[Refactor]: Unified parameter initialization

[hhaAndroid]

Motivation

Refactoring model initialization methods based MMCV. For more information, please refer to open-mmlab/mmcv#780

Dependency

• [Refactoring] Unified parameters initialization mmcv#780
• [Refactoring]Add Sequential with init_weight mmcv#884
• [Refactoring] Revise function of layers and override keys in init_cfg mmcv#893
• add more detail of warning in basemodule mmcv#899
• [Refactoring] Add Caffe2Xavier Initializer mmcv#902
• [Fix]: Fix initialization bug of override cfg mmcv#918
• [Fix] Revise override in init_cfg mmcv#930
• Remove redundant self.init_cfg mmcv#959
• [Fix] Rename init_weight to init_weights mmcv#971

Models

Models that have been refactored as follow:

• ATSS
• CARAFE
• Cascade R-CNN
• Faster R-CNN
• DCN
• Double Head
• Dynamic R-CNN
• Empirical Attention
• Foveabox
• FreeAnchor
• FASF
• GCNet
• GFL
• GN
• GROIE
• Guided Anchoring
• HRNet
• HTC
• MS R-CNN
• NAS FPN
• PAA
• PAFPN
• PISA
• RepPoints
• ResneSt
• RetinaNet
• GHM
• RPN
• SABL
• SCNet
• TridentNet
• Sparse R-CNN
• VFNet
• YOLOV3
• FPG
• cascade_rpn
• centripetalnet
• cornernet
• detectors
• detr
• nas_fcos
• gn+ws (40.2/36.2) vs (40.6/36.6)
• grid_rcnn
• libra_rcnn
• point_rend
• regnet
• res2net
• ssd
• yolact

All models have been checked and passed！

[ZwwWayne]

1. Fix lint
2. Involve @MeowZheng to have a look.

[hhaAndroid]

@MeowZheng @jshilong Please review the one-stage algorithm first, I tagged todo where I couldn't be sure.

[jshilong]

Please set self._is_init as True in all init_weight your overload in this PR.

[MeowZheng]

Many thanks for your nice work, it looks good.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ hhaAndroid
❌ qwe12369
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[codecov]

Codecov Report

Merging #4750 (a11e276) into master (0425fb6) will decrease coverage by 1.18%.
The diff coverage is 80.49%.

@@            Coverage Diff             @@
##           master    #4750      +/-   ##
==========================================
- Coverage   66.03%   64.85%   -1.19%     
==========================================
  Files         261      261              
  Lines       20306    19915     -391     
  Branches     3466     3371      -95     
==========================================
- Hits        13410    12916     -494     
- Misses       6156     6251      +95     
- Partials      740      748       +8     

Flag	Coverage Δ
unittests	64.85% <80.49%> (-1.15%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmdet/models/detectors/atss.py	100.00% <ø> (ø)
mmdet/models/detectors/cascade_rcnn.py	50.00% <ø> (ø)
mmdet/models/detectors/cornernet.py	94.87% <ø> (-5.13%)	⬇️
mmdet/models/detectors/detr.py	100.00% <ø> (ø)
mmdet/models/detectors/fast_rcnn.py	43.75% <ø> (ø)
mmdet/models/detectors/faster_rcnn.py	100.00% <ø> (ø)
mmdet/models/detectors/fcos.py	100.00% <ø> (ø)
mmdet/models/detectors/fovea.py	100.00% <ø> (ø)
mmdet/models/detectors/fsaf.py	100.00% <ø> (ø)
mmdet/models/detectors/gfl.py	100.00% <ø> (ø)
... and 145 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0425fb6...a11e276. Read the comment docs.

[shinya7y]

Will recent detectors be modified?

• Deformable DETR
• YOLOF
• AutoAssign
• Localization Distillation

[hhaAndroid]

Will recent detectors be modified

No modification for the time being, if you think it is necessary, we can consider doing it in advance.

[shinya7y]

Thanks. My understanding is that modification is not essential (except for backbones).