leixy

Hi, @ChenJoya I was reproduced your "Optimal Bias Initialization" on yolov3-voc. It was amazing. Improved mAP about 1%. Thanks for your work. Would you release codes of your another paper...

the detail is very simple: 1. only init bias for last conv befor yolo; 2. init channels corresponding with objectness and classes. 3. I estimated N and Nf very **roughly**....

@xia00100 Sorry, I can't relase codes without authorization from the company.

@plz583585760 我这边之前按照默认配置，复现acc是没有问题的，loss没在意具体是多少；后来就换了backbone，做QAT了。

> > @plz583585760 我这边之前按照默认配置，复现acc是没有问题的，loss没在意具体是多少；后来就换了backbone，做QAT了。 > > 为什么我复现的时候chunksizes只能设置为1才能训练，我用的是windows系统，显卡是3060，训练完后准确率只有85左右，所以很困惑不知道哪里出现了问题，您的chunksizes设置为多少啊？ chunksizes没改，设的是16；我一般都是Ubuntu训练，windows系统没办法把GPU占满，不过这个模型训练不需要那么多的显存就是了