Support corner_pool related custom operators for onnxruntime in mmcv

[v-qjqs]

Hi, this PR supports to export corner_pool related custom operators (TopPool,BottomPool,LeftPool,RightPool ) to onnx, and run it with onnxruntime.
It's working in process during my free time.

[codecov]

Codecov Report

Merging #997 (68bcc9c) into master (c4aa414) will decrease coverage by 0.03%.
The diff coverage is 38.46%.

@@            Coverage Diff             @@
##           master     #997      +/-   ##
==========================================
- Coverage   64.60%   64.57%   -0.04%     
==========================================
  Files         152      152              
  Lines        9779     9792      +13     
  Branches     1779     1779              
==========================================
+ Hits         6318     6323       +5     
- Misses       3133     3141       +8     
  Partials      328      328              

Flag	Coverage Δ
unittests	64.57% <38.46%> (-0.04%)	⬇️

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

Impacted Files	Coverage Δ
mmcv/ops/corner_pool.py	52.77% <38.46%> (-3.16%)	⬇️

---

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 c4aa414...68bcc9c. Read the comment docs.

[v-qjqs]

Hi, any reviewers?

[grimoire]

Thanks for the contribution!
Would it be better to add a cummax custom op instead of CornerPool op? It is more general and can support torch>1.5.0.

[v-qjqs]

Hi @grimoire, I tested the implemented mmcv::CornerPool with my torch1.5.0 environment and it works. I wonder why this line is needed? Isn't it ok for just using self.corner_pool.apply(x)?

[grimoire]

PyTorch built-in ops support more platforms and might get better support from different deploy backends in the future.
As we do plan to add more backends support. It is a good idea to avoid custom ops unless we have to do.

[grimoire]

Hi, could you please add document about your ops in onnxruntime_custom_ops and onnxruntime_op?

[v-qjqs]

Thanks for the contribution!
Would it be better to add a cummax custom op instead of CornerPool op? It is more general and can support torch>1.5.0.

The implementation of mmcv::CornerPool helps to those who use self.corner_pool.apply(x) as the forward function.
I'm newer to onnx, perhaps a custom cummax op will be supported in next PR during my free time, following the official pytorch implementation.

[CLAassistant]

All committers have signed the CLA.

[grimoire]

LGTM!

[grimoire]

@ZwwWayne any comments?