XuKong
XuKong
> I have converted other datasets to VOC format, can I train it? if so.can you explain in detail how to modify which part of the code. Thank you very...
请教一下博主,如果batch size设置为4,输入分辨率缩放到800x800这样,去训练VOC数据集,测试mAP会不会很低。其实是我用tensorflow把您的代码写了一遍,在debug时用您的代码做了一下对比,然后生成的标签、以及损失值都是相同的,但我用上面的训练设置去训练VOC的时候,测试mAP只有0.65左右,所以想问一下是不是FCOS对batch size要求比较大呢。
> > > > > > > > 博主,你好,我在用VOC训练的时候,将num_class设置成20,然后去训练得到的map是0.63左右,然后您提供的78.7.pth测试的时候也需要将num_classes设置为80才能测试成功,这是什么原因呢 > > > > > > > > > > > > > > > 请问您找到原因了吗?我也遇到了同样的问题 > > > > >...
> Positive and negative classes have their own attention rpn. The targets of negative class rpn are all 0. But there are still some proposals generated by the negative class...
Thanks, I find my loss hard to drop,thanks for your advice, I will check my code.
> can you tell me the map on voc2007test? thank you! I had Writteen codes of tf version according to this codes,and trained/tested in voc07,got map about 0.67.
对了,我的resolution是640x640
> @JiXuKong 您好~很抱歉这么晚才回复您的issue,我做过一次实验,在voc上大概是78左右,输入时800x800。最近,我这个YOLOF项目已经实现好了,README有了更新,您若是还对YOLOF感兴趣,可以看一看,支持voc上的训练。我这边由于一些客观因素,更多的结果没法给出了,还望见谅~ thanks!