Loss WiSE_UB on batch

[cryax]

Hi I have another question about Loss of WiSE_UB why don't you divide both BCE and KLD by batch but only BCE? (like VAE)

[yoyololicon]

I want to weight 2 losses exactly just like VAE so their performances can be comparable.
I know it slightly deviate from the original formula but whether divide KLD by batch or not didn't show much differences, in my experiments.

[cryax]

I see, so the current weights in WiSE_UB make the two losses comparable