PSENet-Paddle
基于Paddle框架的PSENet复现
本项目基于paddlepaddle框架复现PSENet，并参加百度第三届论文复现赛，将在2021年5月15日比赛完后提供AIStudio链接～敬请期待
参考项目：
环境配置
本项目利用AIstudio平台，采用paddlepaddle: 2.0.2-gpu Version，除此之外你需要通过pip install mmcv editdistance Polygon3 pyclipper或者pip install -r requirement.txt来安装依赖包
数据集
本项目已搭载PSENet比赛指定数据集，你可以在此找到搭载的数据集，包含ICDAR2015 Task4以及Total-Text
工程目录
注意到你需要将submitPSENet重命名为PSENet
/home/aistudio/PSENet
|───data(解压的data.zip)
└───config
└───models
└───dataset
└───eval
└───utils
└───compile.sh
└───__init__.py
└───test.py
└───train.py
└───requirement.txt
└───logo.gif
项目配置**
注意：由于aistudio的docker环境并不适配本项目的编译，所以你需要在本地计算机编译完成后上传编译文件，在本地计算机我才用如下配置，你可以使用gcc --version和g++ --version查看配置
| AIStudio | Local PC |
|---|---|
| gcc (Ubuntu 7.5.0-3ubuntu1~16.04) 7.5.0 Copyright (C) 2017 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. |
gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0 Copyright (C) 2017 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. |
| g++ (Ubuntu 5.4.0-6ubuntu1~16.04.12) 5.4.0 20160609 Copyright (C) 2015 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. |
g++ (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0 Copyright (C) 2017 Free Software Foundation, Inc. This is free software; see the source for copying conditions. There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. |
可以发现AIStudio的g++版本不适配，注意：你需要相同的架构，系统以及python版本，(Ubuntu)linux-x86_64&python3.7
`./compile.sh` or `bash compile.sh` if come out bash: ./compile.sh: Permission denied
或者直接进入指定目录，手动编译
cd /home/aistudio/PSENet/models/post_processing/pse
python setup.py build_ext --inplace
编译完成后你会在/home/aistudio/PSENet/models/post_processing/pse得到build/temp.linux-x86_64-3.7/pse.o文件和pse.cpython-37m-x86_64-linux-gnu.so文件
注意：本项目已经全部配置完成，这一步无需操作
训练
需要注意的是，在paddlepaddle-2.0.2中并不支持字典数据读取，因此我在/home/aistudio/PSENet/utils/data_loader.py利用迭代器重写了DataLoader这拉慢了数据读取的速度，会导致训练速度略慢，例如在使用psenet_r50_ic15_1024_finetune.py训练一个epoch需要512.4秒，另外paddlepaddle2.0.2暂不支持Identity方法，因此我在/home/aistudio/PSENet/models/utils/fuse_conv_bn.py通过继承Paddle.nn.Layer写了Identity类
cd /home/aistudio/PSENet/
python train.py ${CONFIG_FILE}
例如：
cd /home/aistudio/PSENet/
python train.py config/psenet/psenet_r50_ic15_736.py
训练开启时，会生成一个类似/home/aistudio/PSENet/checkpoints/psenet_r50_ic15_1024_finetune的文件夹，里面将保存权重和优化器参数
测试
cd /home/aistudio/PSENet/
python test.py ${CONFIG_FILE} ${CHECKPOINT_FILE}
例如：
cd /home/aistudio/PSENet/
python test.py config/psenet/psenet_r50_ic15_736.py PSENet/PretrainedModel/checkpoint_ic15_736.pdparams
评估
你需要注意的是：测试和评估是递进的，通过测试生成文件后，进行评估
ICDAR 2015
cd /home/aistudio/PSENet/eval
`./eval_ic15.sh` or `bash ./eval_ic15.sh`
你会得到如下类似信息：
Calculated!{"precision": 0.8620689655172413, "recall": 0.7944150216658642, "hmean": 0.826860435980957, "AP": 0}
以下是paddlepaddle预训练模型测试指标
| Method | Backbone | Fine-tuning | Scale | Config | Precision (%) | Recall (%) | F-measure (%) | Model |
|---|---|---|---|---|---|---|---|---|
| PSENet | ResNet50 | N | Shorter Side: 736 | psenet_r50_ic15_736.py | 83.6 | 74.0 | 78.5 | checkpoint_ic15_736 |
| PSENet | ResNet50 | N | Shorter Side: 1024 | psenet_r50_ic15_1024.py | 84.4 | 76.3 | 80.2 | checkpoint_ic15_1024 |
| PSENet | ResNet50 | Y | Shorter Side: 736 | psenet_r50_ic15_736_finetune.py | 85.3 | 76.8 | 80.9 | checkpoint_ic15_736_finetune |
| PSENet | ResNet50 | Y | Shorter Side: 1024 | psenet_r50_ic15_1024_finetune.py | 86.2 | 79.4 | 82.7 | checkpoint_ic15_1024_finetune |
Total-Text
Text detection
cd /home/aistudio/PSENet/eval
./eval_tt.sh or `bash ./eval_tt.sh`
你会得到如下类似信息：
Precision:_0.8727937336814604_______/Recall:_0.7786751361161512/Hmean:_0.8230524859472805
pb
以下是paddlepaddle预训练模型测试指标
| Method | Backbone | Fine-tuning | Config | Precision (%) | Recall (%) | F-measure (%) | Model |
|---|---|---|---|---|---|---|---|
| PSENet | ResNet50 | N | psenet_r50_tt.py | 87.3 | 77.9 | 82.3 | checkpoint_tt |
| PSENet | ResNet50 | Y | psenet_r50_tt_finetune.py | 89.3 | 79.6 | 84.2 | checkpoint_tt_finetune |
速度测试
python test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --report_speed
例如：
cd /home/aistudio/PSENet/
python test.py config/psenet/psenet_r50_ic15_736.py PSENet/PretrainedModel/checkpoint_ic15_736.pdparams --report_speed
你会得到如下类似信息
Testing 283/3000
backbone_time: 0.0152
neck_time: 0.0029
det_head_time: 0.0005
det_pse_time: 0.0660
FPS: 11.8
Testing 284/3000
backbone_time: 0.0152
neck_time: 0.0029
det_head_time: 0.0005
det_pse_time: 0.0660
FPS: 11.8
Testing 285/3000
backbone_time: 0.0152
neck_time: 0.0029
det_head_time: 0.0005
det_pse_time: 0.0660
FPS: 11.8
Testing 286/3000
backbone_time: 0.0152
neck_time: 0.0029
det_head_time: 0.0005
det_pse_time: 0.0660
FPS: 11.8
Citation
@inproceedings{wang2019shape,
title={Shape robust text detection with progressive scale expansion network},
author={Wang, Wenhai and Xie, Enze and Li, Xiang and Hou, Wenbo and Lu, Tong and Yu, Gang and Shao, Shuai},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={9336--9345},
year={2019}
}
194 Dec 28, 2022
2 Nov 06, 2021
4 Aug 06, 2021
4 Mar 02, 2022
3 Nov 17, 2021
235 Dec 22, 2022
9 Nov 11, 2022
2.5k Jan 04, 2023
285 Dec 08, 2022
170 Dec 27, 2022
17 Dec 30, 2022
309 Dec 06, 2022
4 Nov 02, 2022
6 Feb 03, 2022
1.3k Dec 22, 2022
5 May 05, 2022
356 Dec 08, 2022
28 Nov 03, 2022
111 Nov 27, 2022
230 Dec 21, 2022