Allow non-tensor labels in SanitizeBoundingBoxes and SanitizeKeyPoints

[YeonwooSung]

Description

Honor the existing TODOs in SanitizeBoundingBoxes and SanitizeKeyPoints: labels no longer need to be tensors. They only need to be bool-indexable, or sequences that can be subset with the validity mask.

labels_getter may now return:

• a single torch.Tensor, numpy.ndarray, or sequence (list/tuple) of per-box / per-keypoint values
• a tuple/list of such entries (for sanitizing extra fields such as COCO iscrowd or area)

Length is taken from entry.shape[0] when a leading dimension exists, otherwise len(entry). Tensors and ndarrays are indexed with the 1-D bool validity mask; lists and tuples are reconstructed as the same type, keeping only the valid items. None entries and non-indexable objects (for example a bare int) are still rejected.

Box and mask sanitization is unchanged. Sequence labels are treated as pytree leaves so they are subset as a whole rather than flattened into individual values.

Test plan

• python -m pytest test/test_transforms_v2.py -k "SanitizeBoundingBoxes or SanitizeKeyPoints or sanitize" -q --tb=short
• Covers existing tensor labels, list/tuple labels, numpy int arrays, mixed multi-entry labels, and the updated error cases for non-indexable / None entries.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9656

• 📄 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.