Can I substitute the bridge（topk，fc or 1x1conv） betwenn LQE and Regbranch with some simple operators?such as variance or some parameters which can describe DFL's shape
HAOCHENYE opened this issue · comments
Hi! I've been following gfl for some time, and it's a great work.
Actually, the uncertainty of bbox's distribution could produce some bad results sometime in gfl1. It seems that this problem could be optimized in gflv2. However， there are some special operators which could hardly be deployed in some hardware， such as topK.
In some way, gflv2 use the concentration of bbox's distribution to refine LQE, can I use some simple parameters, such as the variance of distribution to supervise LQE？ Will this work for gflv2 theoretically?
你好,我在工程中并没有到GFLv2相应的代码.想问下你找到了吗?
你好,我在工程中并没有到GFLv2相应的代码.想问下你找到了吗?
主体代码在mmdet models densehead/losses 下的gfocalxx下
Hi! I've been following gfl for some time, and it's a great work.
Actually, the uncertainty of bbox's distribution could produce some bad results sometime in gfl1. It seems that this problem could be optimized in gflv2. However， there are some special operators which could hardly be deployed in some hardware， such as topK.
In some way, gflv2 use the concentration of bbox's distribution to refine LQE, can I use some simple parameters, such as the variance of distribution to supervise LQE？ Will this work for gflv2 theoretically?
是的，TopK虽然基本上在所有主流框架中都很常见，但是在芯片上目前还支持的不太好。根据实验结果，一般Top2就比较有效果了，不知道芯片这块啥时候能支持Top2这样的操作。。。其实Top2距离普遍支持的max_pool也就多一个max的操作。。。
我们的ablation study跑了variance，貌似效果不明显，如果光使用variance涨幅是比较小的。
Hi! I've been following gfl for some time, and it's a great work.
Actually, the uncertainty of bbox's distribution could produce some bad results sometime in gfl1. It seems that this problem could be optimized in gflv2. However， there are some special operators which could hardly be deployed in some hardware， such as topK.
In some way, gflv2 use the concentration of bbox's distribution to refine LQE, can I use some simple parameters, such as the variance of distribution to supervise LQE？ Will this work for gflv2 theoretically?是的，TopK虽然基本上在所有主流框架中都很常见，但是在芯片上目前还支持的不太好。根据实验结果，一般Top2就比较有效果了，不知道芯片这块啥时候能支持Top2这样的操作。。。其实Top2距离普遍支持的max_pool也就多一个max的操作。。。
我们的ablation study跑了variance，貌似效果不明显，如果光使用variance涨幅是比较小的。
好的吧~。如果实在不行就手写一下后处理了，也就俩Conv1X1，应该也不会太麻烦！！，感谢翔哥的工作哈，要是提点了就替一替。
话说还想请教一丢丢炼丹心得：如果用SGD去learning from scratch初始学习率设置有木有啥参考呀。痛不欲生ING，想用cosine learning rate结果loss下降的让我虎躯一震，虎目含泪。或者说是我epoch设置的不够多(180)？