标签：centerX KD -- 检测 resdcn18 开源 sh CenterX coco
作者：CPFLAME
项目代码：https://github.com/CPFLAME/centerX
项目介绍：https://zhuanlan.zhihu.com/p/323814368
CenterX
This repo is implemented based on detectron2 and CenterNet
What’s new
- Support imgaug data augmentation
- Support swa
- Support Knowledge Distill, teacher-student, designed by myself
- Support other LR_SCHEDULER
- Support Optimizer RangerLars, not convergence in COCO
- We provide some examples and scripts to convert centerX to Caffe, ONNX and TensorRT format in projects/speedup
What’s comming
- [️✔] Support simple inference
- [✔] Support to caffe, onnx, tensorRT
- Support keypoints
Requirements
- Python >= 3.7
- PyTorch >= 1.5
- torchvision that matches the PyTorch installation.
- OpenCV
- pycocotools
pip install cython; pip install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
- GCC >= 4.9
gcc --version
- detectron2
pip install -U 'git+https://github.com/CPFLAME/detectron2.git'
pip install git+https://github.com/pabloppp/pytorch-tools -U
Data prepare
the same as detectron2
Training
modify your yamls in run.sh
sh run.sh
Testing and Evaluation
modify your yamls in run.sh,
add eval-only and MODEL.WEIGHTS in your setting
sh run.sh
Performance
coco
This repo use less training time to get a competitive performance compared to other versions
Backbone ResNet-50
| Code | mAP | epoch |
|---|---|---|
| centerX | 33.2 | 70 |
| centerX | 34.3 | 140 |
| centernet-better | 34.9 | 140 |
Backbone ResNet-18
centerX_KD means ResNet-50(33.2) as teacher, ResNet-18(27.9) as student, Knowledge Distill for 70 epoch in coco.
| Code | mAP | epoch |
|---|---|---|
| centerX | 30.2 | 140 |
| centerX | 27.9 | 70 |
| centerX_KD | 31.0 | 70 |
| centernet-better | 29.8 | 140 |
| centernet | 28.1 | 140 |
crowd human
- optim: SGD
- lr: 0.02
- scheduler: WarmupMultiStepLR ,drop 0.1 in (50, 62) for 80 epoch; (90 ,120) for 140 epoch
- train size: 512 max size
- test size: 512 max size
- batch size: 64
- woGT: KD only use teacher loss
| Backbone | mAP | mAP50 | mAP75 | epoch | teacher | student_pretrain |
|---|---|---|---|---|---|---|
| resdcn18 | 31.2 | 56.6 | 30.8 | 80 | - | - |
| resdcn18_swa | 31.1 | 56.6 | 30.4 | 80 | - | - |
| resdcn18_syncBN | 31.3 | 56.6 | 30.7 | 80 | - | - |
| resdcn18_imgaug | 29.6 | 54.7 | 28.9 | 80 | - | - |
| resdcn18_KD | 34.5 | 60.2 | 34.3 | 80 | resdcn50 | resdcn18 |
| resdcn18_KD_woGT | 33.0 | 58.3 | 32.7 | 80 | resdcn50 | resdcn18 |
| resdcn18_KD_woGT_scratch | 32.8 | 58.1 | 32.6 | 140 | resdcn50 | imagenet |
| resdcn50 | 35.1 | 61.2 | 35.3 | 80 | - | - |
KD exp
crowd human KD
Generalization performance for Knowledge Distill
| Backbone | crowd mAP | coco_person mAP | epoch | teacher | student_pretrain | train_set |
|---|---|---|---|---|---|---|
| resdcn50 | 35.1 | 35.7 | 80 | - | - | crowd |
| resdcn18(baseline) | 31.2 | 31.2 | 80 | - | - | crowd |
| resdcn18_KD | 34.5 | 34.9 | 80 | resdcn50 | resdcn18 | crowd |
| resdcn18_KD_woGT_scratch | 32.8 | 34.2 | 140 | resdcn50 | imagenet | crowd |
| resdcn18_KD_woGT_scratch | 34.1 | 36.3 | 140 | resdcn50 | imagenet | crowd+coco |
multi teacher KD
| Backbone | mAP crowd | mAP coco_car | epoch | teacher | student_pretrain | train_set |
|---|---|---|---|---|---|---|
| 1.resdcn50 | 35.1 | - | 80 | - | - | crowd |
| 2.resdcn18 | 31.7 | - | 70 | - | - | crowd |
| 3.resdcn50 | - | 31.6 | 70 | - | - | coco_car |
| 4.resdcn18 | - | 27.8 | 70 | - | - | coco_car |
| resdcn18_KD_woGT_scratch | 31.6 | 29.4 | 140 | 1,3 | imagenet | crowd+coco_car |
| Backbone | mAP crowd_human | mAP widerface | epoch | teacher | student_pretrain | train_set |
|---|---|---|---|---|---|---|
| 1.resdcn50 | 35.1 | - | 80 | - | - | crowd |
| 2.resdcn18 | 31.7 | - | 70 | - | - | crowd |
| 3.resdcn50 | - | 32.9 | 70 | - | - | widerface |
| 4.resdcn18 | - | 29.6 | 70 | - | - | widerface |
| 5.resdcn18_ignore_nolabel | 29.1 | 24.2 | 140 | - | - | crowd+wider |
| 6.resdcn18_pseudo_label | 28.9 | 27.7 | 140 | - | - | crowd+wider |
| 7.resdcn18_KD_woGT_scratch | 31.3 | 32.1 | 140 | 1,3 | imagenet | crowd+wider |
License
centerX is released under the Apache 2.0 license.
Acknowledgement
标签：centerX,KD,--,检测,resdcn18,开源,sh,CenterX,coco 来源： https://blog.csdn.net/weixin_47477271/article/details/110455112
本站声明： 1. iCode9 技术分享网（下文简称本站）提供的所有内容，仅供技术学习、探讨和分享； 2. 关于本站的所有留言、评论、转载及引用，纯属内容发起人的个人观点，与本站观点和立场无关； 3. 关于本站的所有言论和文字，纯属内容发起人的个人观点，与本站观点和立场无关； 4. 本站文章均是网友提供，不完全保证技术分享内容的完整性、准确性、时效性、风险性和版权归属；如您发现该文章侵犯了您的权益，可联系我们第一时间进行删除； 5. 本站为非盈利性的个人网站，所有内容不会用来进行牟利，也不会利用任何形式的广告来间接获益，纯粹是为了广大技术爱好者提供技术内容和技术思想的分享性交流网站。