Add optional ultrafast COCO backend to detection references

[developer0hye]

Adds --coco-backend ultrafast to the Detection reference training/test-only scripts, coco_backend="ultrafast" to engine.evaluate(), and backend="ultrafast" to CocoEvaluator. The default remains pycocotools; dataset loading/transforms still use pycocotools. Install the optional evaluator with pip install "ultrafast-pycocotools>=0.1.11,<0.2".

Closes #9665. This is a draft for scope discussion; maintainer agreement is pending. I maintain ultrafast-pycocotools (BSD-2-Clause).

The native evaluator combines evaluation and accumulation. Its distributed path therefore gathers prepared predictions, keeps the first occurrence of each image ID in rank/batch order, sorts the IDs, and evaluates once during synchronization. It explicitly requests per-image records instead of trying to recompute native curves by assigning merged evalImgs. The existing pycocotools evaluation path is retained, with empty-batch/rank handling added. Backend imports are local to each instance, and the source dataset is copied before adaptation.

Validation

• 9 passed on macOS, Python 3.12 / Torch 2.14, and 9 passed on Linux, Python 3.12 / Torch 2.10 CPU.
• Exact bbox, segmentation and keypoint precision/recall/scores/stats and per-image matching records; false positives, tied scores, crowd/empty images, conflicting duplicate image IDs, repeated accumulation and pickle before synchronization.
• Two actual Gloo ranks, including an empty rank; distributed curves also match a single-process reference over unique images.
• CLI argument parsing and engine.evaluate() with fixed model outputs verify option propagation.
• Pinned ufmt/black/usort formatting checks and flake8 passed on Python 3.10; git diff --check passed.
• A focused CPU CI workflow runs these standalone reference-script tests on Python 3.10/3.12. Local tests use installed torchvision wheels; the full torchvision native build, model training and unrelated test suite were not run.

The README explains the different timing boundary: per-batch evaluator_time does not include ultrafast's final synchronization/evaluation work. No end-to-end speedup or memory reduction is claimed from standalone library benchmarks.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9666

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

This comment was automatically generated by Dr. CI and updates every 15 minutes.