// the find
bubbliiiing/yolov4-pytorch
这是一个YoloV4-pytorch的源码，可以用于训练自己的模型。
A PyTorch implementation of YOLOv4 by bubbliiiing, one of several YOLO reimplementations this author maintains in parallel. Aimed at Chinese-speaking practitioners who want to train object detection on their own data with minimal setup — VOC format in, weights out.
The implementation is complete: CSPDarkNet53 backbone, SPP+PAN neck, Mosaic augmentation, CIoU loss, cosine LR decay, and multi-GPU training are all there. Pretrained weights hit 46.1 mAP on COCO, which is consistent with the original paper. The training pipeline is straightforward — edit a txt file, run voc_annotation.py, run train.py — which lowers the barrier for people who just want to fine-tune on custom data. The k-means anchor recalculation script (kmeans_for_anchors.py) is included, which is something many similar repos forget.
Pinned to torch==1.2.0, which is four major versions behind current PyTorch — you will hit compatibility issues immediately on any modern environment and there is no sign this will be updated. Last touched October 2023, and the original paper's era (2020) means you're two or three generations behind YOLOv8/YOLO11 in both accuracy and inference speed for most practical use cases. Weights and datasets are hosted on Baidu Pan, which is inaccessible without a Chinese account — a real blocker for anyone outside China. The codebase is in Chinese throughout (comments, README, issue tracker), so non-Chinese readers lose the explainability value that is presumably the main reason to pick this over Ultralytics.