ from .table_master_loss import TableMasterLoss
# vqa token loss
from .vqa_token_layoutlm_loss import VQASerTokenLayoutLMLoss
# sr loss
from .stroke_focus_loss import StrokeFocusLoss
def build_loss(config):
support_dict = [
......@@ -64,7 +67,7 @@ def build_loss(config):
'ClsLoss', 'AttentionLoss', 'SRNLoss', 'PGLoss', 'CombinedLoss',
'CELoss', 'TableAttentionLoss', 'SARLoss', 'AsterLoss', 'SDMGRLoss',
'VQASerTokenLayoutLMLoss', 'LossFromOutput', 'PRENLoss', 'MultiLoss',
'TableMasterLoss', 'SPINAttentionLoss', 'VLLoss'
'TableMasterLoss', 'SPINAttentionLoss', 'VLLoss', 'StrokeFocusLoss'
]
config = copy.deepcopy(config)
module_name = config.pop('name')
......
# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This code is refer from:
https://github.com/FudanVI/FudanOCR/blob/main/text-gestalt/loss/stroke_focus_loss.py
"""
import cv2
import sys
import time
import string
import random
import numpy as np
import paddle.nn as nn
import paddle
class StrokeFocusLoss(nn.Layer):
def __init__(self, character_dict_path=None, **kwargs):
super(StrokeFocusLoss, self).__init__(character_dict_path)
self.mse_loss = nn.MSELoss()
self.ce_loss = nn.CrossEntropyLoss()
self.l1_loss = nn.L1Loss()
self.english_stroke_alphabet = '0123456789'
self.english_stroke_dict = {}
for index in range(len(self.english_stroke_alphabet)):
self.english_stroke_dict[self.english_stroke_alphabet[
index]] = index
stroke_decompose_lines = open(character_dict_path, 'r').readlines()
self.dic = {}
for line in stroke_decompose_lines:
line = line.strip()
character, sequence = line.split()
self.dic[character] = sequence
def forward(self, pred, data):
sr_img = pred["sr_img"]
hr_img = pred["hr_img"]
mse_loss = self.mse_loss(sr_img, hr_img)
word_attention_map_gt = pred["word_attention_map_gt"]
word_attention_map_pred = pred["word_attention_map_pred"]
hr_pred = pred["hr_pred"]
sr_pred = pred["sr_pred"]
attention_loss = paddle.nn.functional.l1_loss(word_attention_map_gt,
word_attention_map_pred)
loss = (mse_loss + attention_loss * 50) * 100
return {
"mse_loss": mse_loss,
"attention_loss": attention_loss,
"loss": loss
}
......@@ -30,13 +30,13 @@ from .table_metric import TableMetric
from .kie_metric import KIEMetric
from .vqa_token_ser_metric import VQASerTokenMetric
from .vqa_token_re_metric import VQAReTokenMetric
from .sr_metric import SRMetric
def build_metric(config):
support_dict = [
"DetMetric", "DetFCEMetric", "RecMetric", "ClsMetric", "E2EMetric",
"DistillationMetric", "TableMetric", 'KIEMetric', 'VQASerTokenMetric',
'VQAReTokenMetric'
'VQAReTokenMetric', 'SRMetric'
]
config = copy.deepcopy(config)
......
......@@ -16,6 +16,7 @@ import Levenshtein
import string
class RecMetric(object):
def __init__(self,
main_indicator='acc',
......
# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
https://github.com/FudanVI/FudanOCR/blob/main/text-gestalt/utils/ssim_psnr.py
"""
from math import exp
import paddle
import paddle.nn.functional as F
import paddle.nn as nn
import string
class SSIM(nn.Layer):
def __init__(self, window_size=11, size_average=True):
super(SSIM, self).__init__()
self.window_size = window_size
self.size_average = size_average
self.channel = 1
self.window = self.create_window(window_size, self.channel)
def gaussian(self, window_size, sigma):
gauss = paddle.to_tensor([
exp(-(x - window_size // 2)**2 / float(2 * sigma**2))
for x in range(window_size)
])
return gauss / gauss.sum()
def create_window(self, window_size, channel):
_1D_window = self.gaussian(window_size, 1.5).unsqueeze(1)
_2D_window = _1D_window.mm(_1D_window.t()).unsqueeze(0).unsqueeze(0)
window = _2D_window.expand([channel, 1, window_size, window_size])
return window
def _ssim(self, img1, img2, window, window_size, channel,
size_average=True):
mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)
mu2 = F.conv2d(img2, window, padding=window_size // 2, groups=channel)
mu1_sq = mu1.pow(2)
mu2_sq = mu2.pow(2)
mu1_mu2 = mu1 * mu2
sigma1_sq = F.conv2d(
img1 * img1, window, padding=window_size // 2,
groups=channel) - mu1_sq
sigma2_sq = F.conv2d(
img2 * img2, window, padding=window_size // 2,
groups=channel) - mu2_sq
sigma12 = F.conv2d(
img1 * img2, window, padding=window_size // 2,
groups=channel) - mu1_mu2
C1 = 0.01**2
C2 = 0.03**2
ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / (
(mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))
if size_average:
return ssim_map.mean()
else:
return ssim_map.mean([1, 2, 3])
def ssim(self, img1, img2, window_size=11, size_average=True):
(_, channel, _, _) = img1.shape
window = self.create_window(window_size, channel)
return self._ssim(img1, img2, window, window_size, channel,
size_average)
def forward(self, img1, img2):
(_, channel, _, _) = img1.shape
if channel == self.channel and self.window.dtype == img1.dtype:
window = self.window
else:
window = self.create_window(self.window_size, channel)
self.window = window
self.channel = channel
return self._ssim(img1, img2, window, self.window_size, channel,
self.size_average)
class SRMetric(object):
def __init__(self, main_indicator='all', **kwargs):
self.main_indicator = main_indicator
self.eps = 1e-5
self.psnr_result = []
self.ssim_result = []
self.calculate_ssim = SSIM()
self.reset()
def reset(self):
self.correct_num = 0
self.all_num = 0
self.norm_edit_dis = 0
self.psnr_result = []
self.ssim_result = []
def calculate_psnr(self, img1, img2):
# img1 and img2 have range [0, 1]
mse = ((img1 * 255 - img2 * 255)**2).mean()
if mse == 0:
return float('inf')
return 20 * paddle.log10(255.0 / paddle.sqrt(mse))
def _normalize_text(self, text):
text = ''.join(
filter(lambda x: x in (string.digits + string.ascii_letters), text))
return text.lower()
def __call__(self, pred_label, *args, **kwargs):
metric = {}
images_sr = pred_label["sr_img"]
images_hr = pred_label["hr_img"]
psnr = self.calculate_psnr(images_sr, images_hr)
ssim = self.calculate_ssim(images_sr, images_hr)
self.psnr_result.append(psnr)
self.ssim_result.append(ssim)
def get_metric(self):
"""
return metrics {
'acc': 0,
'norm_edit_dis': 0,
}
"""
self.psnr_avg = sum(self.psnr_result) / len(self.psnr_result)
self.psnr_avg = round(self.psnr_avg.item(), 6)
self.ssim_avg = sum(self.ssim_result) / len(self.ssim_result)
self.ssim_avg = round(self.ssim_avg.item(), 6)
self.all_avg = self.psnr_avg + self.ssim_avg
self.reset()
return {
'psnr_avg': self.psnr_avg,
"ssim_avg": self.ssim_avg,
"all": self.all_avg
}
......@@ -14,6 +14,7 @@
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from paddle import nn
from ppocr.modeling.transforms import build_transform
from ppocr.modeling.backbones import build_backbone
......@@ -46,6 +47,10 @@ class BaseModel(nn.Layer):
in_channels = self.transform.out_channels
# build backbone, backbone is need for del, rec and cls
if 'Backbone' not in config or config['Backbone'] is None:
self.use_backbone = False
else:
self.use_backbone = True
config["Backbone"]['in_channels'] = in_channels
self.backbone = build_backbone(config["Backbone"], model_type)
in_channels = self.backbone.out_channels
......@@ -77,6 +82,7 @@ class BaseModel(nn.Layer):
y = dict()
if self.use_transform:
x = self.transform(x)
if self.use_backbone:
x = self.backbone(x)
if isinstance(x, dict):
y.update(x)
......
......@@ -113,7 +113,6 @@ class LayoutLMv2ForSer(NLPBaseModel):
pretrained,
checkpoints,
num_classes=num_classes)
self.use_visual_backbone = True
if hasattr(self.model.layoutlmv2, "use_visual_backbone"
) and self.model.layoutlmv2.use_visual_backbone is False:
self.use_visual_backbone = False
......@@ -155,7 +154,9 @@ class LayoutXLMForSer(NLPBaseModel):
pretrained,
checkpoints,
num_classes=num_classes)
self.use_visual_backbone = True
if hasattr(self.model.layoutxlm, "use_visual_backbone"
) and self.model.layoutxlm.use_visual_backbone is False:
self.use_visual_backbone = False
def forward(self, x):
if self.use_visual_backbone is True:
......@@ -185,6 +186,9 @@ class LayoutLMv2ForRe(NLPBaseModel):
super(LayoutLMv2ForRe, self).__init__(
LayoutLMv2Model, LayoutLMv2ForRelationExtraction, mode, "re",
pretrained, checkpoints)
if hasattr(self.model.layoutlmv2, "use_visual_backbone"
) and self.model.layoutlmv2.use_visual_backbone is False:
self.use_visual_backbone = False
def forward(self, x):
x = self.model(
......@@ -207,7 +211,6 @@ class LayoutXLMForRe(NLPBaseModel):
super(LayoutXLMForRe, self).__init__(
LayoutXLMModel, LayoutXLMForRelationExtraction, mode, "re",
pretrained, checkpoints)
self.use_visual_backbone = True
if hasattr(self.model.layoutxlm, "use_visual_backbone"
) and self.model.layoutxlm.use_visual_backbone is False:
self.use_visual_backbone = False
......
# copyright (c) 2022 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This code is refer from:
https://github.com/FudanVI/FudanOCR/blob/main/text-gestalt/loss/transformer_english_decomposition.py
"""
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
import math, copy
import numpy as np
# stroke-level alphabet
alphabet = '0123456789'
def get_alphabet_len():
return len(alphabet)
def subsequent_mask(size):
"""Generate a square mask for the sequence. The masked positions are filled with float('-inf').
Unmasked positions are filled with float(0.0).
"""
mask = paddle.ones([1, size, size], dtype='float32')
mask_inf = paddle.triu(
paddle.full(
shape=[1, size, size], dtype='float32', fill_value='-inf'),
diagonal=1)
mask = mask + mask_inf
padding_mask = paddle.equal(mask, paddle.to_tensor(1, dtype=mask.dtype))
return padding_mask
def clones(module, N):
return nn.LayerList([copy.deepcopy(module) for _ in range(N)])
def masked_fill(x, mask, value):
y = paddle.full(x.shape, value, x.dtype)
return paddle.where(mask, y, x)
def attention(query, key, value, mask=None, dropout=None, attention_map=None):
d_k = query.shape[-1]
scores = paddle.matmul(query,
paddle.transpose(key, [0, 1, 3, 2])) / math.sqrt(d_k)
if mask is not None:
scores = masked_fill(scores, mask == 0, float('-inf'))
else:
pass
p_attn = F.softmax(scores, axis=-1)
if dropout is not None:
p_attn = dropout(p_attn)
return paddle.matmul(p_attn, value), p_attn
class MultiHeadedAttention(nn.Layer):
def __init__(self, h, d_model, dropout=0.1, compress_attention=False):
super(MultiHeadedAttention, self).__init__()
assert d_model % h == 0
self.d_k = d_model // h
self.h = h
self.linears = clones(nn.Linear(d_model, d_model), 4)
self.attn = None
self.dropout = nn.Dropout(p=dropout, mode="downscale_in_infer")
self.compress_attention = compress_attention
self.compress_attention_linear = nn.Linear(h, 1)
def forward(self, query, key, value, mask=None, attention_map=None):
if mask is not None:
mask = mask.unsqueeze(1)
nbatches = query.shape[0]
query, key, value = \
[paddle.transpose(l(x).reshape([nbatches, -1, self.h, self.d_k]), [0,2,1,3])
for l, x in zip(self.linears, (query, key, value))]
x, attention_map = attention(
query,
key,
value,
mask=mask,
dropout=self.dropout,
attention_map=attention_map)
x = paddle.reshape(
paddle.transpose(x, [0, 2, 1, 3]),
[nbatches, -1, self.h * self.d_k])
return self.linears[-1](x), attention_map
class ResNet(nn.Layer):
def __init__(self, num_in, block, layers):
super(ResNet, self).__init__()
self.conv1 = nn.Conv2D(num_in, 64, kernel_size=3, stride=1, padding=1)
self.bn1 = nn.BatchNorm2D(64, use_global_stats=True)
self.relu1 = nn.ReLU()
self.pool = nn.MaxPool2D((2, 2), (2, 2))
self.conv2 = nn.Conv2D(64, 128, kernel_size=3, stride=1, padding=1)
self.bn2 = nn.BatchNorm2D(128, use_global_stats=True)
self.relu2 = nn.ReLU()
self.layer1_pool = nn.MaxPool2D((2, 2), (2, 2))
self.layer1 = self._make_layer(block, 128, 256, layers[0])
self.layer1_conv = nn.Conv2D(256, 256, 3, 1, 1)
self.layer1_bn = nn.BatchNorm2D(256, use_global_stats=True)
self.layer1_relu = nn.ReLU()
self.layer2_pool = nn.MaxPool2D((2, 2), (2, 2))
self.layer2 = self._make_layer(block, 256, 256, layers[1])
self.layer2_conv = nn.Conv2D(256, 256, 3, 1, 1)
self.layer2_bn = nn.BatchNorm2D(256, use_global_stats=True)
self.layer2_relu = nn.ReLU()
self.layer3_pool = nn.MaxPool2D((2, 2), (2, 2))
self.layer3 = self._make_layer(block, 256, 512, layers[2])
self.layer3_conv = nn.Conv2D(512, 512, 3, 1, 1)
self.layer3_bn = nn.BatchNorm2D(512, use_global_stats=True)
self.layer3_relu = nn.ReLU()
self.layer4_pool = nn.MaxPool2D((2, 2), (2, 2))
self.layer4 = self._make_layer(block, 512, 512, layers[3])
self.layer4_conv2 = nn.Conv2D(512, 1024, 3, 1, 1)
self.layer4_conv2_bn = nn.BatchNorm2D(1024, use_global_stats=True)
self.layer4_conv2_relu = nn.ReLU()
def _make_layer(self, block, inplanes, planes, blocks):
if inplanes != planes:
downsample = nn.Sequential(
nn.Conv2D(inplanes, planes, 3, 1, 1),
nn.BatchNorm2D(
planes, use_global_stats=True), )
else:
downsample = None
layers = []
layers.append(block(inplanes, planes, downsample))
for i in range(1, blocks):
layers.append(block(planes, planes, downsample=None))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu1(x)
x = self.pool(x)
x = self.conv2(x)
x = self.bn2(x)
x = self.relu2(x)
x = self.layer1_pool(x)
x = self.layer1(x)
x = self.layer1_conv(x)
x = self.layer1_bn(x)
x = self.layer1_relu(x)
x = self.layer2(x)
x = self.layer2_conv(x)
x = self.layer2_bn(x)
x = self.layer2_relu(x)
x = self.layer3(x)
x = self.layer3_conv(x)
x = self.layer3_bn(x)
x = self.layer3_relu(x)
x = self.layer4(x)
x = self.layer4_conv2(x)
x = self.layer4_conv2_bn(x)
x = self.layer4_conv2_relu(x)
return x
class Bottleneck(nn.Layer):
def __init__(self, input_dim):
super(Bottleneck, self).__init__()
self.conv1 = nn.Conv2D(input_dim, input_dim, 1)
self.bn1 = nn.BatchNorm2D(input_dim, use_global_stats=True)
self.relu = nn.ReLU()
self.conv2 = nn.Conv2D(input_dim, input_dim, 3, 1, 1)
self.bn2 = nn.BatchNorm2D(input_dim, use_global_stats=True)
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out += residual
out = self.relu(out)
return out
class PositionalEncoding(nn.Layer):
"Implement the PE function."
def __init__(self, dropout, dim, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout, mode="downscale_in_infer")
pe = paddle.zeros([max_len, dim])
position = paddle.arange(0, max_len, dtype=paddle.float32).unsqueeze(1)
div_term = paddle.exp(
paddle.arange(0, dim, 2).astype('float32') *
(-math.log(10000.0) / dim))
pe[:, 0::2] = paddle.sin(position * div_term)
pe[:, 1::2] = paddle.cos(position * div_term)
pe = paddle.unsqueeze(pe, 0)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:, :paddle.shape(x)[1]]
return self.dropout(x)
class PositionwiseFeedForward(nn.Layer):
"Implements FFN equation."
def __init__(self, d_model, d_ff, dropout=0.1):
super(PositionwiseFeedForward, self).__init__()
self.w_1 = nn.Linear(d_model, d_ff)
self.w_2 = nn.Linear(d_ff, d_model)
self.dropout = nn.Dropout(dropout, mode="downscale_in_infer")
def forward(self, x):
return self.w_2(self.dropout(F.relu(self.w_1(x))))
class Generator(nn.Layer):
"Define standard linear + softmax generation step."
def __init__(self, d_model, vocab):
super(Generator, self).__init__()
self.proj = nn.Linear(d_model, vocab)
self.relu = nn.ReLU()
def forward(self, x):
out = self.proj(x)
return out
class Embeddings(nn.Layer):
def __init__(self, d_model, vocab):
super(Embeddings, self).__init__()
self.lut = nn.Embedding(vocab, d_model)
self.d_model = d_model
def forward(self, x):
embed = self.lut(x) * math.sqrt(self.d_model)
return embed
class LayerNorm(nn.Layer):
"Construct a layernorm module (See citation for details)."
def __init__(self, features, eps=1e-6):
super(LayerNorm, self).__init__()
self.a_2 = self.create_parameter(
shape=[features],
default_initializer=paddle.nn.initializer.Constant(1.0))
self.b_2 = self.create_parameter(
shape=[features],
default_initializer=paddle.nn.initializer.Constant(0.0))
self.eps = eps
def forward(self, x):
mean = x.mean(-1, keepdim=True)
std = x.std(-1, keepdim=True)
return self.a_2 * (x - mean) / (std + self.eps) + self.b_2
class Decoder(nn.Layer):
def __init__(self):
super(Decoder, self).__init__()
self.mask_multihead = MultiHeadedAttention(
h=16, d_model=1024, dropout=0.1)
self.mul_layernorm1 = LayerNorm(1024)
self.multihead = MultiHeadedAttention(h=16, d_model=1024, dropout=0.1)
self.mul_layernorm2 = LayerNorm(1024)
self.pff = PositionwiseFeedForward(1024, 2048)
self.mul_layernorm3 = LayerNorm(1024)
def forward(self, text, conv_feature, attention_map=None):
text_max_length = text.shape[1]
mask = subsequent_mask(text_max_length)
result = text
result = self.mul_layernorm1(result + self.mask_multihead(
text, text, text, mask=mask)[0])
b, c, h, w = conv_feature.shape
conv_feature = paddle.transpose(
conv_feature.reshape([b, c, h * w]), [0, 2, 1])
word_image_align, attention_map = self.multihead(
result,
conv_feature,
conv_feature,
mask=None,
attention_map=attention_map)
result = self.mul_layernorm2(result + word_image_align)
result = self.mul_layernorm3(result + self.pff(result))
return result, attention_map
class BasicBlock(nn.Layer):
def __init__(self, inplanes, planes, downsample):
super(BasicBlock, self).__init__()
self.conv1 = nn.Conv2D(
inplanes, planes, kernel_size=3, stride=1, padding=1)
self.bn1 = nn.BatchNorm2D(planes, use_global_stats=True)
self.relu = nn.ReLU()
self.conv2 = nn.Conv2D(
planes, planes, kernel_size=3, stride=1, padding=1)
self.bn2 = nn.BatchNorm2D(planes, use_global_stats=True)
self.downsample = downsample
def forward(self, x):
residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample != None:
residual = self.downsample(residual)
out += residual
out = self.relu(out)
return out
class Encoder(nn.Layer):
def __init__(self):
super(Encoder, self).__init__()
self.cnn = ResNet(num_in=1, block=BasicBlock, layers=[1, 2, 5, 3])
def forward(self, input):
conv_result = self.cnn(input)
return conv_result
class Transformer(nn.Layer):
def __init__(self, in_channels=1):
super(Transformer, self).__init__()
word_n_class = get_alphabet_len()
self.embedding_word_with_upperword = Embeddings(512, word_n_class)
self.pe = PositionalEncoding(dim=512, dropout=0.1, max_len=5000)
self.encoder = Encoder()
self.decoder = Decoder()
self.generator_word_with_upperword = Generator(1024, word_n_class)
for p in self.parameters():
if p.dim() > 1:
nn.initializer.XavierNormal(p)
def forward(self, image, text_length, text_input, attention_map=None):
if image.shape[1] == 3:
R = image[:, 0:1, :, :]
G = image[:, 1:2, :, :]
B = image[:, 2:3, :, :]
image = 0.299 * R + 0.587 * G + 0.114 * B
conv_feature = self.encoder(image) # batch, 1024, 8, 32
max_length = max(text_length)
text_input = text_input[:, :max_length]
text_embedding = self.embedding_word_with_upperword(
text_input) # batch, text_max_length, 512
postion_embedding = self.pe(
paddle.zeros(text_embedding.shape)) # batch, text_max_length, 512
text_input_with_pe = paddle.concat([text_embedding, postion_embedding],
2) # batch, text_max_length, 1024
batch, seq_len, _ = text_input_with_pe.shape
text_input_with_pe, word_attention_map = self.decoder(
text_input_with_pe, conv_feature)
word_decoder_result = self.generator_word_with_upperword(
text_input_with_pe)
if self.training:
total_length = paddle.sum(text_length)
probs_res = paddle.zeros([total_length, get_alphabet_len()])
start = 0
for index, length in enumerate(text_length):
length = int(length.numpy())
probs_res[start:start + length, :] = word_decoder_result[
index, 0:0 + length, :]
start = start + length
return probs_res, word_attention_map, None
else:
return word_decoder_result
......@@ -18,10 +18,10 @@ __all__ = ['build_transform']
def build_transform(config):
from .tps import TPS
from .stn import STN_ON
from .tsrn import TSRN
from .gaspin_transformer import GA_SPIN_Transformer as GA_SPIN
support_dict = ['TPS', 'STN_ON', 'GA_SPIN']
support_dict = ['TPS', 'STN_ON', 'GA_SPIN', 'TSRN']
module_name = config.pop('name')
assert module_name in support_dict, Exception(
......
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
This code is refer from:
https://github.com/FudanVI/FudanOCR/blob/main/text-gestalt/model/tsrn.py
"""
import math
import paddle
import paddle.nn.functional as F
from paddle import nn
from collections import OrderedDict
import sys
import numpy as np
import warnings
import math, copy
import cv2
warnings.filterwarnings("ignore")
from .tps_spatial_transformer import TPSSpatialTransformer
from .stn import STN as STN_model
from ppocr.modeling.heads.sr_rensnet_transformer import Transformer
class TSRN(nn.Layer):
def __init__(self,
in_channels,
scale_factor=2,
width=128,
height=32,
STN=False,
srb_nums=5,
mask=False,
hidden_units=32,
infer_mode=False,
**kwargs):
super(TSRN, self).__init__()
in_planes = 3
if mask:
in_planes = 4
assert math.log(scale_factor, 2) % 1 == 0
upsample_block_num = int(math.log(scale_factor, 2))
self.block1 = nn.Sequential(
nn.Conv2D(
in_planes, 2 * hidden_units, kernel_size=9, padding=4),
nn.PReLU())
self.srb_nums = srb_nums
for i in range(srb_nums):
setattr(self, 'block%d' % (i + 2),
RecurrentResidualBlock(2 * hidden_units))
setattr(
self,
'block%d' % (srb_nums + 2),
nn.Sequential(
nn.Conv2D(
2 * hidden_units,
2 * hidden_units,
kernel_size=3,
padding=1),
nn.BatchNorm2D(2 * hidden_units)))
block_ = [
UpsampleBLock(2 * hidden_units, 2)
for _ in range(upsample_block_num)
]
block_.append(
nn.Conv2D(
2 * hidden_units, in_planes, kernel_size=9, padding=4))
setattr(self, 'block%d' % (srb_nums + 3), nn.Sequential(*block_))
self.tps_inputsize = [height // scale_factor, width // scale_factor]
tps_outputsize = [height // scale_factor, width // scale_factor]
num_control_points = 20
tps_margins = [0.05, 0.05]
self.stn = STN
if self.stn:
self.tps = TPSSpatialTransformer(
output_image_size=tuple(tps_outputsize),
num_control_points=num_control_points,
margins=tuple(tps_margins))
self.stn_head = STN_model(
in_channels=in_planes,
num_ctrlpoints=num_control_points,
activation='none')
self.out_channels = in_channels
self.r34_transformer = Transformer()
for param in self.r34_transformer.parameters():
param.trainable = False
self.infer_mode = infer_mode
def forward(self, x):
output = {}
if self.infer_mode:
output["lr_img"] = x
y = x
else:
output["lr_img"] = x[0]
output["hr_img"] = x[1]
y = x[0]
if self.stn and self.training:
_, ctrl_points_x = self.stn_head(y)
y, _ = self.tps(y, ctrl_points_x)
block = {'1': self.block1(y)}
for i in range(self.srb_nums + 1):
block[str(i + 2)] = getattr(self,
'block%d' % (i + 2))(block[str(i + 1)])
block[str(self.srb_nums + 3)] = getattr(self, 'block%d' % (self.srb_nums + 3)) \
((block['1'] + block[str(self.srb_nums + 2)]))
sr_img = paddle.tanh(block[str(self.srb_nums + 3)])
output["sr_img"] = sr_img
if self.training:
hr_img = x[1]
length = x[2]
input_tensor = x[3]
# add transformer
sr_pred, word_attention_map_pred, _ = self.r34_transformer(
sr_img, length, input_tensor)
hr_pred, word_attention_map_gt, _ = self.r34_transformer(
hr_img, length, input_tensor)
output["hr_img"] = hr_img
output["hr_pred"] = hr_pred
output["word_attention_map_gt"] = word_attention_map_gt
output["sr_pred"] = sr_pred
output["word_attention_map_pred"] = word_attention_map_pred
return output
class RecurrentResidualBlock(nn.Layer):
def __init__(self, channels):
super(RecurrentResidualBlock, self).__init__()
self.conv1 = nn.Conv2D(channels, channels, kernel_size=3, padding=1)
self.bn1 = nn.BatchNorm2D(channels)
self.gru1 = GruBlock(channels, channels)
self.prelu = mish()
self.conv2 = nn.Conv2D(channels, channels, kernel_size=3, padding=1)
self.bn2 = nn.BatchNorm2D(channels)
self.gru2 = GruBlock(channels, channels)
def forward(self, x):
residual = self.conv1(x)
residual = self.bn1(residual)
residual = self.prelu(residual)
residual = self.conv2(residual)
residual = self.bn2(residual)
residual = self.gru1(residual.transpose([0, 1, 3, 2])).transpose(
[0, 1, 3, 2])
return self.gru2(x + residual)
class UpsampleBLock(nn.Layer):
def __init__(self, in_channels, up_scale):
super(UpsampleBLock, self).__init__()
self.conv = nn.Conv2D(
in_channels, in_channels * up_scale**2, kernel_size=3, padding=1)
self.pixel_shuffle = nn.PixelShuffle(up_scale)
self.prelu = mish()
def forward(self, x):
x = self.conv(x)
x = self.pixel_shuffle(x)
x = self.prelu(x)
return x
class mish(nn.Layer):
def __init__(self, ):
super(mish, self).__init__()
self.activated = True
def forward(self, x):
if self.activated:
x = x * (paddle.tanh(F.softplus(x)))
return x
class GruBlock(nn.Layer):
def __init__(self, in_channels, out_channels):
super(GruBlock, self).__init__()
assert out_channels % 2 == 0
self.conv1 = nn.Conv2D(
in_channels, out_channels, kernel_size=1, padding=0)
self.gru = nn.GRU(out_channels,
out_channels // 2,
direction='bidirectional')
def forward(self, x):
# x: b, c, w, h
x = self.conv1(x)
x = x.transpose([0, 2, 3, 1]) # b, w, h, c
batch_size, w, h, c = x.shape
x = x.reshape([-1, h, c]) # b*w, h, c
x, _ = self.gru(x)
x = x.reshape([-1, w, h, c])
x = x.transpose([0, 3, 1, 2])
return x
......@@ -780,7 +780,7 @@ class VLLabelDecode(BaseRecLabelDecode):
) + length[i])].topk(1)[0][:, 0]
preds_prob = paddle.exp(
paddle.log(preds_prob).sum() / (preds_prob.shape[0] + 1e-6))
text.append((preds_text, preds_prob))
text.append((preds_text, preds_prob.numpy()[0]))
if label is None:
return text
label = self.decode(label)
......
......@@ -56,7 +56,7 @@ def load_model(config, model, optimizer=None, model_type='det'):
is_float16 = False
if model_type == 'vqa':
# NOTE: for vqa model, resume training is not supported now
# NOTE: for vqa model dsitillation, resume training is not supported now
if config["Architecture"]["algorithm"] in ["Distillation"]:
return best_model_dict
checkpoints = config['Architecture']['Backbone']['checkpoints']
......@@ -148,10 +148,14 @@ def load_pretrained_params(model, path):
"The {}.pdparams does not exists!".format(path)
params = paddle.load(path + '.pdparams')
state_dict = model.state_dict()
new_state_dict = {}
is_float16 = False
for k1 in params.keys():
if k1 not in state_dict.keys():
logger.warning("The pretrained params {} not in model".format(k1))
else:
......@@ -187,7 +191,6 @@ def save_model(model,
"""
_mkdir_if_not_exist(model_path, logger)
model_prefix = os.path.join(model_path, prefix)
if config['Architecture']["model_type"] != 'vqa':
paddle.save(optimizer.state_dict(), model_prefix + '.pdopt')
if config['Architecture']["model_type"] != 'vqa':
paddle.save(model.state_dict(), model_prefix + '.pdparams')
......
- [关键信息提取(Key Information Extraction)](#关键信息提取key-information-extraction)
- [1. 快速使用](#1-快速使用)
- [2. 执行训练](#2-执行训练)
- [3. 执行评估](#3-执行评估)
- [4. 参考文献](#4-参考文献)
# 关键信息提取(Key Information Extraction)
本节介绍PaddleOCR中关键信息提取SDMGR方法的快速使用和训练方法。
SDMGR是一个关键信息提取算法,将每个检测到的文本区域分类为预定义的类别,如订单ID、发票号码,金额等。
## 1. 快速使用
训练和测试的数据采用wildreceipt数据集,通过如下指令下载数据集:
```
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/wildreceipt.tar && tar xf wildreceipt.tar
```
执行预测:
```
cd PaddleOCR/
wget https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar && tar xf kie_vgg16.tar
python3.7 tools/infer_kie.py -c configs/kie/kie_unet_sdmgr.yml -o Global.checkpoints=kie_vgg16/best_accuracy Global.infer_img=../wildreceipt/1.txt
```
执行预测后的结果保存在`./output/sdmgr_kie/predicts_kie.txt`文件中,可视化结果保存在`/output/sdmgr_kie/kie_results/`目录下。
可视化结果如下图所示:
<div align="center">
<img src="./imgs/0.png" width="800">
</div>
## 2. 执行训练
创建数据集软链到PaddleOCR/train_data目录下:
```
cd PaddleOCR/ && mkdir train_data && cd train_data
ln -s ../../wildreceipt ./
```
训练采用的配置文件是configs/kie/kie_unet_sdmgr.yml,配置文件中默认训练数据路径是`train_data/wildreceipt`,准备好数据后,可以通过如下指令执行训练:
```
python3.7 tools/train.py -c configs/kie/kie_unet_sdmgr.yml -o Global.save_model_dir=./output/kie/
```
## 3. 执行评估
```
python3.7 tools/eval.py -c configs/kie/kie_unet_sdmgr.yml -o Global.checkpoints=./output/kie/best_accuracy
```
## 4. 参考文献
<!-- [ALGORITHM] -->
```bibtex
@misc{sun2021spatial,
title={Spatial Dual-Modality Graph Reasoning for Key Information Extraction},
author={Hongbin Sun and Zhanghui Kuang and Xiaoyu Yue and Chenhao Lin and Wayne Zhang},
year={2021},
eprint={2103.14470},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
......@@ -4,8 +4,7 @@
- [2. OCR和表格识别模型](#2-ocr和表格识别模型)
- [2.1 OCR](#21-ocr)
- [2.2 表格识别模型](#22-表格识别模型)
- [3. VQA模型](#3-vqa模型)
- [4. KIE模型](#4-kie模型)
- [3. KIE模型](#3-kie模型)
<a name="1"></a>
......@@ -38,19 +37,26 @@
|en_ppocr_mobile_v2.0_table_structure|PubTabNet数据集训练的英文表格场景的表格结构预测|18.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
<a name="3"></a>
## 3. VQA模型
|模型名称|模型简介|推理模型大小|下载地址|
| --- | --- | --- | --- |
|ser_LayoutXLM_xfun_zh|基于LayoutXLM在xfun中文数据集上训练的SER模型|1.4G|[推理模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|re_LayoutXLM_xfun_zh|基于LayoutXLM在xfun中文数据集上训练的RE模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
|ser_LayoutLMv2_xfun_zh|基于LayoutLMv2在xfun中文数据集上训练的SER模型|778M|[推理模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar) |
|re_LayoutLMv2_xfun_zh|基于LayoutLMv2在xfun中文数据集上训练的RE模型|765M|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar) |
|ser_LayoutLM_xfun_zh|基于LayoutLM在xfun中文数据集上训练的SER模型|430M|[推理模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) |
## 3. KIE模型
<a name="4"></a>
## 4. KIE模型
在XFUND_zh数据集上,不同模型的精度与V100 GPU上速度信息如下所示。
|模型名称|模型简介|模型大小|下载地址|
| --- | --- | --- | --- |
|SDMGR|关键信息提取模型|78M|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)|
|模型名称|模型简介 | 推理模型大小| 精度(hmean) | 预测耗时(ms) | 下载地址|
| --- | --- | --- |--- |--- | --- |
|ser_VI-LayoutXLM_xfund_zh|基于VI-LayoutXLM在xfund中文数据集上训练的SER模型|1.1G| 93.19% | 15.49 | [推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar) |
|re_VI-LayoutXLM_xfund_zh|基于VI-LayoutXLM在xfund中文数据集上训练的RE模型|1.1G| 83.92% | 15.49 |[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar) |
|ser_LayoutXLM_xfund_zh|基于LayoutXLM在xfund中文数据集上训练的SER模型|1.4G| 90.38% | 19.49 |[推理模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|re_LayoutXLM_xfund_zh|基于LayoutXLM在xfund中文数据集上训练的RE模型|1.4G| 74.83% | 19.49 |[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
|ser_LayoutLMv2_xfund_zh|基于LayoutLMv2在xfund中文数据集上训练的SER模型|778M| 85.44% | 31.46 |[推理模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar) |
|re_LayoutLMv2_xfund_zh|基于LayoutLMv2在xfun中文数据集上训练的RE模型|765M| 67.77% | 31.46 |[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar) |
|ser_LayoutLM_xfund_zh|基于LayoutLM在xfund中文数据集上训练的SER模型|430M| 77.31% | - |[推理模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) |
* 注:上述预测耗时信息仅包含了inference模型的推理耗时,没有统计预处理与后处理耗时,测试环境为`V100 GPU + CUDA 10.2 + CUDNN 8.1.1 + TRT 7.2.3.4`
在wildreceipt数据集上,SDMGR模型精度与下载地址如下所示。
|模型名称|模型简介|模型大小|精度|下载地址|
| --- | --- | --- |--- | --- |
|SDMGR|关键信息提取模型|78M| 86.70% | [推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)|
......@@ -51,6 +51,8 @@ def init_args():
"--ser_dict_path",
type=str,
default="../train_data/XFUND/class_list_xfun.txt")
# need to be None or tb-yx
parser.add_argument("--ocr_order_method", type=str, default=None)
# params for inference
parser.add_argument(
"--mode",
......
# 怎样完成基于图像数据的信息抽取任务
- [1. 简介](#1-简介)
- [1.1 背景](#11-背景)
- [1.2 主流方法](#12-主流方法)
- [2. 关键信息抽取任务流程](#2-关键信息抽取任务流程)
- [2.1 训练OCR模型](#21-训练OCR模型)
- [2.2 训练KIE模型](#22-训练KIE模型)
- [3. 参考文献](#3-参考文献)
## 1. 简介
### 1.1 背景
关键信息抽取 (Key Information Extraction, KIE)指的是是从文本或者图像中,抽取出关键的信息。针对文档图像的关键信息抽取任务作为OCR的下游任务,存在非常多的实际应用场景,如表单识别、车票信息抽取、身份证信息抽取等。然而,使用人力从这些文档图像中提取或者收集关键信息耗时费力,怎样自动化融合图像中的视觉、布局、文字等特征并完成关键信息抽取是一个价值与挑战并存的问题。
对于特定场景的文档图像,其中的关键信息位置、版式等较为固定,因此在研究早期有很多基于模板匹配的方法进行关键信息的抽取,考虑到其流程较为简单,该方法仍然被广泛应用在目前的很多场景中。但是这种基于模板匹配的方法在应用到不同的场景中时,需要耗费大量精力去调整与适配模板,迁移成本较高。
文档图像中的KIE一般包含2个子任务,示意图如下图所示。
* (1)SER: 语义实体识别 (Semantic Entity Recognition),对每一个检测到的文本进行分类,如将其分为姓名,身份证。如下图中的黑色框和红色框。
* (2)RE: 关系抽取 (Relation Extraction),对每一个检测到的文本进行分类,如将其分为问题 (key) 和答案 (value) 。然后对每一个问题找到对应的答案,相当于完成key-value的匹配过程。如下图中的红色框和黑色框分别代表问题和答案,黄色线代表问题和答案之间的对应关系。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/184588654-d87f54f3-13ab-42c4-afc0-da79bead3f14.png" width="800">
</div>
### 1.2 基于深度学习的主流方法
一般的KIE方法基于命名实体识别(Named Entity Recognition,NER)来展开研究,但是此类方法仅使用了文本信息而忽略了位置与视觉特征信息,因此精度受限。近几年大多学者开始融合多个模态的输入信息,进行特征融合,并对多模态信息进行处理,从而提升KIE的精度。主要方法有以下几种
* (1)基于Grid的方法:此类方法主要关注图像层面多模态信息的融合,文本大多大多为字符粒度,对文本与结构结构信息的嵌入方式较为简单,如Chargrid[1]等算法。
* (2)基于Token的方法:此类方法参考NLP中的BERT等方法,将位置、视觉等特征信息共同编码到多模态模型中,并且在大规模数据集上进行预训练,从而在下游任务中,仅需要少量的标注数据便可以获得很好的效果。如LayoutLM[2], LayoutLMv2[3], LayoutXLM[4], StrucText[5]等算法。
* (3)基于GCN的方法:此类方法尝试学习图像、文字之间的结构信息,从而可以解决开集信息抽取的问题(训练集中没有见过的模板),如GCN[6]、SDMGR[7]等算法。
* (4)基于End-to-end的方法:此类方法将现有的OCR文字识别以及KIE信息抽取2个任务放在一个统一的网络中进行共同学习,并在学习过程中相互加强。如Trie[8]等算法。
更多关于该系列算法的详细介绍,请参考“动手学OCR·十讲”课程的课节六部分:[文档分析理论与实践](https://aistudio.baidu.com/aistudio/education/group/info/25207)
## 2. 关键信息抽取任务流程
PaddleOCR中实现了LayoutXLM等算法(基于Token),同时,在PP-Structurev2中,对LayoutXLM多模态预训练模型的网络结构进行简化,去除了其中的Visual backbone部分,设计了视觉无关的VI-LayoutXLM模型,同时引入符合人类阅读顺序的排序逻辑以及UDML知识蒸馏策略,最终同时提升了关键信息抽取模型的精度与推理速度。
下面介绍怎样基于PaddleOCR完成关键信息抽取任务。
在非End-to-end的KIE方法中,完成关键信息抽取,至少需要**2个步骤**:首先使用OCR模型,完成文字位置与内容的提取,然后使用KIE模型,根据图像、文字位置以及文字内容,提取出其中的关键信息。
### 2.1 训练OCR模型
#### 2.1.1 文本检测
**(1)数据**
PaddleOCR中提供的模型大多数为通用模型,在进行文本检测的过程中,相邻文本行的检测一般是根据位置的远近进行区分,如上图,使用PP-OCRv3通用中英文检测模型进行文本检测时,容易将”民族“与“汉”这2个代表不同的字段检测到一起,从而增加后续KIE任务的难度。因此建议在做KIE任务的过程中,首先训练一个针对该文档数据集的检测模型。
在数据标注时,关键信息的标注需要隔开,比上图中的 “民族汉” 3个字相隔较近,此时需要将”民族“与”汉“标注为2个文本检测框,否则会增加后续KIE任务的难度。
对于下游任务,一般来说,`200~300`张的文本训练数据即可保证基本的训练效果,如果没有太多的先验知识,可以先标注 **`200~300`** 张图片,进行后续文本检测模型的训练。
**(2)模型**
在模型选择方面,推荐使用PP-OCRv3_det,关于更多关于检测模型的训练方法介绍,请参考:[OCR文本检测模型训练教程](../../doc/doc_ch/detection.md)[PP-OCRv3 文本检测模型训练教程](../../doc/doc_ch/PPOCRv3_det_train.md)
#### 2.1.2 文本识别
相对自然场景,文档图像中的文本内容识别难度一般相对较低(背景相对不太复杂),因此**优先建议**尝试PaddleOCR中提供的PP-OCRv3通用文本识别模型([PP-OCRv3模型库链接](../../doc/doc_ch/models_list.md))。
**(1)数据**
然而,在部分文档场景中也会存在一些挑战,如身份证场景中存在着罕见字,在发票等场景中的字体比较特殊,这些问题都会增加文本识别的难度,此时如果希望保证或者进一步提升模型的精度,建议基于特定文档场景的文本识别数据集,加载PP-OCRv3模型进行微调。
在模型微调的过程中,建议准备至少`5000`张垂类场景的文本识别图像,可以保证基本的模型微调效果。如果希望提升模型的精度与泛化能力,可以合成更多与该场景类似的文本识别数据,从公开数据集中收集通用真实文本识别数据,一并添加到该场景的文本识别训练任务过程中。在训练过程中,建议每个epoch的真实垂类数据、合成数据、通用数据比例在`1:1:1`左右,这可以通过设置不同数据源的采样比例进行控制。如有3个训练文本文件,分别包含1W、2W、5W条数据,那么可以在配置文件中设置数据如下:
```yml
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/
label_file_list:
- ./train_data/train_list_1W.txt
- ./train_data/train_list_2W.txt
- ./train_data/train_list_5W.txt
ratio_list: [1.0, 0.5, 0.2]
...
```
**(2)模型**
在模型选择方面,推荐使用通用中英文文本识别模型PP-OCRv3_rec,关于更多关于文本识别模型的训练方法介绍,请参考:[OCR文本识别模型训练教程](../../doc/doc_ch/recognition.md)[PP-OCRv3文本识别模型库与配置文件](../../doc/doc_ch/models_list.md)
### 2.2 训练KIE模型
对于识别得到的文字进行关键信息抽取,有2种主要的方法。
(1)直接使用SER,获取关键信息的类别:如身份证场景中,将“姓名“与”张三“分别标记为`name_key``name_value`。最终识别得到的类别为`name_value`对应的**文本字段**即为我们所需要的关键信息。
(2)联合SER与RE进行使用:这种方法中,首先使用SER,获取图像文字内容中所有的key与value,然后使用RE方法,对所有的key与value进行配对,找到映射关系,从而完成关键信息的抽取。
#### 2.2.1 SER
以身份证场景为例, 关键信息一般包含`姓名``性别``民族`等,我们直接将对应的字段标注为特定的类别即可,如下图所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/184526682-8b810397-5a93-4395-93da-37b8b8494c41.png" width="500">
</div>
**注意:**
- 标注过程中,对于无关于KIE关键信息的文本内容,均需要将其标注为`other`类别,相当于背景信息。如在身份证场景中,如果我们不关注性别信息,那么可以将“性别”与“男”这2个字段的类别均标注为`other`
- 标注过程中,需要以**文本行**为单位进行标注,无需标注单个字符的位置信息。
数据量方面,一般来说,对于比较固定的场景,**50张**左右的训练图片即可达到可以接受的效果,可以使用[PPOCRLabel](../../PPOCRLabel/README_ch.md)完成KIE的标注过程。
模型方面,推荐使用PP-Structurev2中提出的VI-LayoutXLM模型,它基于LayoutXLM模型进行改进,去除其中的视觉特征提取模块,在精度基本无损的情况下,进一步提升了模型推理速度。更多教程请参考:[VI-LayoutXLM算法介绍](../../doc/doc_ch/algorithm_kie_vi_layoutxlm.md)[KIE关键信息抽取使用教程](../../doc/doc_ch/kie.md)
#### 2.2.2 SER + RE
该过程主要包含SER与RE 2个过程。SER阶段主要用于识别出文档图像中的所有key与value,RE阶段主要用于对所有的key与value进行匹配。
以身份证场景为例, 关键信息一般包含`姓名``性别``民族`等关键信息,在SER阶段,我们需要识别所有的question (key) 与answer (value) 。标注如下所示。每个字段的类别信息(`label`字段)可以是question、answer或者other(与待抽取的关键信息无关的字段)
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/184526785-c3d2d310-cd57-4d31-b933-912716b29856.jpg" width="500">
</div>
在RE阶段,需要标注每个字段的的id与连接信息,如下图所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/184528728-626f77eb-fd9f-4709-a7dc-5411cc417dab.jpg" width="500">
</div>
每个文本行字段中,需要添加`id``linking`字段信息,`id`记录该文本行的唯一标识,同一张图片中的不同文本内容不能重复,`linking`是一个列表,记录了不同文本之间的连接信息。如字段“出生”的id为0,字段“1996年1月11日”的id为1,那么它们均有[[0, 1]]的`linking`标注,表示该id=0与id=1的字段构成key-value的关系(姓名、性别等字段类似,此处不再一一赘述)。
**注意:**
- 标注过程中,如果value是多个字符,那么linking中可以新增一个key-value对,如`[[0, 1], [0, 2]]`
数据量方面,一般来说,对于比较固定的场景,**50张**左右的训练图片即可达到可以接受的效果,可以使用PPOCRLabel完成KIE的标注过程。
模型方面,推荐使用PP-Structurev2中提出的VI-LayoutXLM模型,它基于LayoutXLM模型进行改进,去除其中的视觉特征提取模块,在精度基本无损的情况下,进一步提升了模型推理速度。更多教程请参考:[VI-LayoutXLM算法介绍](../../doc/doc_ch/algorithm_kie_vi_layoutxlm.md)[KIE关键信息抽取使用教程](../../doc/doc_ch/kie.md)
## 3. 参考文献
[1] Katti A R, Reisswig C, Guder C, et al. Chargrid: Towards understanding 2d documents[J]. arXiv preprint arXiv:1809.08799, 2018.
[2] Xu Y, Li M, Cui L, et al. Layoutlm: Pre-training of text and layout for document image understanding[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020: 1192-1200.
[3] Xu Y, Xu Y, Lv T, et al. LayoutLMv2: Multi-modal pre-training for visually-rich document understanding[J]. arXiv preprint arXiv:2012.14740, 2020.
[4]: Xu Y, Lv T, Cui L, et al. Layoutxlm: Multimodal pre-training for multilingual visually-rich document understanding[J]. arXiv preprint arXiv:2104.08836, 2021.
[5] Li Y, Qian Y, Yu Y, et al. StrucTexT: Structured Text Understanding with Multi-Modal Transformers[C]//Proceedings of the 29th ACM International Conference on Multimedia. 2021: 1912-1920.
[6] Liu X, Gao F, Zhang Q, et al. Graph convolution for multimodal information extraction from visually rich documents[J]. arXiv preprint arXiv:1903.11279, 2019.
[7] Sun H, Kuang Z, Yue X, et al. Spatial Dual-Modality Graph Reasoning for Key Information Extraction[J]. arXiv preprint arXiv:2103.14470, 2021.
[8] Zhang P, Xu Y, Cheng Z, et al. Trie: End-to-end text reading and information extraction for document understanding[C]//Proceedings of the 28th ACM International Conference on Multimedia. 2020: 1413-1422.
......@@ -40,14 +40,16 @@ logger = get_logger()
class SerPredictor(object):
def __init__(self, args):
self.ocr_engine = PaddleOCR(use_angle_cls=False, show_log=False)
self.ocr_engine = PaddleOCR(
use_angle_cls=False, show_log=False, use_gpu=args.use_gpu)
pre_process_list = [{
'VQATokenLabelEncode': {
'algorithm': args.vqa_algorithm,
'class_path': args.ser_dict_path,
'contains_re': False,
'ocr_engine': self.ocr_engine
'ocr_engine': self.ocr_engine,
'order_method': args.ocr_order_method,
}
}, {
'VQATokenPad': {
......
......@@ -78,20 +78,10 @@ def export_single_model(model,
shape=[None, 3, 64, 512], dtype="float32"),
]
model = to_static(model, input_spec=other_shape)
elif arch_config["algorithm"] == "RobustScanner":
max_text_length = arch_config["Head"]["max_text_length"]
elif arch_config["model_type"] == "sr":
other_shape = [
paddle.static.InputSpec(
shape=[None, 3, 48, 160], dtype="float32"),
[
paddle.static.InputSpec(
shape=[None, ],
dtype="float32"),
paddle.static.InputSpec(
shape=[None, max_text_length],
dtype="int64")
]
shape=[None, 3, 16, 64], dtype="float32")
]
model = to_static(model, input_spec=other_shape)
elif arch_config["algorithm"] == "ViTSTR":
......@@ -119,6 +109,22 @@ def export_single_model(model,
shape=[None, 3, 64, 256], dtype="float32"),
]
model = to_static(model, input_spec=other_shape)
elif arch_config["algorithm"] == "RobustScanner":
max_text_length = arch_config["Head"]["max_text_length"]
other_shape = [
paddle.static.InputSpec(
shape=[None, 3, 48, 160], dtype="float32"),
[
paddle.static.InputSpec(
shape=[None, ],
dtype="float32"),
paddle.static.InputSpec(
shape=[None, max_text_length],
dtype="int64")
]
]
model = to_static(model, input_spec=other_shape)
elif arch_config["algorithm"] in ["LayoutLM", "LayoutLMv2", "LayoutXLM"]:
input_spec = [
paddle.static.InputSpec(
......@@ -132,7 +138,7 @@ def export_single_model(model,
paddle.static.InputSpec(
shape=[None, 3, 224, 224], dtype="int64"), # image
]
if arch_config["algorithm"] == "LayoutLM":
if model.backbone.use_visual_backbone is False:
input_spec.pop(4)
model = to_static(model, input_spec=[input_spec])
else:
......@@ -211,6 +217,9 @@ def main():
else: # base rec model
config["Architecture"]["Head"]["out_channels"] = char_num
# for sr algorithm
if config["Architecture"]["model_type"] == "sr":
config['Architecture']["Transform"]['infer_mode'] = True
model = build_model(config["Architecture"])
load_model(config, model, model_type=config['Architecture']["model_type"])
model.eval()
......
......@@ -349,8 +349,7 @@ class TextRecognizer(object):
for beg_img_no in range(0, img_num, batch_num):
end_img_no = min(img_num, beg_img_no + batch_num)
norm_img_batch = []
# imgC, imgH, imgW = self.rec_image_shape
imgH, imgW = self.rec_image_shape[-2:]
imgC, imgH, imgW = self.rec_image_shape[:3]
max_wh_ratio = imgW / imgH
# max_wh_ratio = 0
for ino in range(beg_img_no, end_img_no):
......
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import sys
from PIL import Image
__dir__ = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, __dir__)
sys.path.insert(0, os.path.abspath(os.path.join(__dir__, '../..')))
os.environ["FLAGS_allocator_strategy"] = 'auto_growth'
import cv2
import numpy as np
import math
import time
import traceback
import paddle
import tools.infer.utility as utility
from ppocr.postprocess import build_post_process
from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, check_and_read_gif
logger = get_logger()
class TextSR(object):
def __init__(self, args):
self.sr_image_shape = [int(v) for v in args.sr_image_shape.split(",")]
self.sr_batch_num = args.sr_batch_num
self.predictor, self.input_tensor, self.output_tensors, self.config = \
utility.create_predictor(args, 'sr', logger)
self.benchmark = args.benchmark
if args.benchmark:
import auto_log
pid = os.getpid()
gpu_id = utility.get_infer_gpuid()
self.autolog = auto_log.AutoLogger(
model_name="sr",
model_precision=args.precision,
batch_size=args.sr_batch_num,
data_shape="dynamic",
save_path=None, #args.save_log_path,
inference_config=self.config,
pids=pid,
process_name=None,
gpu_ids=gpu_id if args.use_gpu else None,
time_keys=[
'preprocess_time', 'inference_time', 'postprocess_time'
],
warmup=0,
logger=logger)
def resize_norm_img(self, img):
imgC, imgH, imgW = self.sr_image_shape
img = img.resize((imgW // 2, imgH // 2), Image.BICUBIC)
img_numpy = np.array(img).astype("float32")
img_numpy = img_numpy.transpose((2, 0, 1)) / 255
return img_numpy
def __call__(self, img_list):
img_num = len(img_list)
batch_num = self.sr_batch_num
st = time.time()
st = time.time()
all_result = [] * img_num
if self.benchmark:
self.autolog.times.start()
for beg_img_no in range(0, img_num, batch_num):
end_img_no = min(img_num, beg_img_no + batch_num)
norm_img_batch = []
imgC, imgH, imgW = self.sr_image_shape
for ino in range(beg_img_no, end_img_no):
norm_img = self.resize_norm_img(img_list[ino])
norm_img = norm_img[np.newaxis, :]
norm_img_batch.append(norm_img)
norm_img_batch = np.concatenate(norm_img_batch)
norm_img_batch = norm_img_batch.copy()
if self.benchmark:
self.autolog.times.stamp()
self.input_tensor.copy_from_cpu(norm_img_batch)
self.predictor.run()
outputs = []
for output_tensor in self.output_tensors:
output = output_tensor.copy_to_cpu()
outputs.append(output)
if len(outputs) != 1:
preds = outputs
else:
preds = outputs[0]
all_result.append(outputs)
if self.benchmark:
self.autolog.times.end(stamp=True)
return all_result, time.time() - st
def main(args):
image_file_list = get_image_file_list(args.image_dir)
text_recognizer = TextSR(args)
valid_image_file_list = []
img_list = []
# warmup 2 times
if args.warmup:
img = np.random.uniform(0, 255, [16, 64, 3]).astype(np.uint8)
for i in range(2):
res = text_recognizer([img] * int(args.sr_batch_num))
for image_file in image_file_list:
img, flag = check_and_read_gif(image_file)
if not flag:
img = Image.open(image_file).convert("RGB")
if img is None:
logger.info("error in loading image:{}".format(image_file))
continue
valid_image_file_list.append(image_file)
img_list.append(img)
try:
preds, _ = text_recognizer(img_list)
for beg_no in range(len(preds)):
sr_img = preds[beg_no][1]
lr_img = preds[beg_no][0]
for i in (range(sr_img.shape[0])):
fm_sr = (sr_img[i] * 255).transpose(1, 2, 0).astype(np.uint8)
fm_lr = (lr_img[i] * 255).transpose(1, 2, 0).astype(np.uint8)
img_name_pure = os.path.split(valid_image_file_list[
beg_no * args.sr_batch_num + i])[-1]
cv2.imwrite("infer_result/sr_{}".format(img_name_pure),
fm_sr[:, :, ::-1])
logger.info("The visualized image saved in infer_result/sr_{}".
format(img_name_pure))
except Exception as E:
logger.info(traceback.format_exc())
logger.info(E)
exit()
if args.benchmark:
text_recognizer.autolog.report()
if __name__ == "__main__":
main(utility.parse_args())
......@@ -121,6 +121,11 @@ def init_args():
parser.add_argument("--use_pdserving", type=str2bool, default=False)
parser.add_argument("--warmup", type=str2bool, default=False)
# SR parmas
parser.add_argument("--sr_model_dir", type=str)
parser.add_argument("--sr_image_shape", type=str, default="3, 32, 128")
parser.add_argument("--sr_batch_num", type=int, default=1)
#
parser.add_argument(
"--draw_img_save_dir", type=str, default="./inference_results")
......@@ -156,6 +161,8 @@ def create_predictor(args, mode, logger):
model_dir = args.table_model_dir
elif mode == 'ser':
model_dir = args.ser_model_dir
elif mode == "sr":
model_dir = args.sr_model_dir
else:
model_dir = args.e2e_model_dir
......@@ -205,17 +212,23 @@ def create_predictor(args, mode, logger):
workspace_size=1 << 30,
precision_mode=precision,
max_batch_size=args.max_batch_size,
min_subgraph_size=args.min_subgraph_size, # skip the minmum trt subgraph
min_subgraph_size=args.
min_subgraph_size, # skip the minmum trt subgraph
use_calib_mode=False)
# collect shape
if args.shape_info_filename is not None:
if not os.path.exists(args.shape_info_filename):
config.collect_shape_range_info(args.shape_info_filename)
logger.info(f"collect dynamic shape info into : {args.shape_info_filename}")
logger.info(
f"collect dynamic shape info into : {args.shape_info_filename}"
)
else:
logger.info(f"dynamic shape info file( {args.shape_info_filename} ) already exists, not need to generate again.")
config.enable_tuned_tensorrt_dynamic_shape(args.shape_info_filename, True)
logger.info(
f"dynamic shape info file( {args.shape_info_filename} ) already exists, not need to generate again."
)
config.enable_tuned_tensorrt_dynamic_shape(
args.shape_info_filename, True)
use_dynamic_shape = True
if mode == "det":
......
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import os
import sys
import json
from PIL import Image
import cv2
__dir__ = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, __dir__)
sys.path.insert(0, os.path.abspath(os.path.join(__dir__, '..')))
os.environ["FLAGS_allocator_strategy"] = 'auto_growth'
import paddle
from ppocr.data import create_operators, transform
from ppocr.modeling.architectures import build_model
from ppocr.postprocess import build_post_process
from ppocr.utils.save_load import load_model
from ppocr.utils.utility import get_image_file_list
import tools.program as program
def main():
global_config = config['Global']
# build post process
post_process_class = build_post_process(config['PostProcess'],
global_config)
# sr transform
config['Architecture']["Transform"]['infer_mode'] = True
model = build_model(config['Architecture'])
load_model(config, model)
# create data ops
transforms = []
for op in config['Eval']['dataset']['transforms']:
op_name = list(op)[0]
if 'Label' in op_name:
continue
elif op_name in ['SRResize']:
op[op_name]['infer_mode'] = True
elif op_name == 'KeepKeys':
op[op_name]['keep_keys'] = ['imge_lr']
transforms.append(op)
global_config['infer_mode'] = True
ops = create_operators(transforms, global_config)
save_res_path = config['Global'].get('save_res_path', "./infer_result")
if not os.path.exists(os.path.dirname(save_res_path)):
os.makedirs(os.path.dirname(save_res_path))
model.eval()
for file in get_image_file_list(config['Global']['infer_img']):
logger.info("infer_img: {}".format(file))
img = Image.open(file).convert("RGB")
data = {'image_lr': img}
batch = transform(data, ops)
images = np.expand_dims(batch[0], axis=0)
images = paddle.to_tensor(images)
preds = model(images)
sr_img = preds["sr_img"][0]
lr_img = preds["lr_img"][0]
fm_sr = (sr_img.numpy() * 255).transpose(1, 2, 0).astype(np.uint8)
fm_lr = (lr_img.numpy() * 255).transpose(1, 2, 0).astype(np.uint8)
img_name_pure = os.path.split(file)[-1]
cv2.imwrite("infer_result/sr_{}".format(img_name_pure),
fm_sr[:, :, ::-1])
logger.info("The visualized image saved in infer_result/sr_{}".format(
img_name_pure))
logger.info("success!")
if __name__ == '__main__':
config, device, logger, vdl_writer = program.preprocess()
main()
......@@ -104,8 +104,6 @@ class SerPredictor(object):
batch = transform(data, self.ops)
batch = to_tensor(batch)
preds = self.model(batch)
if self.algorithm in ['LayoutLMv2', 'LayoutXLM']:
preds = preds[0]
post_result = self.post_process_class(
preds, segment_offset_ids=batch[6], ocr_infos=batch[7])
......
......@@ -25,6 +25,8 @@ import datetime
import paddle
import paddle.distributed as dist
from tqdm import tqdm
import cv2
import numpy as np
from argparse import ArgumentParser, RawDescriptionHelpFormatter
from ppocr.utils.stats import TrainingStats
......@@ -262,6 +264,7 @@ def train(config,
config, 'Train', device, logger, seed=epoch)
max_iter = len(train_dataloader) - 1 if platform.system(
) == "Windows" else len(train_dataloader)
for idx, batch in enumerate(train_dataloader):
profiler.add_profiler_step(profiler_options)
train_reader_cost += time.time() - reader_start
......@@ -289,7 +292,7 @@ def train(config,
else:
if model_type == 'table' or extra_input:
preds = model(images, data=batch[1:])
elif model_type in ["kie", 'vqa']:
elif model_type in ["kie", 'vqa', 'sr']:
preds = model(batch)
else:
preds = model(images)
......@@ -297,11 +300,12 @@ def train(config,
avg_loss = loss['loss']
avg_loss.backward()
optimizer.step()
optimizer.clear_grad()
if cal_metric_during_train and epoch % calc_epoch_interval == 0: # only rec and cls need
batch = [item.numpy() for item in batch]
if model_type in ['kie']:
if model_type in ['kie', 'sr']:
eval_class(preds, batch)
elif model_type in ['table']:
post_result = post_process_class(preds, batch)
......@@ -480,6 +484,7 @@ def eval(model,
leave=True)
max_iter = len(valid_dataloader) - 1 if platform.system(
) == "Windows" else len(valid_dataloader)
sum_images = 0
for idx, batch in enumerate(valid_dataloader):
if idx >= max_iter:
break
......@@ -493,6 +498,20 @@ def eval(model,
preds = model(images, data=batch[1:])
elif model_type in ["kie", 'vqa']:
preds = model(batch)
elif model_type in ['sr']:
preds = model(batch)
sr_img = preds["sr_img"]
lr_img = preds["lr_img"]
for i in (range(sr_img.shape[0])):
fm_sr = (sr_img[i].numpy() * 255).transpose(
1, 2, 0).astype(np.uint8)
fm_lr = (lr_img[i].numpy() * 255).transpose(
1, 2, 0).astype(np.uint8)
cv2.imwrite("output/images/{}_{}_sr.jpg".format(
sum_images, i), fm_sr)
cv2.imwrite("output/images/{}_{}_lr.jpg".format(
sum_images, i), fm_lr)
else:
preds = model(images)
else:
......@@ -500,6 +519,20 @@ def eval(model,
preds = model(images, data=batch[1:])
elif model_type in ["kie", 'vqa']:
preds = model(batch)
elif model_type in ['sr']:
preds = model(batch)
sr_img = preds["sr_img"]
lr_img = preds["lr_img"]
for i in (range(sr_img.shape[0])):
fm_sr = (sr_img[i].numpy() * 255).transpose(
1, 2, 0).astype(np.uint8)
fm_lr = (lr_img[i].numpy() * 255).transpose(
1, 2, 0).astype(np.uint8)
cv2.imwrite("output/images/{}_{}_sr.jpg".format(
sum_images, i), fm_sr)
cv2.imwrite("output/images/{}_{}_lr.jpg".format(
sum_images, i), fm_lr)
else:
preds = model(images)
......@@ -517,12 +550,15 @@ def eval(model,
elif model_type in ['table', 'vqa']:
post_result = post_process_class(preds, batch_numpy)
eval_class(post_result, batch_numpy)
elif model_type in ['sr']:
eval_class(preds, batch_numpy)
else:
post_result = post_process_class(preds, batch_numpy[1])
eval_class(post_result, batch_numpy)
pbar.update(1)
total_frame += len(images)
sum_images += 1
# Get final metric,eg. acc or hmean
metric = eval_class.get_metric()
......@@ -616,7 +652,8 @@ def preprocess(is_train=False):
'EAST', 'DB', 'SAST', 'Rosetta', 'CRNN', 'STARNet', 'RARE', 'SRN',
'CLS', 'PGNet', 'Distillation', 'NRTR', 'TableAttn', 'SAR', 'PSE',
'SEED', 'SDMGR', 'LayoutXLM', 'LayoutLM', 'LayoutLMv2', 'PREN', 'FCE',
'SVTR', 'ViTSTR', 'ABINet', 'DB++', 'TableMaster', 'SPIN', 'VisionLAN', 'RobustScanner'
'SVTR', 'ViTSTR', 'ABINet', 'DB++', 'TableMaster', 'SPIN', 'VisionLAN',
'Gestalt', 'RobustScanner'
]
if use_xpu:
......
......@@ -119,6 +119,7 @@ def main(config, device, logger, vdl_writer):
config['Loss']['ignore_index'] = char_num - 1
model = build_model(config['Architecture'])
model = apply_to_static(model, config, logger)
# build loss
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册