[Feature] follow-up work to enable exporting to onnx for PointRend (#4977)

[AronLin]

This PR is the follow-up work to enable exporting to ONNX for PointRend.
The origin PR is #4977 and open-mmlab/mmcv#953.

Motivation

The origin PR completes the exporting work but we refactor the code logic for ONNX exporting.
So I completed the follow-up logic reconstruction work for PointRend.

Test

Here is the result:

Model	Type	box mAP	box AP50	box AP75	mask mAP	mask AP50	mask AP75
point_rend_r50_caffe_fpn_mstrain_1x_coco	PyTorch	38.4	59.0	41.8	36.3	56.9	38.7
point_rend_r50_caffe_fpn_mstrain_1x_coco	ONNX	38.4	59.0	41.8	35.2	56.5	37.1

It seems all right.

[codecov]

Codecov Report

Merging #5440 (d5ed09e) into master (269bb9e) will decrease coverage by 0.10%.
The diff coverage is 8.64%.

@@            Coverage Diff             @@
##           master    #5440      +/-   ##
==========================================
- Coverage   64.90%   64.80%   -0.11%     
==========================================
  Files         276      276              
  Lines       21347    21527     +180     
  Branches     3543     3562      +19     
==========================================
+ Hits        13856    13951      +95     
- Misses       6733     6820      +87     
+ Partials      758      756       -2     

Flag	Coverage Δ
unittests	64.77% <8.64%> (-0.10%)	⬇️

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

Impacted Files	Coverage Δ
...det/models/roi_heads/mask_heads/mask_point_head.py	31.53% <0.00%> (-1.18%)	⬇️
mmdet/models/roi_heads/standard_roi_head.py	50.31% <ø> (ø)
mmdet/models/roi_heads/test_mixins.py	57.25% <0.00%> (ø)
mmdet/models/roi_heads/point_rend_roi_head.py	16.09% <10.14%> (-3.54%)	⬇️
mmdet/models/detectors/single_stage.py	66.03% <0.00%> (-5.67%)	⬇️
mmdet/models/detectors/cornernet.py	94.87% <0.00%> (-5.13%)	⬇️
mmdet/models/dense_heads/corner_head.py	67.48% <0.00%> (-2.46%)	⬇️
mmdet/models/roi_heads/cascade_roi_head.py	74.69% <0.00%> (-1.68%)	⬇️
mmdet/models/backbones/ssd_vgg.py	64.44% <0.00%> (-1.47%)	⬇️
mmdet/models/backbones/darknet.py	57.69% <0.00%> (-1.29%)	⬇️
... and 14 more

---

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 269bb9e...d5ed09e. Read the comment docs.

[AronLin]

The tests for single image inference, batch inference and ONNX inference have passed. The PR can be merged now.

[fschvart]

Hi, Thanks for all the hard work!
I've been going insane since I can't make the 3x schedule of PointRend convert to ONNX. Either I'm having issues with the grid_sample, or get other weird errors if I use TensorRT. Is it supposed to work? Am I missing anything?