docs(lora): add YOLO-Master domain adaptation configs

[xie-guangzhen]

Description | 描述

Adds YOLO-Master-EsMoE-N LoRA adaptation examples for two visually different vertical scenarios requested in #50:

• VisDrone dense aerial detection
• brain-tumor sparse medical detection

The PR includes scenario-specific LoRA configs, a reproducible rank sweep launcher, and README guidance for rank selection, target module policy, MoE routing handling, and common pitfalls.

Related Issue | 关联 Issue

Fixes #50

Change Type | 修改类型

• Bug fix | Bug 修复
• New feature | 新功能
• Documentation update | 文档更新
• Training/example configuration | 训练/示例配置

Files Added | 新增文件

• examples/lora_examples/yolo_master_visdrone_lora.yaml
• examples/lora_examples/yolo_master_brain_tumor_lora.yaml
• examples/lora_examples/domain_lora_rank_sweep.py

Implementation Details | 实现说明

VisDrone

• Uses VisDrone.yaml
• Uses YOLO-Master v0.5 nano detection config: ultralytics/cfg/models/master/v0_5/det/yolo-master-n.yaml
• Default rank: r=16, alpha=32
• Enables lora_include_moe=True and lora_include_attention=True
• Uses larger imgsz=960, multi_scale=True, and max_det=500 for dense small-object aerial scenes

brain-tumor

• Uses brain-tumor.yaml
• Uses YOLO-Master v0.5 nano detection config: ultralytics/cfg/models/master/v0_5/det/yolo-master-n.yaml
• Default rank: r=8, alpha=16
• Enables lora_include_moe=True, keeps lora_include_attention=False
• Freezes BN and disables mosaic-style late scheduling for sparse medical data stability

MoE routing policy

Both configs intentionally do not explicitly target routing modules. The router controls top-k expert assignment and can overfit or collapse under short few-shot schedules. The configs adapt Conv/Linear feature transforms and MoE expert/projection paths while keeping routing decisions stable.

Rank Sweep | Rank 对比

Added domain_lora_rank_sweep.py to run the required r=4,8,16 comparisons:

python examples/lora_examples/domain_lora_rank_sweep.py --scenario visdrone --device 0
python examples/lora_examples/domain_lora_rank_sweep.py --scenario brain_tumor --device 0

The script launches one yolo train command per rank and writes a CSV summary including:

• mAP50-95
• train time
• peak memory when available from logs/results
• run directory
• trainable parameter note for trainer logs

The README includes the comparison table template and scenario recommendations.

Self-test Checklist | 自测清单

• Verified Python syntax for the sweep script
• Verified VisDrone dry-run commands for r=4,8,16
• Verified brain-tumor dry-run commands for r=4,8,16
• Verified diff has no whitespace errors

Commands run locally:

python3 -m py_compile examples/lora_examples/domain_lora_rank_sweep.py
python3 examples/lora_examples/domain_lora_rank_sweep.py --scenario visdrone --dry-run
python3 examples/lora_examples/domain_lora_rank_sweep.py --scenario brain_tumor --dry-run
git diff --check

[xie-guangzhen]

Superseded by #70 after the valid RhinoBird claim window opened on July 1.