[Feature] Support Dense Teacher

[shufanwu]

This pr implement dense teacher for mmdetection. paper: https://arxiv.org/abs/2207.02541v2

[CLAassistant]

All committers have signed the CLA.

[Czm369]

Excuse me, are the experimental results in line with expectations?

[hhaAndroid]

@shufanwu Do you have any progress?

[shufanwu]

@shufanwu Do you have any progress?

I have asked the author for the official model weight and aligned accuracy

[lzx101]

@shufanwu 你好 请问一下我想复现一下这个代码有版本要求吗

[shufanwu]

@shufanwu 你好 请问一下我想复现一下这个代码有版本要求吗

你是想自己再实现一遍dense teacher还是用我实现的代码训练？

[ltyooo]

@shufanwu 你好，我想用你的代码复现一遍，然后提示我mmdet没有注册是什么意思呢

[ltyooo]

方便加您一个联系方式吗

[ltyooo]

KeyError: 'DenseTeacher is not in the mmdet::model registry. Please check whether the value of DenseTeacher is correct or it was registered as expected. More details can be found at https://mmengine.readthedocs.io/en/latest/advanced_tutorials/config.html#import-the-custom-module'出现了这个错误

[ltyooo]

我检查过了注册列表，没有问题，可能是依赖项的问题，方便问一下版本吗

[lzx101]

大佬你好，我用了你的代码跑了一遍，效果很差 我这边用的是v100，1个gpu，学习率被我改为0.005，请问我还需要更改什么来提高效果吗 2023/10/14 11:01:21 - mmengine - INFO - Iter(val) [5000/5000] teacher/coco/bbox_mAP: 0.0650 teacher/coco/bbox_mAP_50: 0.1340 teacher/coco/bbox_mAP_75: 0.0550 teacher/coco/bbox_mAP_s: 0.0390 teacher/coco/bbox_mAP_m: 0.0810 teacher/coco/bbox_mAP_l: 0.0820 student/coco/bbox_mAP: 0.2400 student/coco/bbox_mAP_50: 0.3870 student/coco/bbox_mAP_75: 0.2540 student/coco/bbox_mAP_s: 0.1170 student/coco/bbox_mAP_m: 0.2590 student/coco/bbox_mAP_l: 0.3230 data_time: 0.0135 time: 0.0480

[lzx101]

验证集的效果可以，但是测试集效果不行 2023/10/14 11:18:00 - mmengine - INFO - Iter(test) [5000/5000] eta: 0:00:00 time: 0.0352 data_time: 0.0010 memory: 477
2023/10/14 11:18:08 - mmengine - INFO - Evaluating bbox... 2023/10/14 11:19:00 - mmengine - INFO - bbox_mAP_copypaste: 0.065 0.134 0.055 0.039 0.081 0.082 2023/10/14 11:19:00 - mmengine - INFO - Iter(test) [5000/5000] coco/bbox_mAP: 0.0650 coco/bbox_mAP_50: 0.1340 coco/bbox_mAP_75: 0.0550 coco/bbox_mAP_s: 0.0390 coco/bbox_mAP_m: 0.0810 coco/bbox_mAP_l: 0.0820 data_time: 0.0009 time: 0.0354

[FionaZhangJ]

大佬你好，我用了你的代码跑了一遍，效果很差 我这边用的是v100，1个gpu，学习率被我改为0.005，请问我还需要更改什么来提高效果吗 2023/10/14 11:01:21 - mmengine - INFO - Iter(val) [5000/5000] teacher/coco/bbox_mAP: 0.0650 teacher/coco/bbox_mAP_50: 0.1340 teacher/coco/bbox_mAP_75: 0.0550 teacher/coco/bbox_mAP_s: 0.0390 teacher/coco/bbox_mAP_m: 0.0810 teacher/coco/bbox_mAP_l: 0.0820 student/coco/bbox_mAP: 0.2400 student/coco/bbox_mAP_50: 0.3870 student/coco/bbox_mAP_75: 0.2540 student/coco/bbox_mAP_s: 0.1170 student/coco/bbox_mAP_m: 0.2590 student/coco/bbox_mAP_l: 0.3230 data_time: 0.0135 time: 0.0480

你好，我按照这个代码也跑了，效果确实不好，mAp在burn in后一直在往下掉，想问下你解决这个问题了吗？