// the find
bubbliiiing/yolo3-pytorch
这是一个yolo3-pytorch的源码，可以用于训练自己的模型。
A PyTorch implementation of YOLOv3 for object detection, primarily targeting Chinese-speaking developers learning or fine-tuning YOLO models on custom datasets. It covers the full pipeline: data prep, training, prediction, and mAP evaluation, all in one repo.
Multi-GPU training support and adaptive learning rate scheduling based on batch size are included, which are often missing from educational implementations. The anchor recalculation script (kmeans_for_anchors.py) lets you retune priors for your specific dataset rather than reusing COCO anchors blindly. mAP 0.5:0.95 of 38.0 on COCO Val2017 is respectable for vanilla YOLOv3 at 416x416. The author maintains a consistent family of repos across YOLOv3 through v7, so the codebase style is familiar if you move between them.
The repo pins torch==1.2.0 in the README, which is from 2019 — requirements.txt claims broader compatibility but this is already a smell. YOLOv3 is genuinely obsolete for production use; YOLOv8, RT-DETR, or even YOLOv5 outperform it substantially with similar inference cost. Pretrained weights are hosted on Baidu Pan, which is inaccessible without a Chinese account — a real adoption barrier outside China. No ONNX or TorchScript export path is provided, so you're stuck with Python inference unless you wire it up yourself.