This commit is contained in:
Binary file not shown.
|
After Width: | Height: | Size: 234 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 143 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 125 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 347 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 404 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 247 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 565 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 736 KiB |
BIN
docs/version2.x/ppstructure/model_train/images/layout.jpg
Normal file
BIN
docs/version2.x/ppstructure/model_train/images/layout.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 535 KiB |
BIN
docs/version2.x/ppstructure/model_train/images/layout_res.jpg
Normal file
BIN
docs/version2.x/ppstructure/model_train/images/layout_res.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 451 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 385 KiB |
BIN
docs/version2.x/ppstructure/model_train/images/recovery_ch.jpg
Normal file
BIN
docs/version2.x/ppstructure/model_train/images/recovery_ch.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 315 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 369 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 772 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 467 KiB |
Binary file not shown.
|
After Width: | Height: | Size: 552 KiB |
BIN
docs/version2.x/ppstructure/model_train/images/zh_val_42_ser.jpg
Normal file
BIN
docs/version2.x/ppstructure/model_train/images/zh_val_42_ser.jpg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 692 KiB |
184
docs/version2.x/ppstructure/model_train/recovery_to_doc.en.md
Normal file
184
docs/version2.x/ppstructure/model_train/recovery_to_doc.en.md
Normal file
@@ -0,0 +1,184 @@
|
||||
---
|
||||
typora-copy-images-to: images
|
||||
comments: true
|
||||
---
|
||||
|
||||
# Layout Recovery
|
||||
|
||||
## 1. Introduction
|
||||
|
||||
The layout recovery module is used to restore the image or pdf to an
|
||||
editable Word file consistent with the original image layout.
|
||||
|
||||
Two layout recovery methods are provided, you can choose by PDF format:
|
||||
|
||||
- **Standard PDF parse(the input is standard PDF)**: Python based PDF to word library [pdf2docx](https://github.com/dothinking/pdf2docx) is optimized, the method extracts data from PDF with PyMuPDF, then parse layout with rule, finally, generate docx with python-docx.
|
||||
|
||||
- **Image format PDF parse(the input can be standard PDF or image format PDF)**: Layout recovery combines [layout analysis](./train_layout.en.md)、[table recognition](./train_table.en.md) to better recover images, tables, titles, etc. supports input files in PDF and document image formats in Chinese and English.
|
||||
|
||||
The input formats and application scenarios of the two methods are as follows:
|
||||
|
||||
| method | input formats | application scenarios/problem |
|
||||
| :-----: | :----------: | :----------------------------------------------------------: |
|
||||
| Standard PDF parse | pdf | Advantages: Better recovery for non-paper documents, each page remains on the same page after restoration<br>Disadvantages: English characters in some Chinese documents are garbled, some contents are still beyond the current page, the whole page content is restored to the table format, and the recovery effect of some pictures is not good |
|
||||
| Image format PDF parse( | pdf、picture | Advantages: More suitable for paper document content recovery, OCR recognition effect is more good<br>Disadvantages: Currently, the recovery is based on rules, the effect of content typesetting (spacing, fonts, etc.) need to be further improved, and the effect of layout recovery depends on layout analysis |
|
||||
|
||||
The following figure shows the effect of restoring the layout of documents by using PDF parse:
|
||||
|
||||
|
||||
|
||||
The following figures show the effect of restoring the layout of English and Chinese documents by using OCR technique:
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
## 2. Install
|
||||
|
||||
### 2.1 Install PaddlePaddle
|
||||
|
||||
```bash linenums="1"
|
||||
python3 -m pip install --upgrade pip
|
||||
|
||||
# If you have cuda9 or cuda10 installed on your machine, please run the following command to install
|
||||
python3 -m pip install "paddlepaddle-gpu" -i https://mirror.baidu.com/pypi/simple
|
||||
|
||||
# CPU installation
|
||||
python3 -m pip install "paddlepaddle" -i https://mirror.baidu.com/pypi/simple
|
||||
````
|
||||
|
||||
For more requirements, please refer to the instructions in [Installation Documentation](https://www.paddlepaddle.org.cn/en/install/quick?docurl=/documentation/docs/en/install/pip/macos-pip_en.html).
|
||||
|
||||
### 2.2 Install PaddleOCR
|
||||
|
||||
- **(1) Download source code**
|
||||
|
||||
```bash linenums="1"
|
||||
[Recommended] git clone https://github.com/PaddlePaddle/PaddleOCR
|
||||
|
||||
# If the pull cannot be successful due to network problems, you can also choose to use the hosting on the code cloud:
|
||||
git clone https://gitee.com/paddlepaddle/PaddleOCR
|
||||
|
||||
# Note: Code cloud hosting code may not be able to synchronize the update of this github project in real time, there is a delay of 3 to 5 days, please use the recommended method first.
|
||||
````
|
||||
|
||||
- **(2) Install recovery `requirements`**
|
||||
|
||||
The layout restoration is exported as docx files, so python-docx API need to be installed, and PyMuPDF api([requires Python >= 3.7](https://pypi.org/project/PyMuPDF/)) need to be installed to process the input files in pdf format.
|
||||
|
||||
Install all the libraries by running the following command:
|
||||
|
||||
```bash linenums="1"
|
||||
python3 -m pip install -r ppstructure/recovery/requirements.txt
|
||||
````
|
||||
|
||||
And if using pdf parse method, we need to install pdf2docx api.
|
||||
|
||||
```bash linenums="1"
|
||||
wget https://paddleocr.bj.bcebos.com/whl/pdf2docx-0.0.0-py3-none-any.whl
|
||||
pip3 install pdf2docx-0.0.0-py3-none-any.whl
|
||||
```
|
||||
|
||||
## 3. Quick Start using standard PDF parse
|
||||
|
||||
`use_pdf2docx_api` use PDF parse for layout recovery, The whl package is also provided for quick use, follow the above code, for more information please refer to [quickstart](../quick_start.en.md) for details.
|
||||
|
||||
```bash linenums="1"
|
||||
# install paddleocr
|
||||
pip3 install "paddleocr>=2.6"
|
||||
paddleocr --image_dir=ppstructure/docs/recovery/UnrealText.pdf --type=structure --recovery=true --use_pdf2docx_api=true
|
||||
```
|
||||
|
||||
Command line:
|
||||
|
||||
```bash linenums="1"
|
||||
python3 predict_system.py \
|
||||
--image_dir=ppstructure/docs/recovery/UnrealText.pdf \
|
||||
--recovery=True \
|
||||
--use_pdf2docx_api=True \
|
||||
--output=../output/
|
||||
```
|
||||
|
||||
## 4. Quick Start using image format PDF parse
|
||||
|
||||
Through layout analysis, we divided the image/PDF documents into regions, located the key regions, such as text, table, picture, etc., and recorded the location, category, and regional pixel value information of each region. Different regions are processed separately, where:
|
||||
|
||||
- OCR detection and recognition is performed in the text area, and the coordinates of the OCR detection box and the text content information are added on the basis of the previous information
|
||||
|
||||
- The table area identifies tables and records html and text information of tables
|
||||
- Save the image directly
|
||||
|
||||
We can restore the test picture through the layout information, OCR detection and recognition structure, table information, and saved pictures.
|
||||
|
||||
The whl package is also provided for quick use, follow the above code, for more information please refer to [quickstart](../quick_start.en.md) for details.
|
||||
|
||||
```bash linenums="1"
|
||||
paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --recovery=true --lang='en'
|
||||
```
|
||||
|
||||
### 4.1 Download models
|
||||
|
||||
If input is English document, download English models:
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleOCR/ppstructure
|
||||
|
||||
# download model
|
||||
mkdir inference && cd inference
|
||||
# Download the detection model of the ultra-lightweight English PP-OCRv3 model and unzip it
|
||||
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar && tar xf en_PP-OCRv3_det_infer.tar
|
||||
# Download the recognition model of the ultra-lightweight English PP-OCRv3 model and unzip it
|
||||
wget https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/en_PP-OCRv3_mobile_rec_infer.tar && tar xf en_PP-OCRv3_mobile_rec_infer.tar
|
||||
# Download the ultra-lightweight English table inch model and unzip it
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/paddle3.0b2/en_ppstructure_mobile_v2.0_SLANet_infer.tar
|
||||
tar xf en_ppstructure_mobile_v2.0_SLANet_infer.tar
|
||||
# Download the layout model of publaynet dataset and unzip it
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_infer.tar
|
||||
tar xf picodet_lcnet_x1_0_fgd_layout_infer.tar
|
||||
cd ..
|
||||
```
|
||||
|
||||
If input is Chinese document，download Chinese models:
|
||||
[Chinese and English ultra-lightweight PP-OCRv3 model](../../ppocr/model_list.md)
|
||||
|
||||
### 4.2 Layout recovery
|
||||
|
||||
```bash linenums="1"
|
||||
python3 predict_system.py \
|
||||
--image_dir=./docs/table/1.png \
|
||||
--det_model_dir=inference/en_PP-OCRv3_det_infer \
|
||||
--rec_model_dir=inference/en_PP-OCRv3_mobile_rec_infer \
|
||||
--rec_char_dict_path=../ppocr/utils/en_dict.txt \
|
||||
--table_model_dir=inference/en_ppstructure_mobile_v2.0_SLANet_infer \
|
||||
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \
|
||||
--layout_model_dir=inference/picodet_lcnet_x1_0_fgd_layout_infer \
|
||||
--layout_dict_path=../ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt \
|
||||
--vis_font_path=../doc/fonts/simfang.ttf \
|
||||
--recovery=True \
|
||||
--output=../output/
|
||||
```
|
||||
|
||||
After running, the docx of each picture will be saved in the directory specified by the output field
|
||||
|
||||
Field：
|
||||
|
||||
- image_dir：test file， can be picture, picture directory, pdf file, pdf file directory
|
||||
- det_model_dir：OCR detection model path
|
||||
- rec_model_dir：OCR recognition model path
|
||||
- rec_char_dict_path：OCR recognition dict path. If the Chinese model is used, change to "../ppocr/utils/ppocr_keys_v1.txt". And if you trained the model on your own dataset, change to the trained dictionary
|
||||
- table_model_dir：table recognition model path
|
||||
- table_char_dict_path：table recognition dict path. If the Chinese model is used, no need to change
|
||||
- layout_model_dir：layout analysis model path
|
||||
- layout_dict_path：layout analysis dict path. If the Chinese model is used, change to "../ppocr/utils/dict/layout_dict/layout_cdla_dict.txt"
|
||||
- recovery：whether to enable layout of recovery, default False
|
||||
- output：save the recovery result path
|
||||
|
||||
## 5. More
|
||||
|
||||
For training, evaluation and inference tutorial for text detection models, please refer to [text detection doc](../../ppocr/model_train/detection.en.md).
|
||||
|
||||
For training, evaluation and inference tutorial for text recognition models, please refer to [text recognition doc](../../ppocr/model_train/recognition.en.md).
|
||||
|
||||
For training, evaluation and inference tutorial for layout analysis models, please refer to [layout analysis doc](./train_layout.en.md)
|
||||
|
||||
For training, evaluation and inference tutorial for table recognition models, please refer to [table recognition doc](./train_table.en.md)
|
||||
191
docs/version2.x/ppstructure/model_train/recovery_to_doc.md
Normal file
191
docs/version2.x/ppstructure/model_train/recovery_to_doc.md
Normal file
@@ -0,0 +1,191 @@
|
||||
---
|
||||
typora-copy-images-to: images
|
||||
comments: true
|
||||
---
|
||||
|
||||
# 版面恢复
|
||||
|
||||
## 1. 简介
|
||||
|
||||
版面恢复就是将输入的图片、pdf内容仍然像原文档那样排列着，段落不变、顺序不变的输出到word文档中等。
|
||||
|
||||
提供了2种版面恢复方法，可根据输入PDF的格式进行选择：
|
||||
|
||||
- **标准PDF解析(输入须为标准PDF)**：基于Python的pdf转word库[pdf2docx](https://github.com/dothinking/pdf2docx)进行优化，该方法通过PyMuPDF获取页面元素，然后利用规则解析章节、段落、表格等布局及样式，最后通过python-docx将解析的内容元素重建到word文档中。
|
||||
- **图片格式PDF解析(输入可为标准PDF或图片格式PDF)**：结合[版面分析](./train_layout.md)、[表格识别](./train_table.md)技术，从而更好地恢复图片、表格、标题等内容，支持中、英文pdf文档、文档图片格式的输入文件。
|
||||
|
||||
2种方法输入格式、适用场景如下：
|
||||
|
||||
| 方法 | 支持输入文件 | 适用场景/存在问题 |
|
||||
| :-------------: | :----------: | :----------------------------------------------------------: |
|
||||
| 标准PDF解析 | pdf | 优点：非论文文档恢复效果更优、每一页内容恢复后仍在同一页<br>缺点：有些中文文档中的英文乱码、仍存在内容超出当前页面的情况、整页内容恢复为表格格式、部分图片恢复效果不佳 |
|
||||
| 图片格式PDF解析 | pdf、图片 | 优点：更适合论文文档正文内容的恢复、中英文文档OCR识别效果好<br>缺点：目前内容恢复基于规则，内容排版效果(间距、字体等)待进一步提升、版面恢复效果依赖于版面分析效果 |
|
||||
|
||||
下图展示了通过PDF解析版面恢复效果：
|
||||
|
||||
|
||||
|
||||
下图分别展示了通过OCR技术，英文文档和中文文档版面恢复的效果：
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
## 2. 安装
|
||||
|
||||
### 2.1 安装PaddlePaddle
|
||||
|
||||
```bash linenums="1"
|
||||
python3 -m pip install --upgrade pip
|
||||
|
||||
# 您的机器安装的是CUDA9或CUDA10，请运行以下命令安装
|
||||
python3 -m pip install "paddlepaddle-gpu" -i https://mirror.baidu.com/pypi/simple
|
||||
|
||||
# 您的机器是CPU，请运行以下命令安装
|
||||
python3 -m pip install "paddlepaddle" -i https://mirror.baidu.com/pypi/simple
|
||||
```
|
||||
|
||||
更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
|
||||
|
||||
### 2.2 安装PaddleOCR
|
||||
|
||||
- **（1）下载版面恢复源码**
|
||||
|
||||
```bash linenums="1"
|
||||
【推荐】git clone https://github.com/PaddlePaddle/PaddleOCR
|
||||
|
||||
# 如果因为网络问题无法pull成功，也可选择使用码云上的托管：
|
||||
git clone https://gitee.com/paddlepaddle/PaddleOCR
|
||||
|
||||
# 注：码云托管代码可能无法实时同步本github项目更新，存在3~5天延时，请优先使用推荐方式。
|
||||
```
|
||||
|
||||
- **（2）安装recovery的`requirements`**
|
||||
|
||||
版面恢复导出为docx文件，所以需要安装Python处理word文档的python-docx API，同时处理pdf格式的输入文件，需要安装PyMuPDF API([要求Python >= 3.7](https://pypi.org/project/PyMuPDF/))。
|
||||
|
||||
通过如下命令安装全部库：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 -m pip install -r ppstructure/recovery/requirements.txt
|
||||
```
|
||||
|
||||
使用pdf2docx库解析的方式恢复文档需要安装优化的pdf2docx。
|
||||
|
||||
```bash linenums="1"
|
||||
wget https://paddleocr.bj.bcebos.com/whl/pdf2docx-0.0.0-py3-none-any.whl
|
||||
pip3 install pdf2docx-0.0.0-py3-none-any.whl
|
||||
```
|
||||
|
||||
## 3.使用标准PDF解析进行版面恢复
|
||||
|
||||
`use_pdf2docx_api`表示使用PDF解析的方式进行版面恢复，通过whl包的形式方便快速使用，代码如下，更多信息详见 [quickstart](../quick_start.md)。
|
||||
|
||||
```bash linenums="1"
|
||||
# 安装 paddleocr，推荐使用2.6版本
|
||||
pip3 install "paddleocr>=2.6"
|
||||
paddleocr --image_dir=ppstructure/docs/recovery/UnrealText.pdf --type=structure --recovery=true --use_pdf2docx_api=true
|
||||
```
|
||||
|
||||
通过命令行的方式：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 predict_system.py \
|
||||
--image_dir=ppstructure/docs/recovery/UnrealText.pdf \
|
||||
--recovery=True \
|
||||
--use_pdf2docx_api=True \
|
||||
--output=../output/
|
||||
```
|
||||
|
||||
## 4.使用图片格式PDF解析进行版面恢复
|
||||
|
||||
我们通过版面分析对图片/pdf形式的文档进行区域划分，定位其中的关键区域，如文字、表格、图片等，记录每个区域的位置、类别、区域像素值信息。对不同的区域分别处理，其中：
|
||||
|
||||
- 文字区域直接进行OCR检测和识别，在之前信息基础上增加OCR检测框坐标和文本内容信息
|
||||
- 表格区域进行表格识别，记录表格html和文字信息
|
||||
- 图片直接保存
|
||||
|
||||
我们通过版面信息、OCR检测和识别结构、表格信息、保存的图片，对测试图片进行恢复即可。
|
||||
|
||||
提供如下代码实现版面恢复，也提供了whl包的形式方便快速使用，代码如下，更多信息详见 [quickstart](../quick_start.md)。
|
||||
|
||||
```bash linenums="1"
|
||||
# 安装 paddleocr，推荐使用2.6版本
|
||||
pip3 install "paddleocr>=2.6"
|
||||
# 中文测试图
|
||||
paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --recovery=true
|
||||
# 英文测试图
|
||||
paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --recovery=true --lang='en'
|
||||
# pdf测试文件
|
||||
paddleocr --image_dir=ppstructure/docs/recovery/UnrealText.pdf --type=structure --recovery=true --lang='en'
|
||||
```
|
||||
|
||||
### 4.1 下载模型
|
||||
|
||||
如果输入为英文文档类型，下载OCR检测和识别、版面分析、表格识别的英文模型
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleOCR/ppstructure
|
||||
|
||||
# 下载模型
|
||||
mkdir inference && cd inference
|
||||
# 下载英文超轻量PP-OCRv3检测模型并解压
|
||||
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar && tar xf en_PP-OCRv3_det_infer.tar
|
||||
# 下载英文超轻量PP-OCRv3识别模型并解压
|
||||
wget https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/en_PP-OCRv3_mobile_rec_infer.tar && tar xf en_PP-OCRv3_mobile_rec_infer.tar
|
||||
# 下载英文表格识别模型并解压
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/paddle3.0b2/en_ppstructure_mobile_v2.0_SLANet_infer.tar
|
||||
tar xf en_ppstructure_mobile_v2.0_SLANet_infer.tar
|
||||
# 下载英文版面分析模型
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_infer.tar
|
||||
tar xf picodet_lcnet_x1_0_fgd_layout_infer.tar
|
||||
cd ..
|
||||
```
|
||||
|
||||
如果输入为中文文档类型，在下述链接中下载中文模型即可：
|
||||
|
||||
[PP-OCRv3中英文超轻量文本检测和识别模型](../../ppocr/model_list.md)
|
||||
|
||||
### 4.2 版面恢复
|
||||
|
||||
使用下载的模型恢复给定文档的版面，以英文模型为例，执行如下命令：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 predict_system.py \
|
||||
--image_dir=./docs/table/1.png \
|
||||
--det_model_dir=inference/en_PP-OCRv3_det_infer \
|
||||
--rec_model_dir=inference/en_PP-OCRv3_mobile_rec_infer \
|
||||
--rec_char_dict_path=../ppocr/utils/en_dict.txt \
|
||||
--table_model_dir=inference/en_ppstructure_mobile_v2.0_SLANet_infer \
|
||||
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \
|
||||
--layout_model_dir=inference/picodet_lcnet_x1_0_fgd_layout_infer \
|
||||
--layout_dict_path=../ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt \
|
||||
--vis_font_path=../doc/fonts/simfang.ttf \
|
||||
--recovery=True \
|
||||
--output=../output/
|
||||
```
|
||||
|
||||
运行完成后，恢复版面的docx文档会保存到`output`字段指定的目录下
|
||||
|
||||
字段含义：
|
||||
|
||||
- image_dir：测试文件，可以是图片、图片目录、pdf文件、pdf文件目录
|
||||
- det_model_dir：OCR检测模型路径
|
||||
- rec_model_dir：OCR识别模型路径
|
||||
- rec_char_dict_path：OCR识别字典，如果更换为中文模型，需要更改为"../ppocr/utils/ppocr_keys_v1.txt"，如果您在自己的数据集上训练的模型，则更改为训练的字典的文件
|
||||
- table_model_dir：表格识别模型路径
|
||||
- table_char_dict_path：表格识别字典，如果更换为中文模型，不需要更换字典
|
||||
- layout_model_dir：版面分析模型路径
|
||||
- layout_dict_path：版面分析字典，如果更换为中文模型，需要更改为"../ppocr/utils/dict/layout_dict/layout_cdla_dict.txt"
|
||||
- recovery：是否进行版面恢复，默认False
|
||||
- output：版面恢复结果保存路径
|
||||
|
||||
## 5. 更多
|
||||
|
||||
关于OCR检测模型的训练评估与推理，请参考：[文本检测教程](../../ppocr/model_train/detection.md)
|
||||
|
||||
关于OCR识别模型的训练评估与推理，请参考：[文本识别教程](../../ppocr/model_train/recognition.md)
|
||||
|
||||
关于版面分析模型的训练评估与推理，请参考：[版面分析教程](./train_layout.md)
|
||||
|
||||
关于表格识别模型的训练评估与推理，请参考：[表格识别教程](./train_table.md)
|
||||
245
docs/version2.x/ppstructure/model_train/train_kie.en.md
Normal file
245
docs/version2.x/ppstructure/model_train/train_kie.en.md
Normal file
@@ -0,0 +1,245 @@
|
||||
---
|
||||
typora-copy-images-to: images
|
||||
comments: true
|
||||
---
|
||||
|
||||
# Key Information Extraction (KIE)
|
||||
|
||||
## 1. Introduction
|
||||
|
||||
Key information extraction (KIE) refers to extracting key information from text or images. As downstream task of OCR, the key information extraction task of document image has many practical application scenarios, such as form recognition, ticket information extraction, ID card information extraction, etc.
|
||||
|
||||
PP-Structure conducts research based on the LayoutXLM multi-modal, and proposes the VI-LayoutXLM, which gets rid of visual features when finetuning the downstream tasks. An textline sorting method is also utilized to fit in reading order. What's more, UDML knowledge distillation is used for higher accuracy. Finally, the accuracy and inference speed of VI-LayoutXLM surpass those of LayoutXLM.
|
||||
|
||||
The main features of the key information extraction module in PP-Structure are as follows.
|
||||
|
||||
- Integrate multi-modal methods such as [LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf), VI-LayoutXLM, and PP-OCR inference engine.
|
||||
- Supports Semantic Entity Recognition (SER) and Relation Extraction (RE) tasks based on multimodal methods. Based on the SER task, the text recognition and classification in the image can be completed; based on the RE task, the relationship extraction of the text content in the image can be completed, such as judging the problem pair (pair).
|
||||
- Supports custom training for SER tasks and RE tasks.
|
||||
- Supports end-to-end system prediction and evaluation of OCR+SER.
|
||||
- Supports end-to-end system prediction of OCR+SER+RE.
|
||||
- Support SER model export and inference using PaddleInference.
|
||||
|
||||
## 2. Performance
|
||||
|
||||
We evaluate the methods on the Chinese dataset of [XFUND](https://github.com/doc-analysis/XFUND), and the performance is as follows
|
||||
|
||||
|Model | Backbone | Task | Config file | Hmean | Inference time (ms) | Download link|
|
||||
| --- | --- | --- | --- | --- | --- | --- |
|
||||
|VI-LayoutXLM| VI-LayoutXLM-base | SER | [ser_vi_layoutxlm_xfund_zh_udml.yml](https://github.com/PaddlePaddle/PaddleOCR/tree/main/configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)|**93.19%**| 15.49|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)|
|
||||
|LayoutXLM| LayoutXLM-base | SER | [ser_layoutxlm_xfund_zh.yml](https://github.com/PaddlePaddle/PaddleOCR/tree/main/configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml)|90.38%| 19.49 | [trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)|
|
||||
|VI-LayoutXLM| VI-LayoutXLM-base | RE | [re_vi_layoutxlm_xfund_zh_udml.yml](https://github.com/PaddlePaddle/PaddleOCR/tree/main/configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)|**83.92%**| 15.49|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar)|
|
||||
|LayoutXLM| LayoutXLM-base | RE | [re_layoutxlm_xfund_zh.yml](https://github.com/PaddlePaddle/PaddleOCR/tree/main/configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml)|74.83%| 19.49|[trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar)|
|
||||
|
||||
- Note：Inference environment：V100 GPU + cuda10.2 + cudnn8.1.1 + TensorRT 7.2.3.4，tested using fp16.
|
||||
|
||||
For more KIE models in PaddleOCR, please refer to [KIE model zoo](../../algorithm/overview.en.md).
|
||||
|
||||
## 3. Visualization
|
||||
|
||||
There are two main solutions to the key information extraction task based on VI-LayoutXLM series model.
|
||||
|
||||
(1) Text detection + text recognition + semantic entity recognition (SER)
|
||||
|
||||
(2) Text detection + text recognition + semantic entity recognition (SER) + relationship extraction (RE)
|
||||
|
||||
The following images are demo results of the SER and RE models. For more detailed introduction to the above solutions, please refer to [KIE Guide](../blog/how_to_do_kie.en.md).
|
||||
|
||||
### 3.1 SER
|
||||
|
||||
Demo results for SER task are as follows.
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
**Note:** test pictures are from [xfund dataset](https://github.com/doc-analysis/XFUND), [invoice dataset](https://aistudio.baidu.com/aistudio/datasetdetail/165561) and a composite ID card dataset.
|
||||
|
||||
Boxes of different colors in the image represent different categories.
|
||||
|
||||
The invoice and application form images have three categories: `request`, `answer` and `header`. The `question` and `answer` can be used to extract the relationship.
|
||||
|
||||
For the ID card image, the model can directly identify the key information such as `name`, `gender`, `nationality`, so that the subsequent relationship extraction process is not required, and the key information extraction task can be completed using only one model.
|
||||
|
||||
### 3.2 RE
|
||||
|
||||
Demo results for RE task are as follows.
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
Red boxes are questions, blue boxes are answers. The green lines means the two connected objects are a pair.
|
||||
|
||||
## 4. Usage
|
||||
|
||||
### 4.1 Prepare for the environment
|
||||
|
||||
Use the following command to install KIE dependencies.
|
||||
|
||||
```bash linenums="1"
|
||||
git clone https://github.com/PaddlePaddle/PaddleOCR.git
|
||||
cd PaddleOCR
|
||||
pip install -r requirements.txt
|
||||
pip install -r ppstructure/kie/requirements.txt
|
||||
# 安装PaddleOCR引擎用于预测
|
||||
pip install "paddleocr<3.0"
|
||||
```
|
||||
|
||||
NOTE: For KIE tasks, it is necessary to downgrade the Paddle framework version (Paddle<2.6) and the PaddleNLP version (PaddleNLP<2.6).
|
||||
|
||||
The visualized results of SER are saved in the `./output` folder by default. Examples of results are as follows.
|
||||
|
||||
|
||||
|
||||
### 4.2 Quick start
|
||||
|
||||
Here we use XFUND dataset to quickly experience the SER model and RE model.
|
||||
|
||||
#### 4.2.1 Prepare for the dataset
|
||||
|
||||
```bash linenums="1"
|
||||
mkdir train_data
|
||||
cd train_data
|
||||
# download and uncompress the dataset
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar && tar -xf XFUND.tar
|
||||
cd ..
|
||||
```
|
||||
|
||||
#### 4.2.2 Predict images using the trained model
|
||||
|
||||
Use the following command to download the models.
|
||||
|
||||
```bash linenums="1"
|
||||
mkdir pretrained_model
|
||||
cd pretrained_model
|
||||
# download and uncompress the SER trained model
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar && tar -xf ser_vi_layoutxlm_xfund_pretrained.tar
|
||||
|
||||
# download and uncompress the RE trained model
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar && tar -xf re_vi_layoutxlm_xfund_pretrained.tar
|
||||
```
|
||||
|
||||
If you want to use OCR engine to obtain end-to-end prediction results, you can use the following command to predict.
|
||||
|
||||
```bash linenums="1"
|
||||
# just predict using SER trained model
|
||||
python3 tools/infer_kie_token_ser.py \
|
||||
-c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
|
||||
-o Architecture.Backbone.checkpoints=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy \
|
||||
Global.infer_img=./ppstructure/docs/kie/input/zh_val_42.jpg
|
||||
|
||||
# predict using SER and RE trained model at the same time
|
||||
python3 ./tools/infer_kie_token_ser_re.py \
|
||||
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
|
||||
-o Architecture.Backbone.checkpoints=./pretrained_model/re_vi_layoutxlm_xfund_pretrained/best_accuracy \
|
||||
Global.infer_img=./train_data/XFUND/zh_val/image/zh_val_42.jpg \
|
||||
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
|
||||
-o_ser Architecture.Backbone.checkpoints=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
|
||||
```
|
||||
|
||||
The visual result images and the predicted text file will be saved in the `Global.save_res_path` directory.
|
||||
|
||||
If you want to use a custom ocr model, you can set it through the following fields
|
||||
|
||||
- `Global.kie_det_model_dir`: the detection inference model path
|
||||
- `Global.kie_rec_model_dir`: the recognition inference model path
|
||||
|
||||
If you want to load the text detection and recognition results collected before, you can use the following command to predict.
|
||||
|
||||
```bash linenums="1"
|
||||
# just predict using SER trained model
|
||||
python3 tools/infer_kie_token_ser.py \
|
||||
-c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
|
||||
-o Architecture.Backbone.checkpoints=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy \
|
||||
Global.infer_img=./train_data/XFUND/zh_val/val.json \
|
||||
Global.infer_mode=False
|
||||
|
||||
# predict using SER and RE trained model at the same time
|
||||
python3 ./tools/infer_kie_token_ser_re.py \
|
||||
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
|
||||
-o Architecture.Backbone.checkpoints=./pretrained_model/re_vi_layoutxlm_xfund_pretrained/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=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
|
||||
```
|
||||
|
||||
#### 4.2.3 Inference using PaddleInference
|
||||
|
||||
Firstly, download the inference SER inference model.
|
||||
|
||||
```bash linenums="1"
|
||||
mkdir inference
|
||||
cd inference
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar && tar -xf ser_vi_layoutxlm_xfund_infer.tar
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_infer.tar && tar -xf re_vi_layoutxlm_xfund_infer.tar
|
||||
cd ..
|
||||
```
|
||||
|
||||
- SER
|
||||
|
||||
Use the following command for inference.
|
||||
|
||||
```bash linenums="1"
|
||||
cd ppstructure
|
||||
python3 kie/predict_kie_token_ser.py \
|
||||
--kie_algorithm=LayoutXLM \
|
||||
--ser_model_dir=../inference/ser_vi_layoutxlm_xfund_infer \
|
||||
--image_dir=./docs/kie/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"
|
||||
```
|
||||
|
||||
The visual results and text file will be saved in directory `output`.
|
||||
|
||||
- RE
|
||||
|
||||
Use the following command for inference.
|
||||
|
||||
```bash linenums="1"
|
||||
cd ppstructure
|
||||
python3 kie/predict_kie_token_ser_re.py \
|
||||
--kie_algorithm=LayoutXLM \
|
||||
--re_model_dir=../inference/re_vi_layoutxlm_xfund_infer \
|
||||
--ser_model_dir=../inference/ser_vi_layoutxlm_xfund_infer \
|
||||
--use_visual_backbone=False \
|
||||
--image_dir=./docs/kie/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"
|
||||
```
|
||||
|
||||
The visual results and text file will be saved in directory `output`.
|
||||
|
||||
If you want to use a custom ocr model, you can set it through the following fields
|
||||
|
||||
- `--det_model_dir`: the detection inference model path
|
||||
- `--rec_model_dir`: the recognition inference model path
|
||||
|
||||
### 4.3 More
|
||||
|
||||
For training, evaluation and inference tutorial for KIE models, please refer to [KIE doc](../model_train/train_kie.en.md).
|
||||
|
||||
For training, evaluation and inference tutorial for text detection models, please refer to [text detection doc](../../ppocr/model_train/detection.en.md).
|
||||
|
||||
For training, evaluation and inference tutorial for text recognition models, please refer to [text recognition doc](../../ppocr/model_train/recognition.en.md).
|
||||
|
||||
To complete the key information extraction task in your own scenario from data preparation to model selection, please refer to: [Guide to End-to-end KIE](../blog/how_to_do_kie.en.md)。
|
||||
|
||||
## 5. Reference
|
||||
|
||||
- LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, <https://arxiv.org/pdf/2104.08836.pdf>
|
||||
- microsoft/unilm/layoutxlm, <https://github.com/microsoft/unilm/tree/master/layoutxlm>
|
||||
- XFUND dataset, <https://github.com/doc-analysis/XFUND>
|
||||
|
||||
## 6. License
|
||||
|
||||
The content of this project itself is licensed under the [Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/)
|
||||
240
docs/version2.x/ppstructure/model_train/train_kie.md
Normal file
240
docs/version2.x/ppstructure/model_train/train_kie.md
Normal file
@@ -0,0 +1,240 @@
|
||||
---
|
||||
typora-copy-images-to: images
|
||||
comments: true
|
||||
---
|
||||
|
||||
# 关键信息抽取
|
||||
|
||||
## 1. 简介
|
||||
|
||||
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。
|
||||
|
||||
PP-Structure 基于 LayoutXLM 文档多模态系列方法进行研究与优化，设计了视觉特征无关的多模态模型结构VI-LayoutXLM，同时引入符合阅读顺序的文本行排序方法以及UDML联合互学习蒸馏方法，最终在精度与速度均超越LayoutXLM。
|
||||
|
||||
PP-Structure中关键信息抽取模块的主要特性如下：
|
||||
|
||||
- 集成[LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf)、VI-LayoutXLM等多模态模型以及PP-OCR预测引擎。
|
||||
- 支持基于多模态方法的语义实体识别 (Semantic Entity Recognition, SER) 以及关系抽取 (Relation Extraction, RE) 任务。基于 SER 任务，可以完成对图像中的文本识别与分类；基于 RE 任务，可以完成对图象中的文本内容的关系提取，如判断问题对(pair)。
|
||||
- 支持SER任务和RE任务的自定义训练。
|
||||
- 支持OCR+SER的端到端系统预测与评估。
|
||||
- 支持OCR+SER+RE的端到端系统预测。
|
||||
- 支持SER模型的动转静导出与基于PaddleInfernece的模型推理。
|
||||
|
||||
## 2. 精度与性能
|
||||
|
||||
我们在 [XFUND](https://github.com/doc-analysis/XFUND) 的中文数据集上对算法进行了评估，SER与RE上的任务性能如下
|
||||
|
||||
| 模型 | 骨干网络 | 任务 | 配置文件 | hmean | 预测耗时(ms) | 下载链接 |
|
||||
| ------------ | ----------------- | ---- | ------- | ---------- | ------ | ------ |
|
||||
| VI-LayoutXLM | VI-LayoutXLM-base | SER | [ser_vi_layoutxlm_xfund_zh_udml.yml](https://github.com/PaddlePaddle/PaddleOCR/tree/main/configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml) | **93.19%** | 15.49 | [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](https://github.com/PaddlePaddle/PaddleOCR/tree/main/configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml) | 90.38% | 19.49 | [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|
||||
| VI-LayoutXLM | VI-LayoutXLM-base | RE | [re_vi_layoutxlm_xfund_zh_udml.yml](https://github.com/PaddlePaddle/PaddleOCR/tree/main/configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml) | **83.92%** | 15.49 | [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](https://github.com/PaddlePaddle/PaddleOCR/tree/main/configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml) | 74.83% | 19.49 | [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
|
||||
|
||||
- 注：预测耗时测试条件：V100 GPU + cuda10.2 + cudnn8.1.1 + TensorRT 7.2.3.4，使用FP16进行测试。
|
||||
|
||||
更多关于PaddleOCR中关键信息抽取模型的介绍，请参考[关键信息抽取模型库](../../algorithm/overview.md)。
|
||||
|
||||
## 3. 效果演示
|
||||
|
||||
基于多模态模型的关键信息抽取任务有2种主要的解决方案。
|
||||
|
||||
（1）文本检测 + 文本识别 + 语义实体识别(SER)
|
||||
（2）文本检测 + 文本识别 + 语义实体识别(SER) + 关系抽取(RE)
|
||||
|
||||
下面给出SER与RE任务的示例效果，关于上述解决方案的详细介绍，请参考[关键信息抽取全流程指南](../blog/how_to_do_kie.md)。
|
||||
|
||||
### 3.1 SER
|
||||
|
||||
对于SER任务，效果如下所示。
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
**注意：** 测试图片来源于[XFUND数据集](https://github.com/doc-analysis/XFUND)、[发票数据集](https://aistudio.baidu.com/aistudio/datasetdetail/165561)以及合成的身份证数据集。
|
||||
|
||||
图中不同颜色的框表示不同的类别。
|
||||
|
||||
图中的发票以及申请表图像，有`QUESTION`, `ANSWER`, `HEADER` 3种类别，识别的`QUESTION`, `ANSWER`可以用于后续的问题与答案的关系抽取。
|
||||
|
||||
图中的身份证图像，则直接识别出其中的`姓名`、`性别`、`民族`等关键信息，这样就无需后续的关系抽取过程，一个模型即可完成关键信息抽取。
|
||||
|
||||
### 3.2 RE
|
||||
|
||||
对于RE任务，效果如下所示。
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
红色框是问题，蓝色框是答案。绿色线条表示连接的两端为一个key-value的pair。
|
||||
|
||||
## 4. 使用
|
||||
|
||||
### 4.1 准备环境
|
||||
|
||||
使用下面的命令安装运行SER与RE关键信息抽取的依赖。
|
||||
|
||||
```bash linenums="1"
|
||||
git clone https://github.com/PaddlePaddle/PaddleOCR.git
|
||||
cd PaddleOCR
|
||||
pip install -r requirements.txt
|
||||
pip install -r ppstructure/kie/requirements.txt
|
||||
# 安装PaddleOCR引擎用于预测
|
||||
pip install "paddleocr<3.0"
|
||||
```
|
||||
|
||||
NOTE: 对于KIE任务需要降低Paddle框架版本（Paddle<2.6），和PaddleNLP版本（PaddleNLP<2.6）。
|
||||
|
||||
### 4.2 快速开始
|
||||
|
||||
下面XFUND数据集，快速体验SER模型与RE模型。
|
||||
|
||||
#### 4.2.1 准备数据
|
||||
|
||||
```bash linenums="1"
|
||||
mkdir train_data
|
||||
cd train_data
|
||||
# 下载与解压数据
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar && tar -xf XFUND.tar
|
||||
cd ..
|
||||
```
|
||||
|
||||
#### 4.2.2 基于动态图的预测
|
||||
|
||||
首先下载模型。
|
||||
|
||||
```bash linenums="1"
|
||||
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
|
||||
```
|
||||
|
||||
如果希望使用OCR引擎，获取端到端的预测结果，可以使用下面的命令进行预测。
|
||||
|
||||
```bash linenums="1"
|
||||
# 仅预测SER模型
|
||||
python3 tools/infer_kie_token_ser.py \
|
||||
-c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
|
||||
-o Architecture.Backbone.checkpoints=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy \
|
||||
Global.infer_img=./ppstructure/docs/kie/input/zh_val_42.jpg
|
||||
|
||||
# SER + RE模型串联
|
||||
python3 ./tools/infer_kie_token_ser_re.py \
|
||||
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
|
||||
-o Architecture.Backbone.checkpoints=./pretrained_model/re_vi_layoutxlm_xfund_pretrained/best_accuracy \
|
||||
Global.infer_img=./train_data/XFUND/zh_val/image/zh_val_42.jpg \
|
||||
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
|
||||
-o_ser Architecture.Backbone.checkpoints=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
|
||||
```
|
||||
|
||||
`Global.save_res_path`目录中会保存可视化的结果图像以及预测的文本文件。
|
||||
|
||||
如果想使用自定义OCR模型，可通过如下字段进行设置
|
||||
|
||||
- `Global.kie_det_model_dir`: 设置检测inference模型地址
|
||||
- `Global.kie_rec_model_dir`: 设置识别inference模型地址
|
||||
|
||||
如果希望加载标注好的文本检测与识别结果，仅预测可以使用下面的命令进行预测。
|
||||
|
||||
```bash linenums="1"
|
||||
# 仅预测SER模型
|
||||
python3 tools/infer_kie_token_ser.py \
|
||||
-c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
|
||||
-o Architecture.Backbone.checkpoints=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy \
|
||||
Global.infer_img=./train_data/XFUND/zh_val/val.json \
|
||||
Global.infer_mode=False
|
||||
|
||||
# SER + RE模型串联
|
||||
python3 ./tools/infer_kie_token_ser_re.py \
|
||||
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
|
||||
-o Architecture.Backbone.checkpoints=./pretrained_model/re_vi_layoutxlm_xfund_pretrained/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=./pretrained_model/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
|
||||
```
|
||||
|
||||
#### 4.2.3 基于PaddleInference的预测
|
||||
|
||||
首先下载SER和RE的推理模型。
|
||||
|
||||
```bash linenums="1"
|
||||
mkdir inference
|
||||
cd inference
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar && tar -xf ser_vi_layoutxlm_xfund_infer.tar
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_infer.tar && tar -xf re_vi_layoutxlm_xfund_infer.tar
|
||||
cd ..
|
||||
```
|
||||
|
||||
- SER
|
||||
|
||||
执行下面的命令进行预测。
|
||||
|
||||
```bash linenums="1"
|
||||
cd ppstructure
|
||||
python3 kie/predict_kie_token_ser.py \
|
||||
--kie_algorithm=LayoutXLM \
|
||||
--ser_model_dir=../inference/ser_vi_layoutxlm_xfund_infer \
|
||||
--image_dir=./docs/kie/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"
|
||||
```
|
||||
|
||||
可视化结果保存在`output`目录下。
|
||||
|
||||
- RE
|
||||
|
||||
执行下面的命令进行预测。
|
||||
|
||||
```bash linenums="1"
|
||||
cd ppstructure
|
||||
python3 kie/predict_kie_token_ser_re.py \
|
||||
--kie_algorithm=LayoutXLM \
|
||||
--re_model_dir=../inference/re_vi_layoutxlm_xfund_infer \
|
||||
--ser_model_dir=../inference/ser_vi_layoutxlm_xfund_infer \
|
||||
--use_visual_backbone=False \
|
||||
--image_dir=./docs/kie/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"
|
||||
```
|
||||
|
||||
可视化结果保存在`output`目录下。
|
||||
|
||||
如果想使用自定义OCR模型，可通过如下字段进行设置
|
||||
|
||||
- `--det_model_dir`: 设置检测inference模型地址
|
||||
- `--rec_model_dir`: 设置识别inference模型地址
|
||||
|
||||
### 4.3 更多
|
||||
|
||||
关于KIE模型的训练评估与推理，请参考：[关键信息抽取教程](./train_kie.md)。
|
||||
|
||||
关于文本检测模型的训练评估与推理，请参考：[文本检测教程](../../ppocr/model_train/detection.md)。
|
||||
|
||||
关于文本识别模型的训练评估与推理，请参考：[文本识别教程](../../ppocr/model_train/recognition.md)。
|
||||
|
||||
关于怎样在自己的场景中完成关键信息抽取任务，请参考：[关键信息抽取全流程指南](../blog/how_to_do_kie.md)。
|
||||
|
||||
## 5. 参考链接
|
||||
|
||||
- LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, <https://arxiv.org/pdf/2104.08836.pdf>
|
||||
- microsoft/unilm/layoutxlm, <https://github.com/microsoft/unilm/tree/master/layoutxlm>
|
||||
- XFUND dataset, <https://github.com/doc-analysis/XFUND>
|
||||
|
||||
## 6. License
|
||||
|
||||
The content of this project itself is licensed under the [Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/)
|
||||
444
docs/version2.x/ppstructure/model_train/train_layout.en.md
Normal file
444
docs/version2.x/ppstructure/model_train/train_layout.en.md
Normal file
@@ -0,0 +1,444 @@
|
||||
---
|
||||
comments: true
|
||||
---
|
||||
|
||||
# Layout analysis
|
||||
|
||||
## 1. Introduction
|
||||
|
||||
Layout analysis refers to the regional division of documents in the form of pictures and the positioning of key areas, such as text, title, table, picture, etc. The layout analysis algorithm is based on the lightweight model PP-picodet of [PaddleDetection]( https://github.com/PaddlePaddle/PaddleDetection ), including English layout analysis, Chinese layout analysis and table layout analysis models. English layout analysis models can detect document layout elements such as text, title, table, figure, list. Chinese layout analysis models can detect document layout elements such as text, figure, figure caption, table, table caption, header, footer, reference, and equation. Table layout analysis models can detect table regions.
|
||||
|
||||
|
||||
|
||||
## 2. Quick start
|
||||
|
||||
PP-Structure currently provides layout analysis models in Chinese, English and table documents. For the model link, see [models_list](../models_list.en.md). The whl package is also provided for quick use, see [quickstart](../quick_start.en.md) for details.
|
||||
|
||||
## 3. Install
|
||||
|
||||
### 3.1. Install PaddlePaddle
|
||||
|
||||
- **（1) Install PaddlePaddle**
|
||||
|
||||
```bash linenums="1"
|
||||
python3 -m pip install --upgrade pip
|
||||
|
||||
# GPU Install
|
||||
python3 -m pip install "paddlepaddle-gpu>=2.3" -i https://mirror.baidu.com/pypi/simple
|
||||
|
||||
# CPU Install
|
||||
python3 -m pip install "paddlepaddle>=2.3" -i https://mirror.baidu.com/pypi/simple
|
||||
```
|
||||
|
||||
For more requirements, please refer to the instructions in the [Install file](https://www.paddlepaddle.org.cn/install/quick)。
|
||||
|
||||
### 3.2. Install PaddleDetection
|
||||
|
||||
- **（1）Download PaddleDetection Source code**
|
||||
|
||||
```bash linenums="1"
|
||||
git clone https://github.com/PaddlePaddle/PaddleDetection.git
|
||||
```
|
||||
|
||||
- **（2）Install third-party libraries**
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleDetection
|
||||
python3 -m pip install -r requirements.txt
|
||||
```
|
||||
|
||||
## 4. Data preparation
|
||||
|
||||
If you want to experience the prediction process directly, you can skip data preparation and download the pre-training model.
|
||||
|
||||
### 4.1. English data set
|
||||
|
||||
Download document analysis data set [PubLayNet](https://developer.ibm.com/exchanges/data/all/publaynet/)（Dataset 96G），contains 5 classes：`{0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"}`
|
||||
|
||||
```
|
||||
# Download data
|
||||
wget https://dax-cdn.cdn.appdomain.cloud/dax-publaynet/1.0.0/publaynet.tar.gz
|
||||
# Decompress data
|
||||
tar -xvf publaynet.tar.gz
|
||||
```
|
||||
|
||||
Uncompressed **directory structure：**
|
||||
|
||||
```
|
||||
|-publaynet
|
||||
|- test
|
||||
|- PMC1277013_00004.jpg
|
||||
|- PMC1291385_00002.jpg
|
||||
| ...
|
||||
|- train.json
|
||||
|- train
|
||||
|- PMC1291385_00002.jpg
|
||||
|- PMC1277013_00004.jpg
|
||||
| ...
|
||||
|- val.json
|
||||
|- val
|
||||
|- PMC538274_00004.jpg
|
||||
|- PMC539300_00004.jpg
|
||||
| ...
|
||||
```
|
||||
|
||||
**data distribution：**
|
||||
|
||||
| File or Folder | Description | num |
|
||||
| :------------- | :------------- | ------- |
|
||||
| `train/` | Training set pictures | 335,703 |
|
||||
| `val/` | Verification set pictures | 11,245 |
|
||||
| `test/` | Test set pictures | 11,405 |
|
||||
| `train.json` | Training set annotation files | - |
|
||||
| `val.json` | Validation set dimension files | - |
|
||||
|
||||
**Data Annotation**
|
||||
|
||||
The JSON file contains the annotations of all images, and the data is stored in a dictionary nested manner.Contains the following keys：
|
||||
|
||||
- info，represents the dimension file info。
|
||||
|
||||
- licenses，represents the dimension file licenses。
|
||||
|
||||
- images，represents the list of image information in the annotation file，each element is the information of an image。The information of one of the images is as follows:
|
||||
|
||||
```
|
||||
{
|
||||
'file_name': 'PMC4055390_00006.jpg', # file_name
|
||||
'height': 601, # image height
|
||||
'width': 792, # image width
|
||||
'id': 341427 # image id
|
||||
}
|
||||
```
|
||||
|
||||
- annotations， represents the list of annotation information of the target object in the annotation file，each element is the annotation information of a target object。The following is the annotation information of one of the target objects:
|
||||
|
||||
```
|
||||
{
|
||||
|
||||
'segmentation': # Segmentation annotation of objects
|
||||
'area': 60518.099043117836, # Area of object
|
||||
'iscrowd': 0, # iscrowd
|
||||
'image_id': 341427, # image id
|
||||
'bbox': [50.58, 490.86, 240.15, 252.16], # bbox [x1,y1,w,h]
|
||||
'category_id': 1, # category_id
|
||||
'id': 3322348 # image id
|
||||
}
|
||||
```
|
||||
|
||||
### 4.2. More datasets
|
||||
|
||||
We provide CDLA(Chinese layout analysis), TableBank(Table layout analysis)etc. data set download links，process to the JSON format of the above annotation file，that is, the training can be conducted in the same way。
|
||||
|
||||
| dataset | 简介 |
|
||||
| ------------------------------------------------------------ | ------------------------------------------------------------ |
|
||||
| [cTDaR2019_cTDaR](https://cndplab-founder.github.io/cTDaR2019/) | For form detection (TRACKA) and form identification (TRACKB).Image types include historical data sets (beginning with cTDaR_t0, such as CTDAR_T00872.jpg) and modern data sets (beginning with cTDaR_t1, CTDAR_T10482.jpg). |
|
||||
| [IIIT-AR-13K](http://cvit.iiit.ac.in/usodi/iiitar13k.php) | Data sets constructed by manually annotating figures or pages from publicly available annual reports, containing 5 categories:table, figure, natural image, logo, and signature. |
|
||||
| [TableBank](https://github.com/doc-analysis/TableBank) | For table detection and recognition of large datasets, including Word and Latex document formats |
|
||||
| [CDLA](https://github.com/buptlihang/CDLA) | Chinese document layout analysis data set, for Chinese literature (paper) scenarios, including 10 categories:Text, Title, Figure, Figure caption, Table, Table caption, Header, Footer, Reference, Equation |
|
||||
| [DocBank](https://github.com/doc-analysis/DocBank) | Large-scale dataset (500K document pages) constructed using weakly supervised methods for document layout analysis, containing 12 categories:Author, Caption, Date, Equation, Figure, Footer, List, Paragraph, Reference, Section, Table, Title |
|
||||
|
||||
## 5. Start training
|
||||
|
||||
Training scripts, evaluation scripts, and prediction scripts are provided, and the PubLayNet pre-training model is used as an example in this section.
|
||||
|
||||
If you do not want training and directly experience the following process of model evaluation, prediction, motion to static, and inference, you can download the provided pre-trained model (PubLayNet dataset) and skip this part.
|
||||
|
||||
```bash linenums="1"
|
||||
mkdir pretrained_model
|
||||
cd pretrained_model
|
||||
# Download PubLayNet pre-training model（Direct experience model evaluates, predicts, and turns static）
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout.pdparams
|
||||
# Download the PubLaynet inference model（Direct experience model reasoning）
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_infer.tar
|
||||
```
|
||||
|
||||
If the test image is Chinese, the pre-trained model of Chinese CDLA dataset can be downloaded to identify 10 types of document regions：Table, Figure, Figure caption, Table, Table caption, Header, Footer, Reference, Equation，Download the training model and inference model of Model 'picodet_lcnet_x1_0_fgd_layout_cdla' in [layout analysis model](../models_list.en.md)。If only the table area in the image is detected, you can download the pre-trained model of the table dataset, and download the training model and inference model of the 'picodet_LCnet_x1_0_FGd_layout_table' model in [Layout Analysis model](../models_list.en.md)
|
||||
|
||||
### 5.1. Train
|
||||
|
||||
Start training with the PaddleDetection [layout analysis profile](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.5/configs/picodet/legacy_model/application/layout_analysis)
|
||||
|
||||
- Modify Profile
|
||||
|
||||
If you want to train your own data set, you need to modify the data configuration and the number of categories in the configuration file.
|
||||
|
||||
Using 'configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml' as an example, the change is as follows:
|
||||
|
||||
```yaml linenums="1"
|
||||
metric: COCO
|
||||
# Number of categories
|
||||
num_classes: 5
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
# Modify to your own training data directory
|
||||
image_dir: train
|
||||
# Modify to your own training data label file
|
||||
anno_path: train.json
|
||||
# Modify to your own training data root directory
|
||||
dataset_dir: /root/publaynet/
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
# Modify to your own validation data directory
|
||||
image_dir: val
|
||||
# Modify to your own validation data label file
|
||||
anno_path: val.json
|
||||
# Modify to your own validation data root
|
||||
dataset_dir: /root/publaynet/
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
# Modify to your own test data label file
|
||||
anno_path: /root/publaynet/val.json
|
||||
```
|
||||
|
||||
- Start training. During training, PP picodet pre training model will be downloaded by default. There is no need to download in advance.
|
||||
|
||||
```bash linenums="1"
|
||||
# GPU training supports single-card and multi-card training
|
||||
# The training log is automatically saved to the log directory
|
||||
|
||||
# Single card training
|
||||
export CUDA_VISIBLE_DEVICES=0
|
||||
python3 tools/train.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--eval
|
||||
|
||||
# Multi-card training, with the -- GPUS parameter specifying the card number
|
||||
export CUDA_VISIBLE_DEVICES=0,1,2,3
|
||||
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--eval
|
||||
```
|
||||
|
||||
**Attention：**If the video memory is out during training, adjust Batch_size in TrainReader and base_LR in LearningRate. The published config is obtained by 8-card training. If the number of GPU cards is changed to 1, then the base_LR needs to be reduced by 8 times.
|
||||
|
||||
After starting training normally, you will see the following log output:
|
||||
|
||||
```
|
||||
[08/15 04:02:30] ppdet.utils.checkpoint INFO: Finish loading model weights: /root/.cache/paddle/weights/LCNet_x1_0_pretrained.pdparams
|
||||
[08/15 04:02:46] ppdet.engine INFO: Epoch: [0] [ 0/1929] learning_rate: 0.040000 loss_vfl: 1.216707 loss_bbox: 1.142163 loss_dfl: 0.544196 loss: 2.903065 eta: 17 days, 13:50:26 batch_cost: 15.7452 data_cost: 2.9112 ips: 1.5243 images/s
|
||||
[08/15 04:03:19] ppdet.engine INFO: Epoch: [0] [ 20/1929] learning_rate: 0.064000 loss_vfl: 1.180627 loss_bbox: 0.939552 loss_dfl: 0.442436 loss: 2.628206 eta: 2 days, 12:18:53 batch_cost: 1.5770 data_cost: 0.0008 ips: 15.2184 images/s
|
||||
[08/15 04:03:47] ppdet.engine INFO: Epoch: [0] [ 40/1929] learning_rate: 0.088000 loss_vfl: 0.543321 loss_bbox: 1.071401 loss_dfl: 0.457817 loss: 2.057003 eta: 2 days, 0:07:03 batch_cost: 1.3190 data_cost: 0.0007 ips: 18.1954 images/s
|
||||
[08/15 04:04:12] ppdet.engine INFO: Epoch: [0] [ 60/1929] learning_rate: 0.112000 loss_vfl: 0.630989 loss_bbox: 0.859183 loss_dfl: 0.384702 loss: 1.883143 eta: 1 day, 19:01:29 batch_cost: 1.2177 data_cost: 0.0006 ips: 19.7087 images/s
|
||||
```
|
||||
|
||||
- `--eval` indicates that the best model is saved as `output/picodet_lcnet_x1_0_layout/best_accuracy` by default during the evaluation process 。
|
||||
|
||||
**Note that the configuration file for prediction / evaluation must be consistent with the training.**
|
||||
|
||||
### 5.2. FGD Distillation Training
|
||||
|
||||
PaddleDetection supports FGD-based [Focal and Global Knowledge Distillation for Detectors]( https://arxiv.org/abs/2111.11837v1) The training process of the target detection model of distillation, FGD distillation is divided into two parts `Focal` and `Global`. `Focal` Distillation separates the foreground and background of the image, allowing the student model to focus on the key pixels of the foreground and background features of the teacher model respectively;`Global`Distillation section reconstructs the relationships between different pixels and transfers them from the teacher to the student to compensate for the global information lost in `Focal`Distillation.
|
||||
|
||||
Change the dataset and modify the data configuration and number of categories in the [TODO] configuration, referring to 4.1. Start training:
|
||||
|
||||
```bash linenums="1"
|
||||
# Single Card Training
|
||||
export CUDA_VISIBLE_DEVICES=0
|
||||
python3 tools/train.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
|
||||
--eval
|
||||
```
|
||||
|
||||
- `-c`: Specify the model configuration file.
|
||||
- `--slim_config`: Specify the compression policy profile.
|
||||
|
||||
## 6. Model evaluation and prediction
|
||||
|
||||
### 6.1. Indicator evaluation
|
||||
|
||||
Model parameters in training are saved by default in `output/picodet_ Lcnet_ X1_ 0_ Under the layout` directory. When evaluating indicators, you need to set `weights` to point to the saved parameter file.Assessment datasets can be accessed via `configs/picodet/legacy_ Model/application/layout_ Analysis/picodet_ Lcnet_ X1_ 0_ Layout. Yml` . Modify `EvalDataset` : `img_dir`,`anno_ Path`and`dataset_dir` setting.
|
||||
|
||||
```bash linenums="1"
|
||||
# GPU evaluation, weights as weights to be measured
|
||||
python3 tools/eval.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
-o weights=./output/picodet_lcnet_x1_0_layout/best_model
|
||||
```
|
||||
|
||||
The following information will be printed out, such as mAP, AP0.5, etc.
|
||||
|
||||
```python linenums="1"
|
||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.935
|
||||
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.979
|
||||
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.956
|
||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.404
|
||||
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.782
|
||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.969
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.539
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.938
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.949
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.495
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.818
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.978
|
||||
[08/15 07:07:09] ppdet.engine INFO: Total sample number: 11245, averge FPS: 24.405059207157436
|
||||
[08/15 07:07:09] ppdet.engine INFO: Best test bbox ap is 0.935.
|
||||
```
|
||||
|
||||
If you use the provided pre-training model for evaluation or the FGD distillation training model, replace the `weights` model path and execute the following command for evaluation:
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/eval.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
|
||||
-o weights=output/picodet_lcnet_x2_5_layout/best_model
|
||||
```
|
||||
|
||||
- `-c`: Specify the model configuration file.
|
||||
- `--slim_config`: Specify the distillation policy profile.
|
||||
- `-o weights`: Specify the model path trained by the distillation algorithm.
|
||||
|
||||
### 6.2. Test Layout Analysis Results
|
||||
|
||||
The profile predicted to be used must be consistent with the training, for example, if you pass `python3 tools/train'. Py-c configs/picodet/legacy_ Model/application/layout_ Analysis/picodet_ Lcnet_ X1_ 0_ Layout. Yml` completed the training process for the model.
|
||||
|
||||
With trained PaddleDetection model, you can use the following commands to make model predictions.
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/infer.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
-o weights='output/picodet_lcnet_x1_0_layout/best_model.pdparams' \
|
||||
--infer_img='docs/images/layout.jpg' \
|
||||
--output_dir=output_dir/ \
|
||||
--draw_threshold=0.5
|
||||
```
|
||||
|
||||
- `--infer_img`: Reasoning for a single picture can also be done via `--infer_ Dir`Inform all pictures in the file.
|
||||
- `--output_dir`: Specify the path to save the visualization results.
|
||||
- `--draw_threshold`:Specify the NMS threshold for drawing the result box.
|
||||
|
||||
If you use the provided pre-training model for prediction or the FGD distillation training model, change the `weights` model path and execute the following command to make the prediction:
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/infer.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
|
||||
-o weights='output/picodet_lcnet_x2_5_layout/best_model.pdparams' \
|
||||
--infer_img='docs/images/layout.jpg' \
|
||||
--output_dir=output_dir/ \
|
||||
--draw_threshold=0.5
|
||||
```
|
||||
|
||||
## 7. Model Export and Inference
|
||||
|
||||
### 7.1 Model Export
|
||||
|
||||
The inference model (the model saved by `paddle.jit.save`) is generally a solidified model saved after the model training is completed, and is mostly used to give prediction in deployment.
|
||||
|
||||
The model saved during the training process is the checkpoints model, which saves the parameters of the model and is mostly used to resume training.
|
||||
|
||||
Compared with the checkpoints model, the inference model will additionally save the structural information of the model. Therefore, it is easier to deploy because the model structure and model parameters are already solidified in the inference model file, and is suitable for integration with actual systems.
|
||||
|
||||
Layout analysis model to inference model steps are as follows：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/export_model.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
-o weights=output/picodet_lcnet_x1_0_layout/best_model \
|
||||
--output_dir=output_inference/
|
||||
```
|
||||
|
||||
- If no post-export processing is required, specify：`-o export.benchmark=True`（If -o already exists, delete -o here）
|
||||
- If you do not need to export NMS, specify：`-o export.nms=False`
|
||||
|
||||
After successful conversion, there are three files in the directory:
|
||||
|
||||
```
|
||||
output_inference/picodet_lcnet_x1_0_layout/
|
||||
├── model.pdiparams # inference Parameter file for model
|
||||
├── model.pdiparams.info # inference Model parameter information, ignorable
|
||||
└── model.pdmodel # inference Model Structure File for Model
|
||||
```
|
||||
|
||||
If you change the `weights` model path using the provided pre-training model to the Inference model, or using the FGD distillation training model, the model to inference model steps are as follows:
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/export_model.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
|
||||
-o weights=./output/picodet_lcnet_x2_5_layout/best_model \
|
||||
--output_dir=output_inference/
|
||||
```
|
||||
|
||||
### 7.2 Model inference
|
||||
|
||||
Replace model_with the provided inference training model for inference or the FGD distillation training `model_dir`Inference model path, execute the following commands for inference:
|
||||
|
||||
```bash linenums="1"
|
||||
python3 deploy/python/infer.py \
|
||||
--model_dir=output_inference/picodet_lcnet_x1_0_layout/ \
|
||||
--image_file=docs/images/layout.jpg \
|
||||
--device=CPU
|
||||
```
|
||||
|
||||
- --device：Specify the GPU or CPU device
|
||||
|
||||
When model inference is complete, you will see the following log output:
|
||||
|
||||
```
|
||||
------------------------------------------
|
||||
----------- Model Configuration -----------
|
||||
Model Arch: PicoDet
|
||||
Transform Order:
|
||||
--transform op: Resize
|
||||
--transform op: NormalizeImage
|
||||
--transform op: Permute
|
||||
--transform op: PadStride
|
||||
--------------------------------------------
|
||||
class_id:0, confidence:0.9921, left_top:[20.18,35.66],right_bottom:[341.58,600.99]
|
||||
class_id:0, confidence:0.9914, left_top:[19.77,611.42],right_bottom:[341.48,901.82]
|
||||
class_id:0, confidence:0.9904, left_top:[369.36,375.10],right_bottom:[691.29,600.59]
|
||||
class_id:0, confidence:0.9835, left_top:[369.60,608.60],right_bottom:[691.38,736.72]
|
||||
class_id:0, confidence:0.9830, left_top:[369.58,805.38],right_bottom:[690.97,901.80]
|
||||
class_id:0, confidence:0.9716, left_top:[383.68,271.44],right_bottom:[688.93,335.39]
|
||||
class_id:0, confidence:0.9452, left_top:[370.82,34.48],right_bottom:[688.10,63.54]
|
||||
class_id:1, confidence:0.8712, left_top:[370.84,771.03],right_bottom:[519.30,789.13]
|
||||
class_id:3, confidence:0.9856, left_top:[371.28,67.85],right_bottom:[685.73,267.72]
|
||||
save result to: output/layout.jpg
|
||||
Test iter 0
|
||||
------------------ Inference Time Info ----------------------
|
||||
total_time(ms): 2196.0, img_num: 1
|
||||
average latency time(ms): 2196.00, QPS: 0.455373
|
||||
preprocess_time(ms): 2172.50, inference_time(ms): 11.90, postprocess_time(ms): 11.60
|
||||
```
|
||||
|
||||
- Model：model structure
|
||||
- Transform Order：Preprocessing operation
|
||||
- class_id, confidence, left_top, right_bottom：Indicates category id, confidence level, upper left coordinate, lower right coordinate, respectively
|
||||
- save result to：Save path of visual layout analysis results, default save to ./output folder
|
||||
- inference time info：Inference time, where preprocess_time represents the preprocessing time, Inference_time represents the model prediction time, and postprocess_time represents the post-processing time
|
||||
|
||||
The result of visualization layout is shown in the following figure
|
||||
|
||||
|
||||
|
||||
## Citations
|
||||
|
||||
```bibtex
|
||||
@inproceedings{zhong2019publaynet,
|
||||
title={PubLayNet: largest dataset ever for document layout analysis},
|
||||
author={Zhong, Xu and Tang, Jianbin and Yepes, Antonio Jimeno},
|
||||
booktitle={2019 International Conference on Document Analysis and Recognition (ICDAR)},
|
||||
year={2019},
|
||||
volume={},
|
||||
number={},
|
||||
pages={1015-1022},
|
||||
doi={10.1109/ICDAR.2019.00166},
|
||||
ISSN={1520-5363},
|
||||
month={Sep.},
|
||||
organization={IEEE}
|
||||
}
|
||||
|
||||
@inproceedings{yang2022focal,
|
||||
title={Focal and global knowledge distillation for detectors},
|
||||
author={Yang, Zhendong and Li, Zhe and Jiang, Xiaohu and Gong, Yuan and Yuan, Zehuan and Zhao, Danpei and Yuan, Chun},
|
||||
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
|
||||
pages={4643--4652},
|
||||
year={2022}
|
||||
}
|
||||
```
|
||||
441
docs/version2.x/ppstructure/model_train/train_layout.md
Normal file
441
docs/version2.x/ppstructure/model_train/train_layout.md
Normal file
@@ -0,0 +1,441 @@
|
||||
---
|
||||
typora-copy-images-to: images
|
||||
comments: true
|
||||
---
|
||||
|
||||
# 版面分析
|
||||
|
||||
## 1. 简介
|
||||
|
||||
版面分析指的是对图片形式的文档进行区域划分，定位其中的关键区域，如文字、标题、表格、图片等。版面分析算法基于[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection)的轻量模型PP-PicoDet进行开发，包含英文、中文、表格版面分析3类模型。其中，英文模型支持Text、Title、Tale、Figure、List5类区域的检测，中文模型支持Text、Title、Figure、Figure caption、Table、Table caption、Header、Footer、Reference、Equation10类区域的检测，表格版面分析支持Table区域的检测，版面分析效果如下图所示：
|
||||
|
||||
|
||||
|
||||
## 2. 快速开始
|
||||
|
||||
PP-Structure目前提供了中文、英文、表格三类文档版面分析模型，模型链接见 [models_list](../models_list.md)。也提供了whl包的形式方便快速使用，详见 [quickstart](../quick_start.md)。
|
||||
|
||||
## 3. 安装
|
||||
|
||||
### 3.1. 安装PaddlePaddle
|
||||
|
||||
- **（1) 安装PaddlePaddle**
|
||||
|
||||
```bash linenums="1"
|
||||
python3 -m pip install --upgrade pip
|
||||
|
||||
# GPU安装
|
||||
python3 -m pip install "paddlepaddle-gpu>=2.3" -i https://mirror.baidu.com/pypi/simple
|
||||
|
||||
# CPU安装
|
||||
python3 -m pip install "paddlepaddle>=2.3" -i https://mirror.baidu.com/pypi/simple
|
||||
```
|
||||
|
||||
更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
|
||||
|
||||
### 3.2. 安装PaddleDetection
|
||||
|
||||
- **（1）下载PaddleDetection源码**
|
||||
|
||||
```bash linenums="1"
|
||||
git clone https://github.com/PaddlePaddle/PaddleDetection.git
|
||||
```
|
||||
|
||||
- **（2）安装其他依赖**
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleDetection
|
||||
python3 -m pip install -r requirements.txt
|
||||
```
|
||||
|
||||
## 4. 数据准备
|
||||
|
||||
如果希望直接体验预测过程，可以跳过数据准备，下载我们提供的预训练模型。
|
||||
|
||||
### 4.1. 英文数据集
|
||||
|
||||
下载文档分析数据集[PubLayNet](https://developer.ibm.com/exchanges/data/all/publaynet/)（数据集96G），包含5个类：`{0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"}`
|
||||
|
||||
```
|
||||
# 下载数据
|
||||
wget https://dax-cdn.cdn.appdomain.cloud/dax-publaynet/1.0.0/publaynet.tar.gz
|
||||
# 解压数据
|
||||
tar -xvf publaynet.tar.gz
|
||||
```
|
||||
|
||||
解压之后的**目录结构：**
|
||||
|
||||
```
|
||||
|-publaynet
|
||||
|- test
|
||||
|- PMC1277013_00004.jpg
|
||||
|- PMC1291385_00002.jpg
|
||||
| ...
|
||||
|- train.json
|
||||
|- train
|
||||
|- PMC1291385_00002.jpg
|
||||
|- PMC1277013_00004.jpg
|
||||
| ...
|
||||
|- val.json
|
||||
|- val
|
||||
|- PMC538274_00004.jpg
|
||||
|- PMC539300_00004.jpg
|
||||
| ...
|
||||
```
|
||||
|
||||
**数据分布：**
|
||||
|
||||
| File or Folder | Description | num |
|
||||
| :------------- | :------------- | ------- |
|
||||
| `train/` | 训练集图片 | 335,703 |
|
||||
| `val/` | 验证集图片 | 11,245 |
|
||||
| `test/` | 测试集图片 | 11,405 |
|
||||
| `train.json` | 训练集标注文件 | - |
|
||||
| `val.json` | 验证集标注文件 | - |
|
||||
|
||||
**标注格式：**
|
||||
|
||||
json文件包含所有图像的标注，数据以字典嵌套的方式存放，包含以下key：
|
||||
|
||||
- info，表示标注文件info。
|
||||
|
||||
- licenses，表示标注文件licenses。
|
||||
|
||||
- images，表示标注文件中图像信息列表，每个元素是一张图像的信息。如下为其中一张图像的信息：
|
||||
|
||||
```
|
||||
{
|
||||
'file_name': 'PMC4055390_00006.jpg', # file_name
|
||||
'height': 601, # image height
|
||||
'width': 792, # image width
|
||||
'id': 341427 # image id
|
||||
}
|
||||
```
|
||||
|
||||
- annotations，表示标注文件中目标物体的标注信息列表，每个元素是一个目标物体的标注信息。如下为其中一个目标物体的标注信息：
|
||||
|
||||
```
|
||||
{
|
||||
|
||||
'segmentation': # 物体的分割标注
|
||||
'area': 60518.099043117836, # 物体的区域面积
|
||||
'iscrowd': 0, # iscrowd
|
||||
'image_id': 341427, # image id
|
||||
'bbox': [50.58, 490.86, 240.15, 252.16], # bbox [x1,y1,w,h]
|
||||
'category_id': 1, # category_id
|
||||
'id': 3322348 # image id
|
||||
}
|
||||
```
|
||||
|
||||
### 4.2. 更多数据集
|
||||
|
||||
我们提供了CDLA(中文版面分析)、TableBank(表格版面分析)等数据集的下连接，处理为上述标注文件json格式，即可以按相同方式进行训练。
|
||||
|
||||
| dataset | 简介 |
|
||||
| --------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
|
||||
| [cTDaR2019_cTDaR](https://cndplab-founder.github.io/cTDaR2019/) | 用于表格检测(TRACKA)和表格识别(TRACKB)。图片类型包含历史数据集(以cTDaR_t0开头，如cTDaR_t00872.jpg)和现代数据集(以cTDaR_t1开头，cTDaR_t10482.jpg)。 |
|
||||
| [IIIT-AR-13K](http://cvit.iiit.ac.in/usodi/iiitar13k.php) | 手动注释公开的年度报告中的图形或页面而构建的数据集，包含5类：table, figure, natural image, logo, and signature |
|
||||
| [CDLA](https://github.com/buptlihang/CDLA) | 中文文档版面分析数据集，面向中文文献类（论文）场景，包含10类：Text、Title、Figure、Figure caption、Table、Table caption、Header、Footer、Reference、Equation |
|
||||
| [TableBank](https://github.com/doc-analysis/TableBank) | 用于表格检测和识别大型数据集，包含Word和Latex2种文档格式 |
|
||||
| [DocBank](https://github.com/doc-analysis/DocBank) | 使用弱监督方法构建的大规模数据集(500K文档页面)，用于文档布局分析，包含12类：Author、Caption、Date、Equation、Figure、Footer、List、Paragraph、Reference、Section、Table、Title |
|
||||
|
||||
## 5. 开始训练
|
||||
|
||||
提供了训练脚本、评估脚本和预测脚本，本节将以PubLayNet预训练模型为例进行讲解。
|
||||
|
||||
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，可以下载提供的预训练模型(PubLayNet数据集)，并跳过5.1和5.2。
|
||||
|
||||
```bash linenums="1"
|
||||
mkdir pretrained_model
|
||||
cd pretrained_model
|
||||
# 下载PubLayNet预训练模型（直接体验模型评估、预测、动转静）
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout.pdparams
|
||||
# 下载PubLaynet推理模型（直接体验模型推理）
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_infer.tar
|
||||
```
|
||||
|
||||
如果测试图片为中文，可以下载中文CDLA数据集的预训练模型，识别10类文档区域：Table、Figure、Figure caption、Table、Table caption、Header、Footer、Reference、Equation，在[版面分析模型](../models_list.md)中下载`picodet_lcnet_x1_0_fgd_layout_cdla`模型的训练模型和推理模型。如果只检测图片中的表格区域，可以下载表格数据集的预训练模型，在[版面分析模型](../models_list.md)中下载`picodet_lcnet_x1_0_fgd_layout_table`模型的训练模型和推理模型。
|
||||
|
||||
### 5.1. 启动训练
|
||||
|
||||
使用PaddleDetection[版面分析配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/release/2.5/configs/picodet/legacy_model/application/layout_analysis)启动训练
|
||||
|
||||
- 修改配置文件
|
||||
|
||||
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、类别数。
|
||||
|
||||
以`configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml` 为例，修改的内容如下所示。
|
||||
|
||||
```yaml linenums="1"
|
||||
metric: COCO
|
||||
# 类别数
|
||||
num_classes: 5
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
# 修改为你自己的训练数据目录
|
||||
image_dir: train
|
||||
# 修改为你自己的训练数据标签文件
|
||||
anno_path: train.json
|
||||
# 修改为你自己的训练数据根目录
|
||||
dataset_dir: /root/publaynet/
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
# 修改为你自己的验证数据目录
|
||||
image_dir: val
|
||||
# 修改为你自己的验证数据标签文件
|
||||
anno_path: val.json
|
||||
# 修改为你自己的验证数据根目录
|
||||
dataset_dir: /root/publaynet/
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
# 修改为你自己的测试数据标签文件
|
||||
anno_path: /root/publaynet/val.json
|
||||
```
|
||||
|
||||
- 开始训练，在训练时，会默认下载PP-PicoDet预训练模型，这里无需预先下载。
|
||||
|
||||
```bash linenums="1"
|
||||
# GPU训练 支持单卡，多卡训练
|
||||
# 训练日志会自动保存到 log 目录中
|
||||
|
||||
# 单卡训练
|
||||
export CUDA_VISIBLE_DEVICES=0
|
||||
python3 tools/train.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--eval
|
||||
|
||||
# 多卡训练，通过--gpus参数指定卡号
|
||||
export CUDA_VISIBLE_DEVICES=0,1,2,3
|
||||
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--eval
|
||||
```
|
||||
|
||||
**注意：**如果训练时显存out memory，将TrainReader中batch_size调小，同时LearningRate中base_lr等比例减小。发布的config均由8卡训练得到，如果改变GPU卡数为1，那么base_lr需要减小8倍。
|
||||
|
||||
正常启动训练后，会看到以下log输出：
|
||||
|
||||
```bash linenums="1"
|
||||
[08/15 04:02:30] ppdet.utils.checkpoint INFO: Finish loading model weights: /root/.cache/paddle/weights/LCNet_x1_0_pretrained.pdparams
|
||||
[08/15 04:02:46] ppdet.engine INFO: Epoch: [0] [ 0/1929] learning_rate: 0.040000 loss_vfl: 1.216707 loss_bbox: 1.142163 loss_dfl: 0.544196 loss: 2.903065 eta: 17 days, 13:50:26 batch_cost: 15.7452 data_cost: 2.9112 ips: 1.5243 images/s
|
||||
[08/15 04:03:19] ppdet.engine INFO: Epoch: [0] [ 20/1929] learning_rate: 0.064000 loss_vfl: 1.180627 loss_bbox: 0.939552 loss_dfl: 0.442436 loss: 2.628206 eta: 2 days, 12:18:53 batch_cost: 1.5770 data_cost: 0.0008 ips: 15.2184 images/s
|
||||
[08/15 04:03:47] ppdet.engine INFO: Epoch: [0] [ 40/1929] learning_rate: 0.088000 loss_vfl: 0.543321 loss_bbox: 1.071401 loss_dfl: 0.457817 loss: 2.057003 eta: 2 days, 0:07:03 batch_cost: 1.3190 data_cost: 0.0007 ips: 18.1954 images/s
|
||||
[08/15 04:04:12] ppdet.engine INFO: Epoch: [0] [ 60/1929] learning_rate: 0.112000 loss_vfl: 0.630989 loss_bbox: 0.859183 loss_dfl: 0.384702 loss: 1.883143 eta: 1 day, 19:01:29 batch_cost: 1.2177 data_cost: 0.0006 ips: 19.7087 images/s
|
||||
```
|
||||
|
||||
- `--eval`表示训练的同时，进行评估， 评估过程中默认将最佳模型，保存为 `output/picodet_lcnet_x1_0_layout/best_accuracy` 。
|
||||
|
||||
**注意，预测/评估时的配置文件请务必与训练一致。**
|
||||
|
||||
### 5.2. FGD蒸馏训练
|
||||
|
||||
PaddleDetection支持了基于FGD([Focal and Global Knowledge Distillation for Detectors](https://arxiv.org/abs/2111.11837v1))蒸馏的目标检测模型训练过程，FGD蒸馏分为两个部分`Focal`和`Global`。`Focal`蒸馏分离图像的前景和背景，让学生模型分别关注教师模型的前景和背景部分特征的关键像素；`Global`蒸馏部分重建不同像素之间的关系并将其从教师转移到学生，以补偿`Focal`蒸馏中丢失的全局信息。
|
||||
|
||||
更换数据集，修改【TODO】配置中的数据配置、类别数，具体可以参考4.1。启动训练：
|
||||
|
||||
```bash linenums="1"
|
||||
# 单卡训练
|
||||
export CUDA_VISIBLE_DEVICES=0
|
||||
python3 tools/train.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
|
||||
--eval
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件。
|
||||
- `--slim_config`: 指定压缩策略配置文件。
|
||||
|
||||
## 6. 模型评估与预测
|
||||
|
||||
### 6.1. 指标评估
|
||||
|
||||
训练中模型参数默认保存在`output/picodet_lcnet_x1_0_layout`目录下。在评估指标时，需要设置`weights`指向保存的参数文件。评估数据集可以通过 `configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml` 修改`EvalDataset`中的 `image_dir`、`anno_path`和`dataset_dir` 设置。
|
||||
|
||||
```bash linenums="1"
|
||||
# GPU 评估， weights 为待测权重
|
||||
python3 tools/eval.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
-o weights=./output/picodet_lcnet_x1_0_layout/best_model
|
||||
```
|
||||
|
||||
会输出以下信息，打印出mAP、AP0.5等信息。
|
||||
|
||||
```python linenums="1"
|
||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.935
|
||||
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.979
|
||||
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.956
|
||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.404
|
||||
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.782
|
||||
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.969
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.539
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.938
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.949
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.495
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.818
|
||||
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.978
|
||||
[08/15 07:07:09] ppdet.engine INFO: Total sample number: 11245, averge FPS: 24.405059207157436
|
||||
[08/15 07:07:09] ppdet.engine INFO: Best test bbox ap is 0.935.
|
||||
```
|
||||
|
||||
若使用**提供的预训练模型进行评估**，或使用**FGD蒸馏训练的模型**，更换`weights`模型路径，执行如下命令进行评估：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/eval.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
|
||||
-o weights=output/picodet_lcnet_x2_5_layout/best_model
|
||||
```
|
||||
|
||||
- `-c`: 指定模型配置文件。
|
||||
- `--slim_config`: 指定蒸馏策略配置文件。
|
||||
- `-o weights`: 指定蒸馏算法训好的模型路径。
|
||||
|
||||
### 6.2 测试版面分析结果
|
||||
|
||||
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/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml` 完成了模型的训练过程。
|
||||
|
||||
使用 PaddleDetection 训练好的模型，您可以使用如下命令进行模型预测。
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/infer.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
-o weights='output/picodet_lcnet_x1_0_layout/best_model.pdparams' \
|
||||
--infer_img='docs/images/layout.jpg' \
|
||||
--output_dir=output_dir/ \
|
||||
--draw_threshold=0.5
|
||||
```
|
||||
|
||||
- `--infer_img`: 推理单张图片，也可以通过`--infer_dir`推理文件中的所有图片。
|
||||
- `--output_dir`: 指定可视化结果保存路径。
|
||||
- `--draw_threshold`:指定绘制结果框的NMS阈值。
|
||||
|
||||
若使用**提供的预训练模型进行预测**，或使用**FGD蒸馏训练的模型**，更换`weights`模型路径，执行如下命令进行预测：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/infer.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
|
||||
-o weights='output/picodet_lcnet_x2_5_layout/best_model.pdparams' \
|
||||
--infer_img='docs/images/layout.jpg' \
|
||||
--output_dir=output_dir/ \
|
||||
--draw_threshold=0.5
|
||||
```
|
||||
|
||||
## 7. 模型导出与预测
|
||||
|
||||
### 7.1 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。
|
||||
|
||||
版面分析模型转inference模型步骤如下：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/export_model.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
-o weights=output/picodet_lcnet_x1_0_layout/best_model \
|
||||
--output_dir=output_inference/
|
||||
```
|
||||
|
||||
- 如无需导出后处理，请指定：`-o export.benchmark=True`（如果-o已出现过，此处删掉-o）
|
||||
- 如无需导出NMS，请指定：`-o export.nms=False`
|
||||
|
||||
转换成功后，在目录下有三个文件：
|
||||
|
||||
```
|
||||
output_inference/picodet_lcnet_x1_0_layout/
|
||||
├── model.pdiparams # inference模型的参数文件
|
||||
├── model.pdiparams.info # inference模型的参数信息，可忽略
|
||||
└── model.pdmodel # inference模型的模型结构文件
|
||||
```
|
||||
|
||||
若使用**提供的预训练模型转Inference模型**，或使用**FGD蒸馏训练的模型**，更换`weights`模型路径，模型转inference模型步骤如下：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 tools/export_model.py \
|
||||
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
|
||||
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
|
||||
-o weights=./output/picodet_lcnet_x2_5_layout/best_model \
|
||||
--output_dir=output_inference/
|
||||
```
|
||||
|
||||
### 7.2 模型推理
|
||||
|
||||
若使用**提供的推理训练模型推理**，或使用**FGD蒸馏训练的模型**，更换`model_dir`推理模型路径，执行如下命令进行推理：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 deploy/python/infer.py \
|
||||
--model_dir=output_inference/picodet_lcnet_x1_0_layout/ \
|
||||
--image_file=docs/images/layout.jpg \
|
||||
--device=CPU
|
||||
```
|
||||
|
||||
- --device：指定GPU、CPU设备
|
||||
|
||||
模型推理完成，会看到以下log输出
|
||||
|
||||
```
|
||||
------------------------------------------
|
||||
----------- Model Configuration -----------
|
||||
Model Arch: PicoDet
|
||||
Transform Order:
|
||||
--transform op: Resize
|
||||
--transform op: NormalizeImage
|
||||
--transform op: Permute
|
||||
--transform op: PadStride
|
||||
--------------------------------------------
|
||||
class_id:0, confidence:0.9921, left_top:[20.18,35.66],right_bottom:[341.58,600.99]
|
||||
class_id:0, confidence:0.9914, left_top:[19.77,611.42],right_bottom:[341.48,901.82]
|
||||
class_id:0, confidence:0.9904, left_top:[369.36,375.10],right_bottom:[691.29,600.59]
|
||||
class_id:0, confidence:0.9835, left_top:[369.60,608.60],right_bottom:[691.38,736.72]
|
||||
class_id:0, confidence:0.9830, left_top:[369.58,805.38],right_bottom:[690.97,901.80]
|
||||
class_id:0, confidence:0.9716, left_top:[383.68,271.44],right_bottom:[688.93,335.39]
|
||||
class_id:0, confidence:0.9452, left_top:[370.82,34.48],right_bottom:[688.10,63.54]
|
||||
class_id:1, confidence:0.8712, left_top:[370.84,771.03],right_bottom:[519.30,789.13]
|
||||
class_id:3, confidence:0.9856, left_top:[371.28,67.85],right_bottom:[685.73,267.72]
|
||||
save result to: output/layout.jpg
|
||||
Test iter 0
|
||||
------------------ Inference Time Info ----------------------
|
||||
total_time(ms): 2196.0, img_num: 1
|
||||
average latency time(ms): 2196.00, QPS: 0.455373
|
||||
preprocess_time(ms): 2172.50, inference_time(ms): 11.90, postprocess_time(ms): 11.60
|
||||
```
|
||||
|
||||
- Model：模型结构
|
||||
- Transform Order：预处理操作
|
||||
- class_id、confidence、left_top、right_bottom：分别表示类别id、置信度、左上角坐标、右下角坐标
|
||||
- save result to：可视化版面分析结果保存路径，默认保存到`./output`文件夹
|
||||
- Inference Time Info：推理时间，其中preprocess_time表示预处理耗时，inference_time表示模型预测耗时，postprocess_time表示后处理耗时
|
||||
|
||||
可视化版面结果如下图所示
|
||||
|
||||
|
||||
|
||||
## Citations
|
||||
|
||||
```bibtex
|
||||
@inproceedings{zhong2019publaynet,
|
||||
title={PubLayNet: largest dataset ever for document layout analysis},
|
||||
author={Zhong, Xu and Tang, Jianbin and Yepes, Antonio Jimeno},
|
||||
booktitle={2019 International Conference on Document Analysis and Recognition (ICDAR)},
|
||||
year={2019},
|
||||
volume={},
|
||||
number={},
|
||||
pages={1015-1022},
|
||||
doi={10.1109/ICDAR.2019.00166},
|
||||
ISSN={1520-5363},
|
||||
month={Sep.},
|
||||
organization={IEEE}
|
||||
}
|
||||
|
||||
@inproceedings{yang2022focal,
|
||||
title={Focal and global knowledge distillation for detectors},
|
||||
author={Yang, Zhendong and Li, Zhe and Jiang, Xiaohu and Gong, Yuan and Yuan, Zehuan and Zhao, Danpei and Yuan, Chun},
|
||||
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
|
||||
pages={4643--4652},
|
||||
year={2022}
|
||||
}
|
||||
```
|
||||
164
docs/version2.x/ppstructure/model_train/train_table.en.md
Normal file
164
docs/version2.x/ppstructure/model_train/train_table.en.md
Normal file
@@ -0,0 +1,164 @@
|
||||
---
|
||||
comments: true
|
||||
---
|
||||
|
||||
# Table Recognition
|
||||
|
||||
## 1. pipeline
|
||||
|
||||
The table recognition mainly contains three models
|
||||
|
||||
1. Single line text detection-DB
|
||||
2. Single line text recognition-CRNN
|
||||
3. Table structure and cell coordinate prediction-SLANet
|
||||
|
||||
The table recognition flow chart is as follows
|
||||
|
||||
|
||||
|
||||
1. The coordinates of single-line text is detected by DB model, and then sends it to the recognition model to get the recognition result.
|
||||
2. The table structure and cell coordinates is predicted by SLANet model.
|
||||
3. The recognition result of the cell is combined by the coordinates, recognition result of the single line and the coordinates of the cell.
|
||||
4. The cell recognition result and the table structure together construct the html string of the table.
|
||||
|
||||
## 2. Performance
|
||||
|
||||
We evaluated the algorithm on the PubTabNet<sup>[1]</sup> eval dataset, and the performance is as follows:
|
||||
|
||||
|Method|Acc|[TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src)|Speed|
|
||||
| --- | --- | --- | ---|
|
||||
| EDD<sup>[2]</sup> |x| 88.30% |x|
|
||||
| TableRec-RARE(ours) | 71.73%| 93.88% |779ms|
|
||||
| SLANet(ours) | 76.31%| 95.89%|766ms|
|
||||
|
||||
The performance indicators are explained as follows:
|
||||
|
||||
- Acc: The accuracy of the table structure in each image, a wrong token is considered an error.
|
||||
- TEDS: The accuracy of the model's restoration of table information. This indicator evaluates not only the table structure, but also the text content in the table.
|
||||
- Speed: The inference speed of a single image when the model runs on the CPU machine and MKL is enabled.
|
||||
|
||||
## 3. Result
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
## 4. How to use
|
||||
|
||||
### 4.1 Quick start
|
||||
|
||||
PP-Structure currently provides table recognition models in both Chinese and English. For the model link, see [models_list](../models_list.en.md). The whl package is also provided for quick use, see [quickstart](../quick_start.en.md) for details.
|
||||
|
||||
The following takes the Chinese table recognition model as an example to introduce how to recognize a table.
|
||||
|
||||
Use the following commands to quickly complete the identification of a table.
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleOCR/ppstructure
|
||||
|
||||
# download model
|
||||
mkdir inference && cd inference
|
||||
# Download the PP-OCRv3 text detection model and unzip it
|
||||
wget https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv3_mobile_det_infer.tar && tar xf PP-OCRv3_mobile_det_infer.tar
|
||||
# Download the PP-OCRv3 text recognition model and unzip it
|
||||
wget https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv3_mobile_rec_infer.tar && tar xf PP-OCRv3_mobile_rec_infer.tar
|
||||
# Download the PP-StructureV2 form recognition model and unzip it
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/paddle3.0b2/ch_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf ch_ppstructure_mobile_v2.0_SLANet_infer.tar
|
||||
cd ..
|
||||
# run
|
||||
python3 table/predict_table.py \
|
||||
--det_model_dir=inference/PP-OCRv3_mobile_det_infer \
|
||||
--rec_model_dir=inference/PP-OCRv3_mobile_rec_infer \
|
||||
--table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \
|
||||
--rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \
|
||||
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \
|
||||
--image_dir=docs/table/table.jpg \
|
||||
--output=../output/table
|
||||
|
||||
```
|
||||
|
||||
After the operation is completed, the excel table of each image will be saved to the directory specified by the output field, and an html file will be produced in the directory to visually view the cell coordinates and the recognized table.
|
||||
|
||||
**NOTE**
|
||||
|
||||
1. If you want to use the English table recognition model, you need to download the English text detection and recognition model and the English table recognition model in [models_list](../models_list.en.md), and replace `table_structure_dict_ch.txt` with `table_structure_dict.txt`.
|
||||
2. To use the TableRec-RARE model, you need to replace `table_structure_dict_ch.txt` with `table_structure_dict.txt`, and add parameter `--merge_no_span_structure=False`
|
||||
|
||||
### 4.2 Training, Evaluation and Inference
|
||||
|
||||
The training, evaluation and inference process of the text detection model can be referred to [detection](../../ppocr/model_train/detection.en.md)
|
||||
|
||||
The training, evaluation and inference process of the text recognition model can be referred to [recognition](../../ppocr/model_train/recognition.en.md)
|
||||
|
||||
The training, evaluation and inference process of the table recognition model can be referred to [table_recognition](./train_table.en.md)
|
||||
|
||||
### 4.3 Calculate TEDS
|
||||
|
||||
The table uses [TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src) as the evaluation metric of the model. Before the model evaluation, the three models in the pipeline need to be exported as inference models (we have provided them), and the gt for evaluation needs to be prepared. Examples of gt are as follows:
|
||||
|
||||
```txt
|
||||
PMC5755158_010_01.png <html><body><table><thead><tr><td></td><td><b>Weaning</b></td><td><b>Week 15</b></td><td><b>Off-test</b></td></tr></thead><tbody><tr><td>Weaning</td><td>–</td><td>–</td><td>–</td></tr><tr><td>Week 15</td><td>–</td><td>0.17 ± 0.08</td><td>0.16 ± 0.03</td></tr><tr><td>Off-test</td><td>–</td><td>0.80 ± 0.24</td><td>0.19 ± 0.09</td></tr></tbody></table></body></html>
|
||||
```
|
||||
|
||||
Each line in gt consists of the file name and the html string of the table. The file name and the html string of the table are separated by `\t`.
|
||||
|
||||
You can also use the following command to generate an evaluation gt file from the annotation file:
|
||||
|
||||
```bash linenums="1"
|
||||
python3 ppstructure/table/convert_label2html.py --ori_gt_path /path/to/your_label_file --save_path /path/to/save_file
|
||||
```
|
||||
|
||||
Use the following command to evaluate. After the evaluation is completed, the teds indicator will be output.
|
||||
|
||||
```bash linenums="1"
|
||||
python3 table/eval_table.py \
|
||||
--det_model_dir=path/to/det_model_dir \
|
||||
--rec_model_dir=path/to/rec_model_dir \
|
||||
--table_model_dir=path/to/table_model_dir \
|
||||
--image_dir=docs/table/table.jpg \
|
||||
--rec_char_dict_path=../ppocr/utils/dict/table_dict.txt \
|
||||
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \
|
||||
--det_limit_side_len=736 \
|
||||
--det_limit_type=min \
|
||||
--gt_path=path/to/gt.txt
|
||||
```
|
||||
|
||||
Evaluate on the PubLatNet dataset using the English model
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleOCR/ppstructure
|
||||
# Download the model
|
||||
mkdir inference && cd inference
|
||||
# Download the text detection model trained on the PubTabNet dataset and unzip it
|
||||
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar && tar xf en_ppocr_mobile_v2.0_table_det_infer.tar
|
||||
# Download the text recognition model trained on the PubTabNet dataset and unzip it
|
||||
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar && tar xf en_ppocr_mobile_v2.0_table_rec_infer.tar
|
||||
# Download the table recognition model trained on the PubTabNet dataset and unzip it
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/paddle3.0b2/en_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf en_ppstructure_mobile_v2.0_SLANet_infer.tar
|
||||
cd ..
|
||||
|
||||
python3 table/eval_table.py \
|
||||
--det_model_dir=inference/en_ppocr_mobile_v2.0_table_det_infer \
|
||||
--rec_model_dir=inference/en_ppocr_mobile_v2.0_table_rec_infer \
|
||||
--table_model_dir=inference/en_ppstructure_mobile_v2.0_SLANet_infer \
|
||||
--image_dir=train_data/table/pubtabnet/val/ \
|
||||
--rec_char_dict_path=../ppocr/utils/dict/table_dict.txt \
|
||||
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \
|
||||
--det_limit_side_len=736 \
|
||||
--det_limit_type=min \
|
||||
--rec_image_shape=3,32,320 \
|
||||
--gt_path=path/to/gt.txt
|
||||
```
|
||||
|
||||
output is
|
||||
|
||||
```bash linenums="1"
|
||||
teds: 95.89
|
||||
```
|
||||
|
||||
## 5. Reference
|
||||
|
||||
1. <https://github.com/ibm-aur-nlp/PubTabNet>
|
||||
2. <https://arxiv.org/pdf/1911.10683>
|
||||
167
docs/version2.x/ppstructure/model_train/train_table.md
Normal file
167
docs/version2.x/ppstructure/model_train/train_table.md
Normal file
@@ -0,0 +1,167 @@
|
||||
---
|
||||
typora-copy-images-to: images
|
||||
comments: true
|
||||
---
|
||||
|
||||
# 表格识别
|
||||
|
||||
## 1. 表格识别 pipeline
|
||||
|
||||
表格识别主要包含三个模型
|
||||
|
||||
1. 单行文本检测-DB
|
||||
2. 单行文本识别-CRNN
|
||||
3. 表格结构和cell坐标预测-SLANet
|
||||
|
||||
具体流程图如下
|
||||
|
||||
|
||||
|
||||
流程说明:
|
||||
|
||||
1. 图片由单行文字检测模型检测到单行文字的坐标，然后送入识别模型拿到识别结果。
|
||||
2. 图片由SLANet模型拿到表格的结构信息和单元格的坐标信息。
|
||||
3. 由单行文字的坐标、识别结果和单元格的坐标一起组合出单元格的识别结果。
|
||||
4. 单元格的识别结果和表格结构一起构造表格的html字符串。
|
||||
|
||||
## 2. 性能
|
||||
|
||||
我们在 PubTabNet<sup>[1]</sup> 评估数据集上对算法进行了评估，性能如下
|
||||
|
||||
| 算法 | Acc | [TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src) | Speed |
|
||||
| ------------------- | ------ | ----------------------------------------------------------------------------------------------------- | ----- |
|
||||
| EDD<sup>[2]</sup> | x | 88.30% | x |
|
||||
| TableRec-RARE(ours) | 71.73% | 93.88% | 779ms |
|
||||
| SLANet(ours) | 76.31% | 95.89% | 766ms |
|
||||
|
||||
性能指标解释如下：
|
||||
|
||||
- Acc: 模型对每张图像里表格结构的识别准确率，错一个token就算错误。
|
||||
- TEDS: 模型对表格信息还原的准确度，此指标评价内容不仅包含表格结构，还包含表格内的文字内容。
|
||||
- Speed: 模型在CPU机器上，开启MKL的情况下，单张图片的推理速度。
|
||||
|
||||
## 3. 效果演示
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
## 4. 使用
|
||||
|
||||
### 4.1 快速开始
|
||||
|
||||
PP-Structure目前提供了中英文两种语言的表格识别模型，模型链接见 [models_list](../models_list.md)。也提供了whl包的形式方便快速使用，详见 [quickstart](../quick_start.md)。
|
||||
|
||||
下面以中文表格识别模型为例，介绍如何识别一张表格。
|
||||
|
||||
使用如下命令即可快速完成一张表格的识别。
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleOCR/ppstructure
|
||||
|
||||
# 下载模型
|
||||
mkdir inference && cd inference
|
||||
# 下载PP-OCRv3文本检测模型并解压
|
||||
wget https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv3_mobile_det_infer.tar && tar xf PP-OCRv3_mobile_det_infer.tar
|
||||
# 下载PP-OCRv3文本识别模型并解压
|
||||
wget https://paddle-model-ecology.bj.bcebos.com/paddlex/official_inference_model/paddle3.0.0/PP-OCRv3_mobile_rec_infer.tar && tar xf PP-OCRv3_mobile_rec_infer.tar
|
||||
# 下载PP-StructureV2中文表格识别模型并解压
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/paddle3.0b2/ch_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf ch_ppstructure_mobile_v2.0_SLANet_infer.tar
|
||||
cd ..
|
||||
# 执行表格识别
|
||||
python table/predict_table.py \
|
||||
--det_model_dir=inference/PP-OCRv3_mobile_det_infer \
|
||||
--rec_model_dir=inference/PP-OCRv3_mobile_rec_infer \
|
||||
--table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \
|
||||
--rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \
|
||||
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \
|
||||
--image_dir=docs/table/table.jpg \
|
||||
--output=../output/table
|
||||
```
|
||||
|
||||
运行完成后，每张图片的excel表格会保存到output字段指定的目录下，同时在该目录下回生产一个html文件，用于可视化查看单元格坐标和识别的表格。
|
||||
|
||||
**NOTE**
|
||||
|
||||
1. 如果想使用英文模型，需要在 [models_list](../../ppocr/model_list.md) 中下载英文文字检测识别模型和英文表格识别模型，同时替换`table_structure_dict_ch.txt`为`table_structure_dict.txt`即可。
|
||||
2. 如需使用TableRec-RARE模型，需要替换`table_structure_dict_ch.txt`为`table_structure_dict.txt`，同时参数`--merge_no_span_structure=False`
|
||||
|
||||
### 4.2 模型训练、评估与推理
|
||||
|
||||
文本检测模型的训练、评估和推理流程可参考 [detection](../../ppocr/model_train/detection.md)
|
||||
|
||||
文本识别模型的训练、评估和推理流程可参考 [recognition](../../ppocr/model_train/recognition.md)
|
||||
|
||||
表格识别模型的训练、评估和推理流程可参考 [table_recognition](../../ppstructure/model_train/train_table.md)
|
||||
|
||||
### 4.3 计算TEDS
|
||||
|
||||
表格使用 [TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src) 作为模型的评估指标。在进行模型评估之前，需要将pipeline中的三个模型分别导出为inference模型(我们已经提供好)，还需要准备评估的gt， gt示例如下:
|
||||
|
||||
```txt
|
||||
PMC5755158_010_01.png <html><body><table><thead><tr><td></td><td><b>Weaning</b></td><td><b>Week 15</b></td><td><b>Off-test</b></td></tr></thead><tbody><tr><td>Weaning</td><td>–</td><td>–</td><td>–</td></tr><tr><td>Week 15</td><td>–</td><td>0.17 ± 0.08</td><td>0.16 ± 0.03</td></tr><tr><td>Off-test</td><td>–</td><td>0.80 ± 0.24</td><td>0.19 ± 0.09</td></tr></tbody></table></body></html>
|
||||
```
|
||||
|
||||
gt每一行都由文件名和表格的html字符串组成，文件名和表格的html字符串之间使用`\t`分隔。
|
||||
|
||||
也可使用如下命令，由标注文件生成评估的gt文件：
|
||||
|
||||
```bash linenums="1"
|
||||
python3 ppstructure/table/convert_label2html.py --ori_gt_path /path/to/your_label_file --save_path /path/to/save_file
|
||||
```
|
||||
|
||||
准备完成后使用如下命令进行评估，评估完成后会输出teds指标。
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleOCR/ppstructure
|
||||
python3 table/eval_table.py \
|
||||
--det_model_dir=path/to/det_model_dir \
|
||||
--rec_model_dir=path/to/rec_model_dir \
|
||||
--table_model_dir=path/to/table_model_dir \
|
||||
--image_dir=docs/table/table.jpg \
|
||||
--rec_char_dict_path=../ppocr/utils/dict/table_dict.txt \
|
||||
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \
|
||||
--det_limit_side_len=736 \
|
||||
--det_limit_type=min \
|
||||
--gt_path=path/to/gt.txt
|
||||
```
|
||||
|
||||
如使用英文表格识别模型在PubLatNet数据集上进行评估
|
||||
|
||||
```bash linenums="1"
|
||||
cd PaddleOCR/ppstructure
|
||||
# 下载模型
|
||||
mkdir inference && cd inference
|
||||
# 下载基于PubTabNet数据集训练的文本检测模型并解压
|
||||
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar && tar xf en_ppocr_mobile_v2.0_table_det_infer.tar
|
||||
# 下载基于PubTabNet数据集训练的文本识别模型并解压
|
||||
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar && tar xf en_ppocr_mobile_v2.0_table_rec_infer.tar
|
||||
# 下载基于PubTabNet数据集训练的表格识别模型并解压
|
||||
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/paddle3.0b2/en_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf en_ppstructure_mobile_v2.0_SLANet_infer.tar
|
||||
cd ..
|
||||
|
||||
python3 table/eval_table.py \
|
||||
--det_model_dir=inference/en_ppocr_mobile_v2.0_table_det_infer \
|
||||
--rec_model_dir=inference/en_ppocr_mobile_v2.0_table_rec_infer \
|
||||
--table_model_dir=inference/en_ppstructure_mobile_v2.0_SLANet_infer \
|
||||
--image_dir=train_data/table/pubtabnet/val/ \
|
||||
--rec_char_dict_path=../ppocr/utils/dict/table_dict.txt \
|
||||
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \
|
||||
--det_limit_side_len=736 \
|
||||
--det_limit_type=min \
|
||||
--rec_image_shape=3,32,320 \
|
||||
--gt_path=path/to/gt.txt
|
||||
```
|
||||
|
||||
将会输出
|
||||
|
||||
```bash linenums="1"
|
||||
teds: 95.89
|
||||
```
|
||||
|
||||
## 5. Reference
|
||||
|
||||
1. <https://github.com/ibm-aur-nlp/PubTabNet>
|
||||
2. <https://arxiv.org/pdf/1911.10683>
|
||||
132
docs/version2.x/ppstructure/model_train/training.en.md
Normal file
132
docs/version2.x/ppstructure/model_train/training.en.md
Normal file
@@ -0,0 +1,132 @@
|
||||
---
|
||||
comments: true
|
||||
---
|
||||
|
||||
# Model Training
|
||||
|
||||
This article will introduce the basic concepts that is necessary for model training and tuning.
|
||||
|
||||
At the same time, it will briefly introduce the structure of the training data and how to prepare the data to fine-tune model in vertical scenes.
|
||||
|
||||
## 1. Yml Configuration
|
||||
|
||||
The PaddleOCR uses configuration files to control network training and evaluation parameters. In the configuration file, you can set the model, optimizer, loss function, and pre- and post-processing parameters of the model. PaddleOCR reads these parameters from the configuration file, and then builds a complete training process to train the model. Fine-tuning can also be completed by modifying the parameters in the configuration file, which is simple and convenient.
|
||||
|
||||
For the complete configuration file description, please refer to [Configuration File](../../ppocr/blog/config.en.md)
|
||||
|
||||
## 2. Basic Concepts
|
||||
|
||||
During the model training process, some hyper-parameters can be manually specified to obtain the optimal result at the least cost. Different data volumes may require different hyper-parameters. When you want to fine-tune the model based on your own data, there are several parameter adjustment strategies for reference:
|
||||
|
||||
### 2.1 Learning Rate
|
||||
|
||||
The learning rate is one of the most important hyper-parameters for training neural networks. It represents the step length of the gradient moving towards the optimal solution of the loss function in each iteration.
|
||||
A variety of learning rate update strategies are provided by PaddleOCR, which can be specified in configuration files. For example,
|
||||
|
||||
```yaml linenums="1"
|
||||
Optimizer:
|
||||
...
|
||||
lr:
|
||||
name: Piecewise
|
||||
decay_epochs : [700, 800]
|
||||
values : [0.001, 0.0001]
|
||||
warmup_epoch: 5
|
||||
```
|
||||
|
||||
`Piecewise` stands for piece-wise constant attenuation. Different learning rates are specified in different learning stages, and the learning rate stay the same in each stage.
|
||||
|
||||
`warmup_epoch` means that in the first 5 epochs, the learning rate will be increased gradually from 0 to base_lr. For all strategies, please refer to the code [learning_rate.py](../../ppocr/optimizer/learning_rate.py).
|
||||
|
||||
### 2.2 Regularization
|
||||
|
||||
Regularization can effectively avoid algorithm over-fitting. PaddleOCR provides L1 and L2 regularization methods.
|
||||
L1 and L2 regularization are the most widely used regularization methods.
|
||||
L1 regularization adds a regularization term to the objective function to reduce the sum of absolute values of the parameters;
|
||||
while in L2 regularization, the purpose of adding a regularization term is to reduce the sum of squared parameters.
|
||||
The configuration method is as follows:
|
||||
|
||||
```yaml linenums="1"
|
||||
Optimizer:
|
||||
...
|
||||
regularizer:
|
||||
name: L2
|
||||
factor: 2.0e-05
|
||||
```
|
||||
|
||||
### 2.3 Evaluation Indicators
|
||||
|
||||
(1) Detection stage: First, evaluate according to the IOU of the detection frame and the labeled frame. If the IOU is greater than a certain threshold, it is judged that the detection is accurate. Here, the detection frame and the label frame are different from the general general target detection frame, and they are represented by polygons. Detection accuracy: the percentage of the correct detection frame number in all detection frames is mainly used to judge the detection index. Detection recall rate: the percentage of correct detection frames in all marked frames, which is mainly an indicator of missed detection.
|
||||
|
||||
(2) Recognition stage: Character recognition accuracy, that is, the ratio of correctly recognized text lines to the number of marked text lines. Only the entire line of text recognition pairs can be regarded as correct recognition.
|
||||
|
||||
(3) End-to-end statistics: End-to-end recall rate: accurately detect and correctly identify the proportion of text lines in all labeled text lines; End-to-end accuracy rate: accurately detect and correctly identify the number of text lines in the detected text lines The standard for accurate detection is that the IOU of the detection box and the labeled box is greater than a certain threshold, and the text in the correctly identified detection box is the same as the labeled text.
|
||||
|
||||
## 3. Data and Vertical Scenes
|
||||
|
||||
### 3.1 Training Data
|
||||
|
||||
The current open source models, data sets and magnitudes are as follows:
|
||||
|
||||
- Detection:
|
||||
- English data set, ICDAR2015
|
||||
- Chinese data set, LSVT street view data set training data 3w pictures
|
||||
|
||||
- Identification:
|
||||
- English data set, MJSynth and SynthText synthetic data, the data volume is tens of millions.
|
||||
- Chinese data set, LSVT street view data set crops the image according to the truth value, and performs position calibration, a total of 30w images. In addition, based on the LSVT corpus, 500w of synthesized data.
|
||||
- Small language data set, using different corpora and fonts, respectively generated 100w synthetic data set, and using ICDAR-MLT as the verification set.
|
||||
|
||||
Among them, the public data sets are all open source, users can search and download by themselves, or refer to [Chinese data set](../../datasets/datasets.en.md), synthetic data is not open source, users can use open source synthesis tools to synthesize by themselves. Synthesis tools include [text_renderer](https://github.com/Sanster/text_renderer), [SynthText](https://github.com/ankush-me/SynthText), [TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) etc.
|
||||
|
||||
### 3.2 Vertical Scene
|
||||
|
||||
PaddleOCR mainly focuses on general OCR. If you have vertical requirements, you can use PaddleOCR + vertical data to train yourself;
|
||||
If there is a lack of labeled data, or if you do not want to invest in research and development costs, it is recommended to directly call the open API, which covers some of the more common vertical categories.
|
||||
|
||||
### 3.3 Build Your Own Dataset
|
||||
|
||||
There are several experiences for reference when constructing the data set:
|
||||
|
||||
(1) The amount of data in the training set:
|
||||
|
||||
a. The data required for detection is relatively small. For Fine-tune based on the PaddleOCR model, 500 sheets are generally required to achieve good results.
|
||||
|
||||
b. Recognition is divided into English and Chinese. Generally, English scenarios require hundreds of thousands of data to achieve good results, while Chinese requires several million or more.
|
||||
|
||||
(2) When the amount of training data is small, you can try the following three ways to get more data:
|
||||
|
||||
a. Manually collect more training data, the most direct and effective way.
|
||||
|
||||
b. Basic image processing or transformation based on PIL and opencv. For example, the three modules of ImageFont, Image, ImageDraw in PIL write text into the background, opencv's rotating affine transformation, Gaussian filtering and so on.
|
||||
|
||||
c. Use data generation algorithms to synthesize data, such as algorithms such as pix2pix.
|
||||
|
||||
## 4. FAQ
|
||||
|
||||
**Q**: How to choose a suitable network input shape when training CRNN recognition?
|
||||
|
||||
A: The general height is 32, the longest width is selected, there are two methods:
|
||||
|
||||
(1) Calculate the aspect ratio distribution of training sample images. The selection of the maximum aspect ratio considers 80% of the training samples.
|
||||
|
||||
(2) Count the number of texts in training samples. The selection of the longest number of characters considers the training sample that satisfies 80%. Then the aspect ratio of Chinese characters is approximately considered to be 1, and that of English is 3:1, and the longest width is estimated.
|
||||
|
||||
**Q**: During the recognition training, the accuracy of the training set has reached 90, but the accuracy of the verification set has been kept at 70, what should I do?
|
||||
|
||||
A: If the accuracy of the training set is 90 and the test set is more than 70, it should be over-fitting. There are two methods to try:
|
||||
|
||||
(1) Add more augmentation methods or increase the [probability] of augmented prob (https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/ppocr/data/imaug/rec_img_aug.py#L341), The default is 0.4.
|
||||
|
||||
(2) Increase the [l2 dcay value] of the system (https://github.com/PaddlePaddle/PaddleOCR/blob/a501603d54ff5513fc4fc760319472e59da25424/configs/rec/ch_ppocr_v1.1/rec_chinese_lite_train_v1.1.yml#L47)
|
||||
|
||||
**Q**: When the recognition model is trained, loss can drop normally, but acc is always 0
|
||||
|
||||
A: It is normal for the acc to be 0 at the beginning of the recognition model training, and the indicator will come up after a longer training period.
|
||||
|
||||
***
|
||||
|
||||
Click the following links for detailed training tutorial:
|
||||
|
||||
- [text detection model training](./detection.en.md)
|
||||
- [text recognition model training](./recognition.en.md)
|
||||
- [text direction classification model training](./angle_class.en.md)
|
||||
128
docs/version2.x/ppstructure/model_train/training.md
Normal file
128
docs/version2.x/ppstructure/model_train/training.md
Normal file
@@ -0,0 +1,128 @@
|
||||
---
|
||||
comments: true
|
||||
---
|
||||
|
||||
# PP-OCR模型训练
|
||||
|
||||
本文将介绍模型训练时需掌握的基本概念，和训练时的调优方法。
|
||||
|
||||
同时会简单介绍PaddleOCR模型训练数据的组成部分，以及如何在垂类场景中准备数据finetune模型。
|
||||
|
||||
## 1. 配置文件说明
|
||||
|
||||
PaddleOCR模型使用配置文件管理网络训练、评估的参数。在配置文件中，可以设置组建模型、优化器、损失函数、模型前后处理的参数，PaddleOCR从配置文件中读取到这些参数，进而组建出完整的训练流程，完成模型训练，在需要对模型进行优化的时，可以通过修改配置文件中的参数完成配置，使用简单且方便修改。
|
||||
|
||||
完整的配置文件说明可以参考[配置文件](../../ppocr/blog/config.md)
|
||||
|
||||
## 2. 基本概念
|
||||
|
||||
模型训练过程中需要手动调整一些超参数，帮助模型以最小的代价获得最优指标。不同的数据量可能需要不同的超参，当您希望在自己的数据上finetune或对模型效果调优时，有以下几个参数调整策略可供参考：
|
||||
|
||||
### 2.1 学习率
|
||||
|
||||
学习率是训练神经网络的重要超参数之一，它代表在每一次迭代中梯度向损失函数最优解移动的步长。
|
||||
在PaddleOCR中提供了多种学习率更新策略,可以通过配置文件修改，例如：
|
||||
|
||||
```yaml linenums="1"
|
||||
Optimizer:
|
||||
...
|
||||
lr:
|
||||
name: Piecewise
|
||||
decay_epochs : [700, 800]
|
||||
values : [0.001, 0.0001]
|
||||
warmup_epoch: 5
|
||||
```
|
||||
|
||||
Piecewise 代表分段常数衰减，在不同的学习阶段指定不同的学习率，在每段内学习率相同。
|
||||
warmup_epoch 代表在前5个epoch中，学习率将逐渐从0增加到base_lr。全部策略可以参考代码[learning_rate.py](../../ppocr/optimizer/learning_rate.py) 。
|
||||
|
||||
### 2.2 正则化
|
||||
|
||||
正则化可以有效的避免算法过拟合，PaddleOCR中提供了L1、L2正则方法，L1 和 L2 正则化是最常用的正则化方法。L1 正则化向目标函数添加正则化项，以减少参数的绝对值总和；而 L2 正则化中，添加正则化项的目的在于减少参数平方的总和。配置方法如下：
|
||||
|
||||
```yaml linenums="1"
|
||||
Optimizer:
|
||||
...
|
||||
regularizer:
|
||||
name: L2
|
||||
factor: 2.0e-05
|
||||
```
|
||||
|
||||
### 2.3 评估指标
|
||||
|
||||
（1）检测阶段：先按照检测框和标注框的IOU评估，IOU大于某个阈值判断为检测准确。这里检测框和标注框不同于一般的通用目标检测框，是采用多边形进行表示。检测准确率：正确的检测框个数在全部检测框的占比，主要是判断检测指标。检测召回率：正确的检测框个数在全部标注框的占比，主要是判断漏检的指标。
|
||||
|
||||
（2）识别阶段： 字符识别准确率，即正确识别的文本行占标注的文本行数量的比例，只有整行文本识别对才算正确识别。
|
||||
|
||||
（3）端到端统计： 端对端召回率：准确检测并正确识别文本行在全部标注文本行的占比； 端到端准确率：准确检测并正确识别文本行在 检测到的文本行数量 的占比； 准确检测的标准是检测框与标注框的IOU大于某个阈值，正确识别的检测框中的文本与标注的文本相同。
|
||||
|
||||
## 3. 数据与垂类场景
|
||||
|
||||
### 3.1 训练数据
|
||||
|
||||
目前开源的模型，数据集和量级如下：
|
||||
|
||||
- 检测：
|
||||
- 英文数据集，ICDAR2015
|
||||
- 中文数据集，LSVT街景数据集训练数据3w张图片
|
||||
|
||||
- 识别：
|
||||
- 英文数据集，MJSynth和SynthText合成数据，数据量上千万。
|
||||
- 中文数据集，LSVT街景数据集根据真值将图crop出来，并进行位置校准，总共30w张图像。此外基于LSVT的语料，合成数据500w。
|
||||
- 小语种数据集，使用不同语料和字体，分别生成了100w合成数据集，并使用ICDAR-MLT作为验证集。
|
||||
|
||||
其中，公开数据集都是开源的，用户可自行搜索下载，也可参考[中文数据集](../../datasets/datasets.md)，合成数据暂不开源，用户可使用开源合成工具自行合成，可参考的合成工具包括[text_renderer](https://github.com/Sanster/text_renderer) 、[SynthText](https://github.com/ankush-me/SynthText) 、[TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) 等。
|
||||
|
||||
### 3.2 垂类场景
|
||||
|
||||
PaddleOCR主要聚焦通用OCR，如果有垂类需求，您可以用PaddleOCR+垂类数据自己训练；
|
||||
如果缺少带标注的数据，或者不想投入研发成本，建议直接调用开放的API，开放的API覆盖了目前比较常见的一些垂类。
|
||||
|
||||
### 3.3 自己构建数据集
|
||||
|
||||
在构建数据集时有几个经验可供参考：
|
||||
|
||||
（1） 训练集的数据量：
|
||||
|
||||
a. 检测需要的数据相对较少，在PaddleOCR模型的基础上进行Fine-tune，一般需要500张可达到不错的效果。
|
||||
|
||||
b. 识别分英文和中文，一般英文场景需要几十万数据可达到不错的效果，中文则需要几百万甚至更多。
|
||||
|
||||
（2）当训练数据量少时，可以尝试以下三种方式获取更多的数据：
|
||||
|
||||
a. 人工采集更多的训练数据，最直接也是最有效的方式。
|
||||
|
||||
b. 基于PIL和opencv基本图像处理或者变换。例如PIL中ImageFont, Image, ImageDraw三个模块将文字写到背景中，opencv的旋转仿射变换，高斯滤波等。
|
||||
|
||||
c. 利用数据生成算法合成数据，例如pix2pix或[StyleText](https://github.com/PFCCLab/StyleText)等算法。
|
||||
|
||||
## 4. 常见问题
|
||||
|
||||
**Q**：训练CRNN识别时，如何选择合适的网络输入shape？
|
||||
|
||||
A：一般高度采用32，最长宽度的选择，有两种方法：
|
||||
|
||||
（1）统计训练样本图像的宽高比分布。最大宽高比的选取考虑满足80%的训练样本。
|
||||
|
||||
（2）统计训练样本文字数目。最长字符数目的选取考虑满足80%的训练样本。然后中文字符长宽比近似认为是1，英文认为3：1，预估一个最长宽度。
|
||||
|
||||
**Q**：识别训练时，训练集精度已经到达90了，但验证集精度一直在70，涨不上去怎么办？
|
||||
|
||||
A：训练集精度90，测试集70多的话，应该是过拟合了，有两个可尝试的方法：
|
||||
|
||||
（1）加入更多的增广方式或者调大增广prob的[概率](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/ppocr/data/imaug/rec_img_aug.py#L341)，默认为0.4。
|
||||
|
||||
（2）调大系统的[l2 dcay值](https://github.com/PaddlePaddle/PaddleOCR/blob/a501603d54ff5513fc4fc760319472e59da25424/configs/rec/ch_ppocr_v1.1/rec_chinese_lite_train_v1.1.yml#L47)
|
||||
|
||||
**Q**: 识别模型训练时，loss能正常下降，但acc一直为0
|
||||
|
||||
A：识别模型训练初期acc为0是正常的，多训一段时间指标就上来了。
|
||||
|
||||
***
|
||||
|
||||
具体的训练教程可点击下方链接跳转：
|
||||
|
||||
- [文本检测模型训练](./detection.md)
|
||||
- [文本识别模型训练](./recognition.md)
|
||||
- [文本方向分类器训练](./angle_class.md)
|
||||
- [知识蒸馏](../model_compress/knowledge_distillation.md)
|
||||
Reference in New Issue
Block a user