search

Tianxiaomo / pytorch-YOLOv4

PyTorch ,ONNX and TensorRT implementation of YOLOv4
Apache License 2.0
4.46k stars 1.49k forks source link

Can Only Achieve 26% AP with 608x608 on MSCOCO 2017? #489

Closed jielyu closed 2 years ago

jielyu commented 2 years ago

Can Only achieve 26% AP with 608x608 on MSCOCO 2017?

But the paper says, it can achieve 43% AP with 608x608.

XuecWu commented 2 years ago

请问您训练训练了多少个epoch？

jielyu commented 2 years ago

@XuecWu 50个epoch左右就基本收敛了，再训练下去就开始过拟合了

XuecWu commented 2 years ago

@jielyu 啊这，我现在还正在训练，我看YOLOX那些都得训练300个epoch，你超参数设定是啥啊兄弟，我是只加载主干权重训练，冰冻主干学习率为0.00125，不冰冻为0.000125，bs都是64

jielyu commented 2 years ago

@XuecWu 参数基本没有修改，按照这份代码跑的。一般是要训将近300epoch，但是才训到50就开始过拟合就不太对劲，所以停了。yolox初步试了一下，比较简单，损失下降也挺明显，没有训练到收敛就先停了，先学习一下损失函数的设计

XuecWu commented 2 years ago

@jielyu 训练到50个epoch拟合就离谱，太玄乎了==