[Feature] Implement ACRN and Focal Loss

[kennymckormick]

No description provided.

[codecov]

Codecov Report

Merging #891 (7e1d8ea) into master (e9b7009) will increase coverage by 0.00%.
The diff coverage is 88.88%.

❗ Current head 7e1d8ea differs from pull request most recent head ac9cf82. Consider uploading reports for the commit ac9cf82 to get more accurate results

@@           Coverage Diff           @@
##           master     #891   +/-   ##
=======================================
  Coverage   83.71%   83.71%           
=======================================
  Files         131      132    +1     
  Lines        9870     9932   +62     
  Branches     1699     1710   +11     
=======================================
+ Hits         8263     8315   +52     
- Misses       1197     1203    +6     
- Partials      410      414    +4     

Flag	Coverage Δ
unittests	83.71% <88.88%> (+<0.01%)	⬆️

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

Impacted Files	Coverage Δ
mmaction/datasets/ava_dataset.py	90.60% <ø> (ø)
mmaction/models/heads/roi_head.py	75.00% <50.00%> (ø)
...maction/models/roi_extractors/single_straight3d.py	74.24% <66.66%> (-3.35%)	⬇️
mmaction/models/heads/misc_head.py	93.75% <93.75%> (ø)
mmaction/models/__init__.py	100.00% <100.00%> (ø)
mmaction/models/heads/__init__.py	100.00% <100.00%> (ø)
mmaction/models/heads/bbox_head.py	91.47% <100.00%> (+0.34%)	⬆️
mmaction/models/heads/fbo_head.py	78.49% <100.00%> (ø)
mmaction/models/heads/lfb_infer_head.py	45.00% <100.00%> (ø)
mmaction/core/evaluation/accuracy.py	92.27% <0.00%> (-0.91%)	⬇️
... and 2 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 e9b7009...ac9cf82. Read the comment docs.

[dreamerlin]

Sth in all config files should be removed.

[congee524]

Why we put acrn_head in misc_head.py

[dreamerlin]

How about moving this part as class FocalLoss and using it by build_loss(xxx)?

[congee524]

like BaseHead

[kennymckormick]

I think we can make FocalLoss a separate module when there is evidence that it can lead to performance gain on other tasks. For now, it is only used in the training of AVA.

[kennymckormick]

Why we put acrn_head in misc_head.py

I will add a series of shared_heads for feature fusion. Currently, I only committed ACRNHead since it's the most successful one.

[dreamerlin]

Why we put acrn_head in misc_head.py

I will add a series of shared_heads for feature fusion. Currently, I only committed ACRNHead since it's the most successful one.

I guess what he means is that why not name it as acrn_head

[kennymckormick]

Why we put acrn_head in misc_head.py

I will add a series of shared_heads for feature fusion. Currently, I only committed ACRNHead since it's the most successful one.

I guess what he means is that why not name it as acrn_head

I know. Since I will add a series of heads in the future, I plan to put all of them in the misc_head.py to avoid creating a new file for each individual.