bubbliiiing / yolov4-tiny-pytorch

这是一个YoloV4-tiny-pytorch的源码,可以用于训练自己的模型。
MIT License
779 stars 185 forks source link

你好,val_loss为1.6,map=0是什么原因 #2

Closed Fafa-DL closed 2 years ago

Fafa-DL commented 4 years ago

用了作者YOLOv4跑出的val_loss是1.7,map=98.1%,然后今天用tiny对相同的数据集设置相同的训练参数最后map结果是0,不知道哪里有问题

bubbliiiing commented 4 years ago

0的话有点上头吧,你预测过有效果嘛

Fafa-DL commented 4 years ago

谢谢回复,已经解决,但是效果不好,只有58%

------------------ 原始邮件 ------------------ 发件人: Bubbliiiing <notifications@github.com> 发送时间: 2020年7月6日 15:18 收件人: bubbliiiing/yolov4-tiny-pytorch <yolov4-tiny-pytorch@noreply.github.com> 抄送: Fafa-DL <516451964@qq.com>, Author <author@noreply.github.com> 主题: 回复:[bubbliiiing/yolov4-tiny-pytorch] 你好,val_loss为1.6,map=0是什么原因 (#2)

bubbliiiing commented 4 years ago

毕竟只是tiny,参数量仅500W,你可以多训练一下,也可以关闭mosaic试试

Fafa-DL commented 4 years ago

好的,谢谢。

------------------ 原始邮件 ------------------ 发件人: Bubbliiiing <notifications@github.com> 发送时间: 2020年7月6日 15:21 收件人: bubbliiiing/yolov4-tiny-pytorch <yolov4-tiny-pytorch@noreply.github.com> 抄送: Fafa-DL <516451964@qq.com>, Author <author@noreply.github.com> 主题: 回复:[bubbliiiing/yolov4-tiny-pytorch] 你好,val_loss为1.6,map=0是什么原因 (#2)

毕竟只是tiny,参数量仅500W,你可以多训练一下,也可以关闭mosaic试试

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

Fafa-DL commented 4 years ago

大佬能把faster rcnn的vgg权重发我吗,百度云是真的太慢了

------------------ 原始邮件 ------------------ 发件人: "Bubbliiiing"<notifications@github.com>; 发送时间: 2020年7月6日(星期一) 下午3:21 收件人: "bubbliiiing/yolov4-tiny-pytorch"<yolov4-tiny-pytorch@noreply.github.com>; 抄送: "邹"<516451964@qq.com>;"Author"<author@noreply.github.com>; 主题: Re: [bubbliiiing/yolov4-tiny-pytorch] 你好,val_loss为1.6,map=0是什么原因 (#2)

毕竟只是tiny,参数量仅500W,你可以多训练一下,也可以关闭mosaic试试

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

Jasonshao0123 commented 2 years ago

请问map=0是怎么解决的呢?我替换了主干,修改了主干,训练之后loss=0.18,但是map为0

bubbliiiing commented 2 years ago

没有预训练吧