In my opinion, authors define L_{vlb} = L_0 + ... + L_T, not L_t.

[unl1002]

          In my opinion, authors define L_{vlb} = L_0 + ... + L_T, not L_t.

Thus, they may calculate the vlb loss with scale factor T (self.num_timestep).

Originally posted by @yhy258 in #114 (comment)

[unl1002]

so, which means we use L_t * T (self. num_timestep) to approximate L_ {vlb}?

[unl1002]