PaddlePaddle / PaddleYOLO

🚀🚀🚀 YOLO series of PaddlePaddle implementation, PP-YOLOE+, RT-DETR, YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOv10, YOLOX, YOLOv5u, YOLOv7u, YOLOv6Lite, RTMDet and so on. 🚀🚀🚀
https://github.com/PaddlePaddle/PaddleYOLO
GNU General Public License v3.0
558 stars 134 forks source link

【飞桨黑客松 Hackathon + No.153】Yolov6 v3.0 论文复现 #114

Closed Yang-Changhui closed 1 year ago

Yang-Changhui commented 1 year ago

目前修复了一些训练时的函数的错误调用,已经可以进行训练, 通过比对训练日志,发现loss有些不同,目前还在寻找原因 配置文件运行:/configs/yolov6/yolov6_s_300e_coco.yml 训练3epoch的结果: 3-epoch

nemonameless commented 1 year ago

https://github.com/PaddlePaddle/PaddleYOLO/issues/104

Yang-Changhui commented 1 year ago

修改了task_aligned_assigner,优化器,权重初始化部分,目前训练20epoch结果: 0cb59e5bc1958c24353806f447655d3

Yang-Changhui commented 1 year ago

更新了EffiDeHead_fuseab代码

Yang-Changhui commented 1 year ago

修改后训练20epoch的结果,map一段时间下降还没解决: image

Yang-Changhui commented 1 year ago

目前训练23个epoch的结果,需要全量训练: image

Yang-Changhui commented 1 year ago

训练300epoch的结果为map 36.2,torch版本yolov6版本精度为map36.9,相差0.7;当使用预热损失的时候,精度会有提高 无标题

Yang-Changhui commented 1 year ago

使用PaddlePaddle2.4复现了Yolov6n模型的完整训练精度,coco数据集36.2和原作者相差0.7 链接:https://pan.baidu.com/s/1j8dGh91uQuuRmkVgnVLkIg?pwd=2p5j 提取码:2p5j