[Enhancement]: Refactor SSD

[RangiLyu]

This PR including these modifications

1. Refactor SSD-VGG backbone. Move extra conv layers to SSDNeck
2. Add SSDNeck.
3. Make SSDHead more flexible, e.g. support depthwise separable conv, support conv tower, support norm layer.
4. Add SSDNeck unit test.

I have trained SSD300 and the mAP is 25.9 (origin 25.6), SSD512 mAP is 29.8 (origin 29.4).

[codecov]

Codecov Report

Merging #5291 (f142df6) into master (87ad480) will decrease coverage by 0.15%.
The diff coverage is 66.66%.

❗ Current head f142df6 differs from pull request most recent head 4b7b094. Consider uploading reports for the commit 4b7b094 to get more accurate results

@@            Coverage Diff             @@
##           master    #5291      +/-   ##
==========================================
- Coverage   65.14%   64.98%   -0.16%     
==========================================
  Files         275      276       +1     
  Lines       21248    21289      +41     
  Branches     3535     3540       +5     
==========================================
- Hits        13841    13835       -6     
- Misses       6654     6698      +44     
- Partials      753      756       +3     

Flag	Coverage Δ
unittests	64.96% <66.66%> (-0.14%)	⬇️

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

Impacted Files	Coverage Δ
mmdet/core/export/model_wrappers.py	0.00% <0.00%> (ø)
mmdet/core/export/pytorch2onnx.py	14.03% <0.00%> (-1.66%)	⬇️
mmdet/models/backbones/ssd_vgg.py	65.90% <40.00%> (-21.75%)	⬇️
mmdet/models/dense_heads/ssd_head.py	87.37% <75.86%> (-5.40%)	⬇️
mmdet/models/necks/__init__.py	100.00% <100.00%> (ø)
mmdet/models/necks/ssd_neck.py	100.00% <100.00%> (ø)
mmdet/models/utils/normed_predictor.py	83.33% <0.00%> (-9.53%)	⬇️
mmdet/models/detectors/cornernet.py	94.87% <0.00%> (-5.13%)	⬇️
mmdet/models/roi_heads/mask_heads/fcn_mask_head.py	69.76% <0.00%> (-3.49%)	⬇️
mmdet/models/dense_heads/corner_head.py	67.48% <0.00%> (-2.46%)	⬇️
... and 1 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 87ad480...4b7b094. Read the comment docs.

[ZwwWayne]

Can we add a model conversion script that can load the previously trained SSD successfully loaded in this PR? We can also do that through _load_from_state_dict without extra scripts.

[RangiLyu]

Can we add a model conversion script that can load the previously trained SSD successfully loaded in this PR? We can also do that through _load_from_state_dict without extra scripts.

Since the extra layers in SSDVGG have been moved to SSDNeck, the _load_from_state_dict function needs to be put into the detector class. But SSD uses SingleStageDetector so I choose to write a convert weight script instead of putting the weight convert function into the base class.

[ZwwWayne]

We also need to update the compatibility.md