monoloxo

maybe different weights can get almost same loss when it is small,but different weights can get different map

you can change the network structure by change the cfg file,but can not load the original weights if you change the network structure,you need train it from stratch.

Object Detection dataset not this structure: `image_dir/ / /

是不是保存的那个模型loss是nan，我的是这种情况，之前yolov2也出现过，这时候就是没有框的。

有时候nan之前的上个模型是可用的。。。

same problems

可以看一下net/yolov2/predict.py中的相关函数

> 需要修改deepstream源码实现和本项目无关 是的，我查资料好像要修改deepstream-app这个主程序，至于到底修改哪里，就不清楚了。

> 我们修改完了 可以给我发邮件你的联系方式和你交流下思路 刚给你发了邮件，请大佬指教呀

> check this the nvidia [deepstream app example](https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/tree/master/runtime_source_add_delete) it is difficult for me,so i would like to get a easy way,but still thank you .