Adds YOLO-Master-EsMoE-N LoRA adaptation examples for brain-tumor and VisDrone

[Alanxtl]

Summary

Fixes #50.

This PR adds YOLO-Master-EsMoE-N LoRA adaptation examples for two vertical scenarios: dense aerial detection on VisDrone and sparse medical detection on brain-tumor. It includes scenario-specific LoRA configs, a reusable rank-sweep script, recorded sweep results, and a guide covering rank selection, target-module choices, and common dataset pitfalls.

Edits to main code

ultralytics/validator.py is currently missing import for torch_distributed_zero_first and LOCAL_RANK. Missing inplement for convert_ndjson_to_yolo_if_needed

I add them

Changes

• Added yolo_master_visdrone_lora.yaml for dense aerial VisDrone LoRA fine-tuning.
• Added yolo_master_brain_tumor_lora.yaml for sparse brain-tumor LoRA fine-tuning.
• Added run_yolo_master_lora_rank_sweep.py to run rank sweeps with r=4,8,16.
• Added yolo_master_lora_rank_sweep_results.csv with recorded metrics, trainable params, training time, and peak VRAM.
• Added yolo_master_lora_scenarios_README.md with:
  • scenario comparison
  • rank recommendations
  • LoRA target-module guidance
  • routing/gating exclusion rationale for MoE
  • common pitfalls for medical and aerial datasets
• Updated the LoRA examples README to reference the new YOLO-Master scenarios.
• Added NDJSON-to-YOLO conversion helper wiring used by validation utilities.

LoRA Design Notes

Both scenario configs expose the requested LoRA controls:

• lora_r
• lora_alpha
• lora_use_rslora
• lora_target_modules
• lora_include_attention
• lora_gradient_checkpointing

For YOLO-Master-EsMoE-N, the default recipes adapt convolution and expert-projection modules while excluding router/gating layers:

lora_exclude_modules: ["router", "routing", "gate", "gating"]

This keeps expert routing stable during short few-shot fine-tuning runs. Router LoRA can still be evaluated separately as an ablation.

Experiments

Rank sweeps were run within the 20-50 epoch constraint.

Scene	Rank	Epochs	Fraction	mAP50-95	Best epoch	Trainable params	Adapter params	Train time	Peak VRAM
brain-tumor	4	40	1.0	0.27129	36	468,290	123,116	13.31 min	3.20 GB
brain-tumor	8	40	1.0	0.31259	40	596,782	251,608	13.74 min	3.21 GB
brain-tumor	16	40	1.0	0.33535	37	848,390	503,216	14.82 min	3.30 GB
VisDrone	4	30	0.25	0.01239	20	472,410	123,116	34.82 min	12.50 GB
VisDrone	8	30	0.25	0.0149	23	600,902	251,608	35.12 min	12.50 GB
VisDrone	16	30	0.25	0.02615	26	852,510	503,216	36.04 min	12.60 GB