pytorch_RVAE icon indicating copy to clipboard operation
pytorch_RVAE copied to clipboard

Loss Function

Open dongqian0206 opened this issue 6 years ago • 2 comments

Hi,

Would you like to tell me why you added scalar '79' in front of the cross entropy (RVAE, Line 110)? loss = 79 * cross_entropy + kld_coef(i) * kld

best

dongqian0206 avatar Jun 27 '18 08:06 dongqian0206

when I remove 79, kld can go to zero very fast.

nguyenvo09 avatar Sep 04 '18 14:09 nguyenvo09

Hi, thanks for your reply. We have to use kl annealing to alleviate kl vanishing issue in this model. I think it is the same.

best

dongqian0206 avatar Sep 04 '18 16:09 dongqian0206