Ponchine Tion

hi,i also meet this question, do you solve it?

i also find this problem,i think you need add this code: ` if 'str' not in str(type(mode)): mode = mode.decode()` @wizyoung

i might have the same problem .please save kid.

smaller loss may be batter.actually,in my experiment,the higher mAP Higher are basically has less loss.

大佬,我也在训练coco2017数据集.我有个问题想请教一下.作者的代码分两阶段训练,我发现开始第二个阶段训练的时候训练集测试集loss会突然上升,然后训练集会慢慢下降,但是测试集下降了几个epoch后就开始上升了,越来越大,请问是我哪里训练有问题么?

我是按照作者设置的config来运行的,不知道哪里有问题.您的config文件方便发给我看看么?

@zhishao 方便微信交流一下么.我的微信是 pc544556

@zhishao 你训练一个epoch要多久?

应该是加载数据方式的原因.我之前训练也是很慢,后面改了一下加载数据的方式,能快个3倍左右吧.

我用mobilenet v2训练过,mAP也很低,只有百分之46左右.