Support Exporting Mask Rcnn to ONNX

[lara-hdr]

In addition to #1401 , these are the last changes to export Maskrcnn to ONNX.
pytorch/pytorch#27566 and pytorch/pytorch#26549.
Like for Faster rcnn, for now we will only be able to export with a batch of 1 (until we support dynamic cases for split).

The test is disabled until the referenced PRs are merged and Resize is implemented in opset 11.

[fmassa]

The PR looks pretty good, thanks!

I'd like to have some more tests for the functions that have custom implementation for ONNX, so that we can enforce that they are not going to be broken in the future by changes to one function or another.

[codecov-io]

Codecov Report

Merging #1461 into master will decrease coverage by 0.65%.
The diff coverage is 45.61%.

@@            Coverage Diff             @@
##           master    #1461      +/-   ##
==========================================
- Coverage   64.34%   63.69%   -0.66%     
==========================================
  Files          83       83              
  Lines        6454     6561     +107     
  Branches      992     1010      +18     
==========================================
+ Hits         4153     4179      +26     
- Misses       2006     2087      +81     
  Partials      295      295

Impacted Files	Coverage Δ
torchvision/models/detection/roi_heads.py	54.36% <45.61%> (-1.78%)	⬇️
torchvision/io/video.py	28.32% <0%> (-46.68%)	⬇️
torchvision/datasets/svhn.py	66.66% <0%> (-2.09%)	⬇️
torchvision/datasets/utils.py	57.66% <0%> (-0.72%)	⬇️
torchvision/io/__init__.py	100% <0%> (ø)	⬆️
torchvision/datasets/mnist.py	51.94% <0%> (+0.23%)	⬆️
torchvision/datasets/video_utils.py	65.08% <0%> (+1.77%)	⬆️
torchvision/__init__.py	70.96% <0%> (+3.22%)	⬆️
torchvision/datasets/samplers/clip_sampler.py	79.54% <0%> (+23.98%)	⬆️
torchvision/io/_video_opt.py	65.85% <0%> (+26.49%)	⬆️

---

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 f46f2c1...bb5842c. Read the comment docs.

[fmassa]

I did one more pass, let me know what you think

[fmassa]

This is almost good to go, thanks!

There is a lint failure, and I'd like to have one more test to make sure that the scripted model indeed captures control-flow as expected.

[fmassa]

LGTM, thanks a lot Lara!

[niaoyu]

LGTM, thanks a lot Lara!

Hi, i just wonder which version of torch and torchvison would be suitable for export maskrcnn to onnx.

I tried the nightly version of both , but all failed

[lara-hdr]

@niaoyu, as explained in pytorch/pytorch#27969 there are some PRs that are not yet merged in PyTorch to export this model.