Hao
hi @kamauz , the priors/anchors needed in your model and the way of branching out paths to detection head might be different.
@kamauz @notabigfish you can also change the number of channels of extra layers to make the network output is consistent with the generated anchors.
没找到LOG。有时间我会更新代码，顺便把新的LOG传上去。一般认为，LOSS不再显著减少时，就可以停止了。你也可以用VALIDATION 数据作为标准。
depends on the value of scale.
Hi @wcwang07 , "CocoDetection" returns "image" and "targets" instead of "image", "boxes", "labels". You can split "targets" into "boxes" and "labels" to make it work. Plz see "coco.loadAnns" to get...
@AlanStark I didn't test the code in multiple GPU environment. https://github.com/pytorch/examples/tree/master/imagenet may be used as a reference. Good luck!
Thanks @hyl-g . That's a big bug indeed. Can make a pull request if it is not too much trouble? So I can merge it.
@HongChow 训练集和新的图片集里含公式的图片或区域差别比较大？
@nightzsky not very high. I don't remember the exact number.
> > @nightzsky not very high. I don't remember the exact number. > > do you remember if it is >0.5? Not sure. But feel free to use this script...