Skip to content

[Reproduce] Add VisDrone & SKU-110K training scripts and full results - #113

Open
Alan-20251210 wants to merge 2 commits into
Tencent:mainfrom
Alan-20251210:reproduce-visdrone-sku110k
Open

[Reproduce] Add VisDrone & SKU-110K training scripts and full results#113
Alan-20251210 wants to merge 2 commits into
Tencent:mainfrom
Alan-20251210:reproduce-visdrone-sku110k

Conversation

@Alan-20251210

Copy link
Copy Markdown

Closes #49

犀牛鸟任务:VisDrone & SKU-110K 复现结果

本 PR 新增了 VisDrone 和 SKU-110K 数据集上的训练脚本及完整复现结果(离线日志提供百度网盘下载)。

新增文件

  • scripts/reproduce/reproduce_visdrone.py:支持 --model v01 / moe,适配 Windows 多进程
  • scripts/reproduce/reproduce_sku110k.py:同上
  • README_reproduce.md:详细说明、训练命令、结果表格及百度网盘离线日志下载链接

复现结果

数据集 模型 输入分辨率 训练轮数 mAP50 mAP50-95 参数量
VisDrone YOLO-Master v0.1-N 800 120 0.360 0.213 7.5M
VisDrone YOLO-Master EsMoE-N 800 120 0.360 0.212 3.4M
SKU-110K YOLO-Master v0.1-N 640 120 0.885 0.564 7.5M
SKU-110K YOLO-Master EsMoE-N 640 120 0.886 0.563 3.4M

所有结果均经过 120 轮完整训练,使用 RTX 5060 Laptop GPU (8GB),具体超参见脚本内注释。

训练日志

  • 在线 Wandb(项目为私有团队,无权限访问将返回 404)
  • 离线日志(百度网盘):已在 README_reproduce.md 中提供永久下载链接及提取码,解压后 wandb sync 即可查看全部训练曲线。

本地复现命令

# VisDrone 基线
python scripts/reproduce/reproduce_visdrone.py --model v01

# VisDrone MoE
python scripts/reproduce/reproduce_visdrone.py --model moe

# SKU-110K 基线
python scripts/reproduce/reproduce_sku110k.py --model v01

# SKU-110K MoE
python scripts/reproduce/reproduce_sku110k.py --model moe

环境配置

  • CUDA Driver: 13.1
  • PyTorch 2.11.0 (CUDA 12.8)
  • 安装命令:
pip install torch==2.11.0 torchvision==0.26.0 torchaudio==2.11.0 --index-url https://download.pytorch.org/whl/cu128
pip install ultralytics wandb

已知问题说明

  • WandB 为 Private 团队,无权限用户访问链接会返回 404,请以上传的离线包为准。
  • Windows 训练需固定 workers=0,否则可能引发 I/O 死锁。
  • VisDrone 密集小目标场景下,imgsz=800 为推荐最低值,降低分辨率可能导致精度下降。


新增VisDrone与SKU-110K数据集复现训练脚本:
1. reproduce_visdrone.py:针对VisDrone数据集的模型训练、评估完整流程
2. reproduce_sku110k.py:SKU-110K商品检测数据集训练脚本
适配当前分支的实验复现流程,可直接运行训练
1. Add reproduce training scripts for VisDrone and SKU-110K datasets, support v0.1-N baseline and EsMoE-N model, compatible with Windows multi-process training.
2. Create README_reproduce.md with complete usage guide: dataset download steps, training commands, quantitative mAP comparison table, offline WandB log download link and environment configuration.
3. Remove large wandb zip packages from repository, store offline logs on Baidu Netdisk to avoid bloating repo storage.
4. Supplement known training issues, hardware environment parameters and dependency installation instructions for 8GB laptop GPU users.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

【2026犀牛鸟开源人才专属】【低难度】模型训练专项-垂类数据集基线训练

1 participant