Why is the weight of type_loss 50?

FengyangZhang / caffe_roomnet

Segnet-like CNN for room layout estimation

Other

27 stars 7 forks source link

Open Fang-Haoshu opened 7 years ago

Fang-Haoshu commented 7 years ago

Hi, I have a question. In the paper, the weight of type_loss is 5. Why are we using 50?

FengyangZhang commented 7 years ago

Hi, because there are two loss heads, so I though it might be reasonable if I set them to the same magnitude, so after calculation I chose 50.

kwangHouzz commented 7 years ago

Haoshu,

Which ratio did you use in the end? I am using 5 instead of 50 right now, and I think 5 is more reasonable.

Thanks!

Fang-Haoshu commented 7 years ago

Hi, I am using 5 now, but I don't see big differences...