[Fix] Fix DCN fp16

[shinya7y]

Motivation

Fix #1004

Modification

Cast DCN weights by type_as, not by autocast.

BC-breaking (Optional)

No.

[CLAassistant]

All committers have signed the CLA.

[shinya7y]

Another solution with autocast (with more manual fix):
(1) Change _addmm to addmm in deform_conv_cuda.cu
https://pytorch.org/docs/stable/amp.html#autocast-op-reference
(2) Apply custom_fwd and custom_bwd
https://pytorch.org/docs/stable/notes/amp_examples.html#autocast-and-custom-autograd-functions

To support PyTorch < 1.6, code for (2) will be long.

[codecov]

Codecov Report

Merging #1014 (8ee2f20) into master (db6b054) will increase coverage by 0.00%.
The diff coverage is 100.00%.

❗ Current head 8ee2f20 differs from pull request most recent head c0edcf6. Consider uploading reports for the commit c0edcf6 to get more accurate results

@@           Coverage Diff           @@
##           master    #1014   +/-   ##
=======================================
  Coverage   64.57%   64.57%           
=======================================
  Files         152      152           
  Lines        9792     9793    +1     
  Branches     1779     1779           
=======================================
+ Hits         6323     6324    +1     
  Misses       3141     3141           
  Partials      328      328           

Flag	Coverage Δ
unittests	64.57% <100.00%> (+<0.01%)	⬆️

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

Impacted Files	Coverage Δ
mmcv/ops/deform_conv.py	61.76% <100.00%> (+0.28%)	⬆️

---

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 db6b054...c0edcf6. Read the comment docs.

[ZwwWayne]

Hi @shinya7y ,
Thanks for the PR. We plan to merge it in a few days. Before that, could you add comments to indicate the reason for the modification or refer the issue in the code?