DSFD: Dual Shot Face Detector
1、Description
DSFD for face detection in general scene with high detection rate
2、Prepare data
1、download WIDER Face Dateset， locate /opt/npu/
2、data process perform the following create data for train
python prepare_wider_data
3、Train
Note: pillow recommends installing a newer version. If the corresponding torchvision version cannot be installed directly, you can use the source code to install the corresponding version. The source code reference link: https://github.com/pytorch/vision，
Suggestion the pillow is 9.1.0 and the torchvision is 0.6.0 1、pretrained weight
download [pretrained weights](链接：https://pan.baidu.com/s/1qbQsOcgD3vuJ5m3Jnu6HTw 提取码：vbo9)
2、train full1p get train_1p log
bash ./test/train_full_1p.sh
3、train 1p performance
bash./test/train_performance_1p.sh
4、train full 8p get train_8p log
bash ./test/train_full_8p.sh
5、train 8p performance
bash ./test/train_performance_8p.sh
6、use resume training
#enable resume training
add --resume "path/to/checkpoint" to .sh
4、Data inference
1.download wider_face_test.mat and wider_face_val.mat in /tools/infer_tools
2.cd /tools/infer_tools
python wider_face_test.py
5、Evalution
1、do setup first
cd /tools/eval_tools
python setup.py build_ext --inplace
2、download ground_truth and unzip to /tools/eval_tools
3、get data evaluation result
python evaluation.py
6、Demo
python demo.py --network 'resnet152'
7、DSFD training result
| Acc | Npu_numbers | epochs | AMP_type |
|---|---|---|---|
| - | 1 | 1 | O2 |
| E:0.9368 M:0.9282 H:0.8460 | 8 | 100 | O2 |
Reference:
| Acc | |
|---|---|
| 参考精度 | E:0.951 M:0.936 H:0.837 |
| GPU 8P 自测精度 | E 0.9473, M 0.9362, H 0.8651 |
Statement
For details about the public address of the code in this repository, you can get from the file public_address_statement.md