// the find
bubbliiiing/yolov8-pytorch
这是一个yolov8-pytorch的仓库，可以用于训练自己的数据集。
A from-scratch PyTorch reimplementation of YOLOv8 by bubbliiiing, who has done the same for v3 through v7. Aimed at people who want to train custom object detectors on VOC-format datasets without the Ultralytics CLI wrapping everything. If you want to understand what's actually happening inside YOLOv8's training loop, this is more readable than the official repo.
The code is straightforward Python without the layers of abstraction in the official ultralytics package — nets/yolo.py and nets/backbone.py are readable without a map. Supports mixed-precision training, EMA, cosine/step LR schedules, and multi-GPU out of the box. COCO mAP numbers are close to official (yolov8_s hits 44.1 mAP 0.5:0.95 vs Ultralytics' 44.9), so it's not a degraded port. The FAQ doc (常见问题汇总.md) covers the most common shape-mismatch and CUDA errors, which saves an afternoon of debugging.
Last commit was August 2023 — it predates YOLOv8's pose, OBB, and instance segmentation heads entirely. The requirements file lists torch==1.2.0, which is three years outdated and will cause silent incompatibilities; the actual minimum that works with AMP is around 1.7. Primary weight downloads are behind Baidu Netdisk, which is inaccessible without a Chinese phone number (GitHub Releases has them too, but the README doesn't lead with that). No ONNX or TensorRT export is built in, so production deployment requires you to wire that up yourself.