[Fix] fix generalized attention fp16

[AronLin]

Motivation

Fix open-mmlab/mmdetection#1241.

The function get_position_embedding in GeneralizedAttention returns float32 values, even if in fp16 mode.

Modification

• Cast the return values of get_position_embedding function
• Add fp16 unit tests for GeneralizedAttention with attention_type='1111'.

Comments

This bug does not exists in pytorch >= 1.6.0 and mmcv >= 1.3.2. It is not a bug with torch.cuda.amp.autocast applied.

BC-breaking (Optional)

No

[codecov]

Codecov Report

Merging #1036 (a4e751d) into master (b8c09f3) will increase coverage by 0.02%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1036      +/-   ##
==========================================
+ Coverage   65.30%   65.32%   +0.02%     
==========================================
  Files         154      154              
  Lines        9891     9895       +4     
  Branches     1801     1801              
==========================================
+ Hits         6459     6464       +5     
- Misses       3099     3100       +1     
+ Partials      333      331       -2     

Flag	Coverage Δ
unittests	65.32% <100.00%> (+0.02%)	⬆️

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

Impacted Files	Coverage Δ
mmcv/cnn/bricks/generalized_attention.py	92.81% <100.00%> (ø)
mmcv/runner/hooks/logger/tensorboard.py	34.37% <0.00%> (-1.11%)	⬇️
mmcv/runner/hooks/logger/pavi.py	71.01% <0.00%> (+0.42%)	⬆️
mmcv/runner/hooks/logger/mlflow.py	81.25% <0.00%> (+0.60%)	⬆️
mmcv/runner/hooks/logger/wandb.py	69.69% <0.00%> (+0.94%)	⬆️
mmcv/runner/hooks/logger/base.py	70.87% <0.00%> (+1.94%)	⬆️

---

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 b8c09f3...a4e751d. Read the comment docs.