[models] Add YOLOv10-N

[frgfm]

Why this model

YOLOv10-N is a 2.3M-parameter detector with end-to-end, NMS-free inference. It complements YOLO-MS-XS by testing a smaller architecture and a distinct consistent-dual-assignment training design.

Paper-reported evidence

These are external COCO results reported by the authors, not Holocron benchmark results.

The official recipe trains on COCO at 640 × 640 for 500 epochs with batch size 256. The model is trained from scratch and uses one-to-many and one-to-one assignments during training, then the one-to-one branch for NMS-free inference.

Source	Model	Dataset/protocol	Params	Compute	Metric	Score
Paper	YOLOv10-N	COCO, 640px, 500 epochs, scratch	2.3M	6.7G	AP	38.5%
Holocron current	YOLOv2	VOC2012, 416px, 40 epochs, frozen backbone	50.65M	Not recorded	Loc/Clf/Det error at IoU 0.5	83.09% / 52.82% / 92.02%
Future Holocron sibling	YOLO-MS-XS	Controlled VOC2012 protocol	Pending	Pending	Loc/Clf/Det error at IoU 0.5	Pending implementation

COCO AP and Holocron's VOC error rates are different metrics and must not be compared numerically.

Existing Holocron comparison

YOLOv2 is the primary current baseline and the only detector with a recorded Holocron run. Its localization, classification, and end-to-end values are error rates, so lower is better. YOLO-MS-XS becomes the second compact comparison after its implementation; this issue must not invent or pre-fill that score.

Holocron implementation benchmark

Use the same controlled VOC2012 campaign as YOLO-MS-XS:

• 416 × 416 inputs for 40 epochs.
• Effective batch size 64, with gradient accumulation when needed.
• TAdam at 5e-4 with OneCycle.
• Scratch initialization for both the candidate and YOLOv2 baseline; no frozen backbone.
• The same verified YOLOv2 controlled run may be referenced by both detector issues.
• Record localization, classification, and detection error at IoU 0.5, parameters, MACs at 416px, peak CUDA VRAM, throughput, hardware/software versions, and elapsed time.

Source	Model	Dataset/protocol	Params	Compute	Metric	Score
Holocron implementation	YOLOv10-N	VOC2012, 416px, 40 epochs, scratch	Pending	Pending	Loc/Clf/Det error @ 0.5	Pending implementation
Controlled baseline	YOLOv2	Identical VOC2012 scratch run	Pending rerun	Pending	Loc/Clf/Det error @ 0.5	Pending rerun
Lightweight sibling	YOLO-MS-XS	Identical VOC2012 run	Pending	Pending	Loc/Clf/Det error @ 0.5	Pending implementation

COCO support and standard COCO AP evaluation are out of scope.

Public API and implementation

• Add a yolov10_n factory and export; do not add larger variants.
• Implement the one-to-many and one-to-one training branches plus consistent dual assignment.
• Use the one-to-one branch for deterministic top-k, NMS-free inference.
• Preserve Holocron's target, loss-dictionary, and prediction-dictionary contracts.
• Reuse YOLO-MS components only where their operations and semantics are demonstrably identical.
• Do not introduce a generic detector framework.

Acceptance criteria

• Tests cover both assignment branches, ordinary and empty targets, finite losses, and backward propagation.
• Evaluation tests prove deterministic top-k output and absence of NMS.
• Default configuration matches the paper parameter count within rounding tolerance.
• ONNX opset-20 export covers raw predictions and the end-to-end path where the exporter supports it.
• The controlled YOLOv10-N and YOLOv2 benchmark rows are recorded with full metadata.
• YOLO-MS-XS is added to the comparison only after its measured result exists.
• Documentation keeps paper COCO AP separate from Holocron VOC errors.
• Checkpoint publication remains separate.

Provenance and license boundary

• Paper: https://proceedings.neurips.cc/paper_files/paper/2024/hash/c34ddd05eb089991f06f3c5dc36836e0-Abstract-Conference.html
• Official implementation: https://github.com/THU-MIG/yolov10
• The reference repository is AGPL-3.0. Implement independently from the paper: do not copy source or weights.
• Related detection roadmap: [models] Implement latest detection SOTA models #164