README.md 8.3 KB
Newer Older
WenmuZhou's avatar
add re  
WenmuZhou committed
1
# 文档视觉问答(DOC-VQA)
littletomatodonkey's avatar
littletomatodonkey committed
2

WenmuZhou's avatar
WenmuZhou committed
3
VQA指视觉问答,主要针对图像内容进行提问和回答,DOC-VQA是VQA任务中的一种,DOC-VQA主要针对文本图像的文字内容提出问题。
WenmuZhou's avatar
add re  
WenmuZhou committed
4
5
6
7

PP-Structure 里的 DOC-VQA算法基于PaddleNLP自然语言处理算法库进行开发。

主要特性如下:
littletomatodonkey's avatar
littletomatodonkey committed
8
9

- 集成[LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf)模型以及PP-OCR预测引擎。
WenmuZhou's avatar
WenmuZhou committed
10
11
- 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。
WenmuZhou's avatar
add re  
WenmuZhou committed
12
13
- 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。
littletomatodonkey's avatar
littletomatodonkey committed
14
15
16
17
18


本项目是 [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/pdf/2104.08836.pdf) 在 Paddle 2.2上的开源实现,
包含了在 [XFUND数据集](https://github.com/doc-analysis/XFUND) 上的微调代码。

WenmuZhou's avatar
add re  
WenmuZhou committed
19
20
## 1 性能

WenmuZhou's avatar
WenmuZhou committed
21
我们在 [XFUN](https://github.com/doc-analysis/XFUND) 的中文数据集上对算法进行了评估,性能如下
WenmuZhou's avatar
add re  
WenmuZhou committed
22

23
| 模型 | 任务 | hmean | 模型下载地址 |
WenmuZhou's avatar
WenmuZhou committed
24
|:---:|:---:|:---:| :---:|
25
26
27
| LayoutXLM | RE | 0.7483 | [链接](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
| LayoutXLM | SER | 0.9038 | [链接](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
| LayoutLM | SER | 0.7731 | [链接](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) |
WenmuZhou's avatar
add re  
WenmuZhou committed
28
29
30
31



## 2. 效果演示
littletomatodonkey's avatar
littletomatodonkey committed
32
33
34

**注意:** 测试图片来源于XFUN数据集。

WenmuZhou's avatar
add re  
WenmuZhou committed
35
### 2.1 SER
littletomatodonkey's avatar
littletomatodonkey committed
36

WenmuZhou's avatar
WenmuZhou committed
37
![](../../doc/vqa/result_ser/zh_val_0_ser.jpg) | ![](../../doc/vqa/result_ser/zh_val_42_ser.jpg)
WenmuZhou's avatar
add re  
WenmuZhou committed
38
---|---
littletomatodonkey's avatar
littletomatodonkey committed
39

WenmuZhou's avatar
add re  
WenmuZhou committed
40
图中不同颜色的框表示不同的类别,对于XFUN数据集,有`QUESTION`, `ANSWER`, `HEADER` 3种类别
littletomatodonkey's avatar
littletomatodonkey committed
41

WenmuZhou's avatar
add re  
WenmuZhou committed
42
43
44
* 深紫色:HEADER
* 浅紫色:QUESTION
* 军绿色:ANSWER
littletomatodonkey's avatar
littletomatodonkey committed
45

WenmuZhou's avatar
add re  
WenmuZhou committed
46
在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
littletomatodonkey's avatar
littletomatodonkey committed
47
48


WenmuZhou's avatar
add re  
WenmuZhou committed
49
### 2.2 RE
littletomatodonkey's avatar
littletomatodonkey committed
50

WenmuZhou's avatar
WenmuZhou committed
51
![](../../doc/vqa/result_re/zh_val_21_re.jpg) | ![](../../doc/vqa/result_re/zh_val_40_re.jpg)
WenmuZhou's avatar
add re  
WenmuZhou committed
52
---|---
littletomatodonkey's avatar
littletomatodonkey committed
53
54


WenmuZhou's avatar
add re  
WenmuZhou committed
55
图中红色框表示问题,蓝色框表示答案,问题和答案之间使用绿色线连接。在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
littletomatodonkey's avatar
littletomatodonkey committed
56

WenmuZhou's avatar
add re  
WenmuZhou committed
57
58
59
60

## 3. 安装

### 3.1 安装依赖
littletomatodonkey's avatar
littletomatodonkey committed
61
62
63
64
65
66
67

- **(1) 安装PaddlePaddle**

```bash
pip3 install --upgrade pip

# GPU安装
68
python3 -m pip install "paddlepaddle-gpu>=2.2" -i https://mirror.baidu.com/pypi/simple
littletomatodonkey's avatar
littletomatodonkey committed
69
70

# CPU安装
71
python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simple
littletomatodonkey's avatar
littletomatodonkey committed
72
73
74
75
76

```
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。


WenmuZhou's avatar
add re  
WenmuZhou committed
77
### 3.2 安装PaddleOCR(包含 PP-OCR 和 VQA )
littletomatodonkey's avatar
littletomatodonkey committed
78
79
80
81

- **(1)pip快速安装PaddleOCR whl包(仅预测)**

```bash
82
python3 -m pip install paddleocr
littletomatodonkey's avatar
littletomatodonkey committed
83
84
```

littletomatodonkey's avatar
littletomatodonkey committed
85
- **(2)下载VQA源码(预测+训练)**
littletomatodonkey's avatar
littletomatodonkey committed
86
87
88
89
90
91
92
93
94
95

```bash
【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。
```

96
- **(3)安装VQA的`requirements`**
littletomatodonkey's avatar
littletomatodonkey committed
97
98

```bash
99
python3 -m pip install -r ppstructure/vqa/requirements.txt
littletomatodonkey's avatar
littletomatodonkey committed
100
101
```

WenmuZhou's avatar
add re  
WenmuZhou committed
102
## 4. 使用
littletomatodonkey's avatar
littletomatodonkey committed
103
104


WenmuZhou's avatar
add re  
WenmuZhou committed
105
### 4.1 数据和预训练模型准备
littletomatodonkey's avatar
littletomatodonkey committed
106

107
108
109
110
如果希望直接体验预测过程,可以下载我们提供的预训练模型,跳过训练过程,直接预测即可。

* 下载处理好的数据集

littletomatodonkey's avatar
littletomatodonkey committed
111
112
113
114
115
116
117
118
119
处理好的XFUN中文数据集下载地址:[https://paddleocr.bj.bcebos.com/dataset/XFUND.tar](https://paddleocr.bj.bcebos.com/dataset/XFUND.tar)


下载并解压该数据集,解压后将数据集放置在当前目录下。

```shell
wget https://paddleocr.bj.bcebos.com/dataset/XFUND.tar
```

120
* 转换数据集
littletomatodonkey's avatar
littletomatodonkey committed
121

122
若需进行其他XFUN数据集的训练,可使用下面的命令进行数据集的转换
littletomatodonkey's avatar
littletomatodonkey committed
123

124
125
126
```bash
python3 ppstructure/vqa/helper/trans_xfun_data.py --ori_gt_path=path/to/json_path --output_path=path/to/save_path
```
littletomatodonkey's avatar
littletomatodonkey committed
127

WenmuZhou's avatar
add re  
WenmuZhou committed
128
### 4.2 SER任务
littletomatodonkey's avatar
littletomatodonkey committed
129

130
131
132
133
134
135
启动训练之前,需要修改下面的四个字段

1. `Train.dataset.data_dir`:指向训练集图片存放目录
2. `Train.dataset.label_file_list`:指向训练集标注文件
3. `Eval.dataset.data_dir`:指指向验证集图片存放目录
4. `Eval.dataset.label_file_list`:指向验证集标注文件
littletomatodonkey's avatar
littletomatodonkey committed
136

137
* 启动训练
littletomatodonkey's avatar
littletomatodonkey committed
138
```shell
139
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/ser/layoutxlm.yml
littletomatodonkey's avatar
littletomatodonkey committed
140
141
```

142
143
最终会打印出`precision`, `recall`, `hmean`等指标。
`./output/ser_layoutxlm/`文件夹中会保存训练日志,最优的模型和最新epoch的模型。
littletomatodonkey's avatar
littletomatodonkey committed
144

zhoujun's avatar
zhoujun committed
145
146
* 恢复训练

147
148
恢复训练需要将之前训练好的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段。

zhoujun's avatar
zhoujun committed
149
```shell
150
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
zhoujun's avatar
zhoujun committed
151
152
153
154
```

* 评估

155
评估需要将待评估的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段。
littletomatodonkey's avatar
littletomatodonkey committed
156
157

```shell
158
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
littletomatodonkey's avatar
littletomatodonkey committed
159
```
160
最终会打印出`precision`, `recall`, `hmean`等指标
littletomatodonkey's avatar
littletomatodonkey committed
161

162
* 使用`OCR引擎 + SER`串联预测
littletomatodonkey's avatar
littletomatodonkey committed
163

164
使用如下命令即可完成`OCR引擎 + SER`的串联预测
littletomatodonkey's avatar
littletomatodonkey committed
165
166

```shell
WenmuZhou's avatar
WenmuZhou committed
167
CUDA_VISIBLE_DEVICES=0 python3 tools/infer_vqa_token_ser.py -c configs/vqa/ser/layoutxlm.yml  -o Architecture.Backbone.checkpoints=PP-Layout_v1.0_ser_pretrained/ Global.infer_img=doc/vqa/input/zh_val_42.jpg
littletomatodonkey's avatar
littletomatodonkey committed
168
169
```

170
171
最终会在`config.Global.save_res_path`字段所配置的目录下保存预测结果可视化图像以及预测结果文本文件,预测结果文本文件名为`infer_results.txt`

littletomatodonkey's avatar
littletomatodonkey committed
172
173
*`OCR引擎 + SER`预测系统进行端到端评估

174
175
首先使用 `tools/infer_vqa_token_ser.py` 脚本完成数据集的预测,然后使用下面的命令进行评估。

littletomatodonkey's avatar
littletomatodonkey committed
176
177
```shell
export CUDA_VISIBLE_DEVICES=0
WenmuZhou's avatar
add re  
WenmuZhou committed
178
python3.7 helper/eval_with_label_end2end.py --gt_json_path XFUND/zh_val/xfun_normalize_val.json  --pred_json_path output_res/infer_results.txt
littletomatodonkey's avatar
littletomatodonkey committed
179
180
181
```


WenmuZhou's avatar
add re  
WenmuZhou committed
182
### 3.3 RE任务
littletomatodonkey's avatar
littletomatodonkey committed
183

WenmuZhou's avatar
add re  
WenmuZhou committed
184
* 启动训练
littletomatodonkey's avatar
littletomatodonkey committed
185

186
启动训练之前,需要修改下面的四个字段
WenmuZhou's avatar
add re  
WenmuZhou committed
187

188
189
190
191
1. `Train.dataset.data_dir`:指向训练集图片存放目录
2. `Train.dataset.label_file_list`:指向训练集标注文件
3. `Eval.dataset.data_dir`:指指向验证集图片存放目录
4. `Eval.dataset.label_file_list`:指向验证集标注文件
zhoujun's avatar
zhoujun committed
192
193

```shell
194
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml
zhoujun's avatar
zhoujun committed
195
196
```

197
198
199
200
201
202
最终会打印出`precision`, `recall`, `hmean`等指标。
`./output/re_layoutxlm/`文件夹中会保存训练日志,最优的模型和最新epoch的模型。

* 恢复训练

恢复训练需要将之前训练好的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段。
WenmuZhou's avatar
add re  
WenmuZhou committed
203

zhoujun's avatar
zhoujun committed
204
```shell
205
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
zhoujun's avatar
zhoujun committed
206
207
```

208
* 评估
zhoujun's avatar
zhoujun committed
209

210
评估需要将待评估的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段。
WenmuZhou's avatar
add re  
WenmuZhou committed
211
212

```shell
213
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
WenmuZhou's avatar
add re  
WenmuZhou committed
214
```
215
最终会打印出`precision`, `recall`, `hmean`等指标
WenmuZhou's avatar
add re  
WenmuZhou committed
216

217
* 使用`OCR引擎 + SER + RE`串联预测
WenmuZhou's avatar
add re  
WenmuZhou committed
218

219
使用如下命令即可完成`OCR引擎 + SER + RE`的串联预测
WenmuZhou's avatar
add re  
WenmuZhou committed
220
221
```shell
export CUDA_VISIBLE_DEVICES=0
WenmuZhou's avatar
WenmuZhou committed
222
python3 tools/infer_vqa_token_ser_re.py -c configs/vqa/re/layoutxlm.yml  -o Architecture.Backbone.checkpoints=PP-Layout_v1.0_re_pretrained/ Global.infer_img=doc/vqa/input/zh_val_21.jpg -c_ser configs/vqa/ser/layoutxlm.yml  -o_ser Architecture.Backbone.checkpoints=PP-Layout_v1.0_ser_pretrained/
WenmuZhou's avatar
add re  
WenmuZhou committed
223
```
littletomatodonkey's avatar
littletomatodonkey committed
224

225
226
227
最终会在`config.Global.save_res_path`字段所配置的目录下保存预测结果可视化图像以及预测结果文本文件,预测结果文本文件名为`infer_results.txt`


littletomatodonkey's avatar
littletomatodonkey committed
228
229
230
231
232
## 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