Clamp neg probability in freeanchor

[ZwwWayne]

This PR tries to resolve the training instability of FreeAnchor when the probabilities are very close to 1 or 0.

[codecov]

Codecov Report

Merging #4082 (1c2b9f5) into master (5853fa4) will increase coverage by 0.07%.
The diff coverage is 70.37%.

@@            Coverage Diff             @@
##           master    #4082      +/-   ##
==========================================
+ Coverage   61.88%   61.96%   +0.07%     
==========================================
  Files         224      225       +1     
  Lines       16619    16644      +25     
  Branches     2809     2814       +5     
==========================================
+ Hits        10285    10313      +28     
+ Misses       5835     5832       -3     
  Partials      499      499              

Flag	Coverage Δ
unittests	61.91% <70.37%> (+0.06%)	⬆️

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

Impacted Files	Coverage Δ
...mdet/models/dense_heads/free_anchor_retina_head.py	21.33% <0.00%> (-0.89%)	⬇️
mmdet/models/necks/channel_mapper.py	84.21% <84.21%> (ø)
mmdet/models/dense_heads/ssd_head.py	92.04% <100.00%> (+0.18%)	⬆️
mmdet/models/necks/__init__.py	100.00% <100.00%> (ø)
mmdet/models/dense_heads/corner_head.py	76.77% <0.00%> (+1.91%)	⬆️
mmdet/models/detectors/cornernet.py	100.00% <0.00%> (+5.12%)	⬆️

---

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 d3cf38d...6f30397. Read the comment docs.