ZLH

Hi, have you solved that?

@shoutOutYangJie Hi, why close it

@shoutOutYangJie Can you leave your mail address please. I find some implementation errors in your previously released code and I hope to communicate with you.

@shoutOutYangJie It seems that GitHub blocks your email, I can only see *@* : ) You can find my email address in my GitHub homepage. Send me an email please~

I dont think I used any tricks training the deformable convnet. I suggest you print the “offsets” of every iter and see if they are reasonable.

> > I dont think I used any tricks training the deformable convnet. I suggest you print the “offsets” of every iter and see if they are reasonable. > >...

@louielu1027 以前跑过，会比这个低1-2个点左右。不过我观察了v1学习的offset和v2学到的offset+mask，offset普遍偏小（1e-1~1e-3） 个人感觉voc上不是很能体现出deformable的优势。因为在stage3、stage4加入dcn会使性能下降，但是在coco上就不一样了，更多的dcn会使性能提升。（从论文实验结果可以看出来，其实v2比v1的提升主要因为使用了更多的dcn，而不是mdcn本身。。）

@KangGrandesty Hi,I read your code, its cool but i have one question. In fractal_join_layer.cpp, I dont understand why "Dtype mult = Dtype(bottom_size)/Dtype(total_undrop_)" I think Dtype mult = Dtype(1)/Dtype(total_undrop_), that makes...

@zgplvyou 同样非常关注这种loss，请问你找到合适新版本的code或者自己解决了编译问题了吗？

@silver-rush 3q