Question about regression loss

[bearinsuke]

Hi,

Thanks for this great contribution.
I am curious why the last term of regression loss is 0.5 * r_log_var rather than 0.5* exp(r_log_var). The reason I am asking this is that when doing a test on this code the regression loss goes to a really big negative value. Do you have any ideas about this?

Thanks

[tonyzyl]

Dear bearinsuke,

Although I am not the author, I am working on modifying this regression model. You may find the derivation in this link useful. Hope it helps.