Add Mask R-CNN

[NielsRogge]

What does this PR do?

This PR is a further development of the Mask R-CNN framework. Supersedes #22973.

Updates:

• improved variable names, docstrings (especially also for the configuration class)
• removed specific __repr__ and __nice__ methods
• favor ONNX-compatible code wherever possible, instead of if-else statements
• torchvision was already leveraged for NMS

Regarding this:

In some of the model-side processing code, there's switching to CPU and casting back and forth between torch and numpy - it's not clear why.

=> this is because when placing masks on the GPU, this would cause OOM errors. Hence those are placed on the CPU.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint.

[amyeroberts]

Thanks for iterating on this!

I've done an initial review - commenting on the things that caught my eye. It wasn't however, extensive and I expect this PR will still need several rounds of review.

General comments:

• Could you add type hints to arguments in the functions?
• Make sure all doc strings are filled out fully, including: descriptions of what the argument is, any shape information, shape information with explicit variable names
• Given the complexity of this model, it needs a lot more tests beyond the defaults in the testing suite. In particular for implementations of loss averaging and NMS
• N is used everywhere - it should be replaced with something more explicit
• As this PR is so necessarily big, I would try to find ways to make it as small as possible and then have follow up PRs for additional functionality. In particular, it would be a lot easier to review without the conditional onnx logic. Could you remove this and then add in a separate PR please?

[amyeroberts]

Thanks for iterating on this!

I've done an initial review - commenting on the things that caught my eye. It wasn't however, extensive and I expect this PR will still need several rounds of review.

General comments:

• Could you add type hints to arguments in the functions?
• Make sure all doc strings are filled out fully, including: descriptions of what the argument is, any shape information, shape information with explicit variable names
• Given the complexity of this model, it needs a lot more tests beyond the defaults in the testing suite. In particular for implementations of loss averaging and NMS
• N is used everywhere - it should be replaced with something more explicit
• As this PR is so necessarily big, I would try to find ways to make it as small as possible and then have follow up PRs for additional functionality. In particular, it would be a lot easier to review without the conditional onnx logic. Could you remove this and then add in a separate PR please?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.