Add more Swin backbones for RetinaNet and VFNet

[dnth]

This PR:

1. Adds Swin-S and Swin-B backbone to RetinaNet models.
2. Adds Swin-T, Swin-S and Swin-B backbones to VFNet.

[codecov]

Codecov Report

Merging #1042 (767f0b2) into master (bfbed41) will decrease coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1042      +/-   ##
==========================================
- Coverage   85.96%   85.95%   -0.02%     
==========================================
  Files         272      281       +9     
  Lines        5794     5922     +128     
==========================================
+ Hits         4981     5090     +109     
- Misses        813      832      +19     

Flag	Coverage Δ
unittests	85.95% <100.00%> (-0.02%)	⬇️

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

Impacted Files	Coverage Δ
icevision/models/mmdet/download_configs.py	80.95% <100.00%> (ø)
...mdet/models/retinanet/backbones/backbone_config.py	100.00% <100.00%> (ø)
...els/mmdet/models/retinanet/backbones/resnet_fpn.py	100.00% <100.00%> (ø)
...on/models/mmdet/models/retinanet/backbones/swin.py	100.00% <100.00%> (ø)
...on/models/mmdet/models/vfnet/backbones/__init__.py	100.00% <100.00%> (ø)
...ls/mmdet/models/vfnet/backbones/backbone_config.py	100.00% <100.00%> (ø)
.../models/mmdet/models/vfnet/backbones/resnet_fpn.py	100.00% <100.00%> (ø)
...vision/models/mmdet/models/vfnet/backbones/swin.py	100.00% <100.00%> (ø)
icevision/models/mmdet/common/utils.py	81.81% <0.00%> (-5.69%)	⬇️
icevision/models/mmdet/utils.py	57.81% <0.00%> (-3.86%)	⬇️
... and 9 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 bfbed41...767f0b2. Read the comment docs.

[ai-fast-track]

We can add your new config files in the airctic/mmdetection_configs

    "https://raw.githubusercontent.com/dnth/mmdetection_configs/custom-configs/configs/swin/retinanet_swin-t-p4-w7_fpn_1x_coco.py",
    "https://raw.githubusercontent.com/dnth/mmdetection_configs/custom-configs/configs/swin/retinanet_swin-s-p4-w7_fpn_1x_coco.py",
    "https://raw.githubusercontent.com/dnth/mmdetection_configs/custom-configs/configs/swin/retinanet_swin-b-p4-w7_fpn_1x_coco.py"

We have to find a way to make sure those files don't get lost when we upgrade to a new mmdet version (e.g. mmdet 2.21.0)

[dnth]

Refactored this after a video conference with Farid. Custom config files are added into the mmdetection_configs in a separate folder.

airctic/mmdetection_configs#4

[ai-fast-track]

@dnth Do you have a notebook where you tested either RetinaNet or VFNet with Swin?

What are the COCO metric for 20 iterations? (it must be low because of the un-trainede Neck)

[dnth]

Here are the COCO metrics on my machine for RetinaNet

Retina-Swin-T

Retina-Swin-S

Retina-Swin-B

[dnth]

Here are the COCO metrics on my machine for VFNet

VFNet-Swin-T

VFNet-Swin-S

VFNet-Swin-B

[dnth]

Here is a Colab notebook I ran
https://colab.research.google.com/drive/1-V_CNycmsvkENxkH08sCcOc0dXjwLtKG?usp=sharing

[dnth]

During my test, I find that for VFNet model to run, we have to keep the mmdet==2.17.0. The latest mmdet==2.20.0 has this issue.

[ai-fast-track]

Thanks a lot @dnth!

We should continue using mmdet==2.17.0 until we solve the issue wit mmdet==2.20.0

🔥 VFNet rocks! 66% with Un-trained Neck is mind blowing! I just confirm my preference towards that model!

I think the PR is ready to get merged🚀