search

denproc / Taming-VAEs

26 stars 4 forks source link

different constraints #1

Open Nintorac opened 6 years ago

Nintorac commented 6 years ago

hey, was wondering if you guys might have any insights on how to apply this method with non reconstruction losses

specifically I have a network with multiple losses modeled after https://arxiv.org/pdf/1807.03653.pdf and was hoping to apply GECO to it. Naievly squaring the loglikelihood and putting that through ends up with all losses increasing.

Long shot I know but any ideas would be appreciated

Nintorac commented 6 years ago

Seems like the constraint does not have to be the same as the loss function and additionally the constraint also does not need to be differentiable...