LUVLI loss cannot convergence
yeqi17 opened this issue · 3 comments
Thanks for the amazing work. I have tried to use the loss in the paper on the facial landmark project. However , I find the loss cannot convergence. In chapter 3.2.1 of the paper, if the predicted landmark's value equal to the target, the loss's T1 part will tends to negative infinity.(Σij will tends to zero).Could you please tell me the correct way to use the loss? ^_^
The code is now publicly available here. The code for the uncertainty in the loss is
https://github.com/abhi1kumar/LUVLi/blob/master/loss/gaussian_loss.py
Where is the code? @abhi1kumar
Where is the code? @abhi1kumar
I received an email from MERL saying that I should not re-host the repo since it violates MERL policies. So, I deleted the files from the repo in the next commit.
The README of the repo directs you to the MERL website. You should officially download the repo from the MERL website after submitting your details.