Created by: chrschorn
Hi,
I have a question/possible fix: shouldn't the MUNIT style loss calculation be slightly different?
Before my proposed change, s_code_12
was compared against style_1
. However, s_code_12
is essentially calculated by s_code_12 = Enc2(Dec2(c_code_1, style_2)
(see lines 177, 181 and 190), i.e. created by style_2
. The cyclic translation result s_code_12
should then be compared against the original style code style_2
. Same applies to the other style loss.