add ort nms op

[tangyanf]

add onnxruntime custom op: nms

[codecov]

Codecov Report

Merging #803 (3c76e98) into master (00658e8) will increase coverage by 0.29%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master     #803      +/-   ##
==========================================
+ Coverage   61.94%   62.23%   +0.29%     
==========================================
  Files         144      144              
  Lines        8467     8506      +39     
  Branches     1520     1521       +1     
==========================================
+ Hits         5245     5294      +49     
+ Misses       2954     2947       -7     
+ Partials      268      265       -3     

Flag	Coverage Δ
unittests	62.23% <0.00%> (+0.29%)	⬆️

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

Impacted Files	Coverage Δ
mmcv/ops/nms.py	34.43% <0.00%> (ø)
mmcv/image/io.py	90.75% <0.00%> (+1.77%)	⬆️
mmcv/runner/checkpoint.py	66.22% <0.00%> (+7.89%)	⬆️

---

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 00658e8...3c76e98. Read the comment docs.

[hellock]

@grimoire Would you help review this PR?

[grimoire]

@hellock Sure, I will do it.

[grimoire]

Hi @tangyanf, thanks for your contribution!
As long as onnxruntime already has an official nms implementation, I think we might not have to create custom one.
The diff between pytorch and onnxruntime is caused by offset, use the code below could fix this:

        if torch.onnx.is_in_onnx_export() and offset == 1:
            # ONNX only support offset == 1
            boxes[:, -2:] += 1
        inds = NMSop.apply(boxes, scores, iou_threshold, offset)
        if torch.onnx.is_in_onnx_export() and offset == 1:
            # ONNX only support offset == 1
            boxes[:, -2:] -= 1