docs(reproduce): add vertical baseline scripts

[xie-guangzhen]

Description | 描述

Adds reproducible baseline scripts for issue #49 to compare YOLO-Master-v0.1-N and YOLO-Master-EsMoE-N on dense vertical datasets:

• VisDrone aerial dense small-object detection
• SKU-110K retail dense product detection

The scripts use the built-in dataset YAML files and release weights from YOLO-Master-v26.02, then collect final metrics from Ultralytics results.csv into per-dataset summary CSV files.

Related Issue | 关联 Issue

Fixes #49

Change Type | 修改类型

• Bug fix | Bug 修复
• New feature | 新功能
• Documentation update | 文档更新
• Reproduction scripts | 复现脚本

Files Added | 新增文件

• scripts/reproduce/reproduce_visdrone.py
• scripts/reproduce/reproduce_sku110k.py
• scripts/reproduce/_domain_baseline_common.py
• scripts/reproduce/README.md

Usage | 使用方式

python scripts/reproduce/reproduce_visdrone.py --epochs 100 --imgsz 640 --device 0 --wandb-project yolo-master-reproduce
python scripts/reproduce/reproduce_sku110k.py --epochs 100 --imgsz 640 --device 0 --wandb-project yolo-master-reproduce

Dry-run command inspection:

python scripts/reproduce/reproduce_visdrone.py --dry-run --epochs 100 --imgsz 640
python scripts/reproduce/reproduce_sku110k.py --dry-run --epochs 100 --imgsz 640

Logged Metrics | 日志指标

The scripts preserve full Ultralytics per-epoch logs in each run directory and summarize final values for:

• metrics/mAP50(B)
• metrics/mAP50-95(B)
• train/box_loss
• train/cls_loss
• train/moe_loss when emitted by the model
• val/box_loss
• val/cls_loss
• val/moe_loss when emitted by the model

W&B can be enabled with --wandb-project; the README includes a result table for adding public W&B URLs after full training.

Notes | 说明

This PR provides reproducible scripts, command templates, logging paths, and result table templates. It does not fabricate full 100-300 epoch training metrics in the repository.

Self-test Checklist | 自测清单

• Python syntax check passed
• VisDrone dry-run emits the expected two-model training commands
• SKU-110K dry-run emits the expected two-model training commands
• git diff --check passed

Commands run locally:

python3 -m py_compile scripts/reproduce/_domain_baseline_common.py scripts/reproduce/reproduce_visdrone.py scripts/reproduce/reproduce_sku110k.py
python3 scripts/reproduce/reproduce_visdrone.py --dry-run --epochs 100 --imgsz 640
python3 scripts/reproduce/reproduce_sku110k.py --dry-run --epochs 100 --imgsz 640
git diff --check

[xie-guangzhen]

Superseded by #71 after the valid RhinoBird claim window opened on July 1.