yolov7自己训练的权重结果不对，但是官方的权重跑下来没问题，这是什么原因，预处理什么都一样的

wang-xinyu / tensorrtx

Implementation of popular deep learning networks with TensorRT network definition API

MIT License

6.93k stars 1.77k forks source link

yolov7自己训练的权重结果不对，但是官方的权重跑下来没问题，这是什么原因，预处理什么都一样的 #1198

Closed Cute-Yang closed 1 year ago

Cute-Yang commented 1 year ago

Env

GPU, e.g. V100, RTX2080, TX2, Xavier NX, Nano, etc.
OS, e.g. Ubuntu16.04, Win10, etc.
Cuda version
TensorRT version

About this repo

which branch/tag/commit are you using?
which model? yolov5, retinaface?

Your problem

what is your command? e.g. sudo ./yolov5 -s
what's your output?
what output do you expect?

wang-xinyu commented 1 year ago

是否试过checkout到v0.1分支再训练？

需要用cfg/deploy/中的yaml文件，不支持cfg/training/中的yaml文件

xyt2008 commented 1 year ago

是否试过checkout到v0.1分支再训练？

需要用cfg/deploy/中的yaml文件，不支持cfg/training/中的yaml文件

我用最新的yolov7版本训练yolov7 w6，精度和曲线都正常，但是转不了tensorrt模型；用v0.1版本yolov7 w6同样的数据，精度特别差。官方给的v0.1版本yolov7 w6精度正常。用的是cfg/deploy/中的yaml文件。最新版本训练曲线： v0.1版本训练曲线： v0.1版本训练参数：

xyt2008 commented 1 year ago

uyolo-ai commented 1 year ago

可能是没有重参数化，转wts模型之前需要重参数化（reparameterization），可以参考https://github.com/WongKinYiu/yolov7/blob/main/tools/reparameterization.ipynb

stale[bot] commented 1 year ago

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.