Use fp16 utils from mmcv

[jin-s13]

Fix #153

[codecov]

Codecov Report

Merging #200 (07204a3) into master (6268fcd) will increase coverage by 0.69%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #200      +/-   ##
==========================================
+ Coverage   83.84%   84.53%   +0.69%     
==========================================
  Files         114      108       -6     
  Lines        7070     6921     -149     
  Branches     1118     1071      -47     
==========================================
- Hits         5928     5851      -77     
+ Misses        941      871      -70     
+ Partials      201      199       -2     

Flag	Coverage Δ
unittests	84.53% <100.00%> (+0.69%)	⬆️

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

Impacted Files	Coverage Δ
mmpose/core/__init__.py	100.00% <ø> (ø)
mmpose/apis/train.py	18.75% <100.00%> (+1.72%)	⬆️
mmpose/core/post_processing/group.py	84.81% <100.00%> (ø)
mmpose/models/detectors/base.py	63.46% <100.00%> (+4.76%)	⬆️
mmpose/models/detectors/bottom_up.py	64.88% <100.00%> (+0.82%)	⬆️
mmpose/models/detectors/mesh.py	92.62% <100.00%> (+0.38%)	⬆️
mmpose/models/detectors/top_down.py	79.14% <100.00%> (ø)
mmpose/models/losses/mesh_loss.py	97.76% <100.00%> (+0.08%)	⬆️
mmpose/models/losses/mse_loss.py	100.00% <100.00%> (ø)
mmpose/models/losses/multi_loss_factory.py	100.00% <100.00%> (ø)

---

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 6268fcd...54eb7de. Read the comment docs.

[innerlee]

I haven't got the chance to try fp16, so @hellock may review this

[hellock]

How about the benchmark?

[jin-s13]

Still have some problems to solve:

1. The training is not stable... The accuracy is normal at first, however, it will suddenly decline to 0 at some point (about 140 epochs).
2. The results may contain NAN values.
3. Currently, the loss calculation & post-processing use fp32.