I see the log file in your repo suggests that there are a total of 7864 steps, and with a batch size of 64 (as mentioned in the config files for that log), it would mean you have a total of 7864*64 identities. Even if I add all the identities of all the datasets (VGGFace2 + Ethnicity) mentioned in the DECA paper, it amounts to a total of roughly 24k identities.
@yfeng95 could you please shed some light on this? Am I overlooking something?
I see the log file in your repo suggests that there are a total of 7864 steps, and with a batch size of 64 (as mentioned in the config files for that log), it would mean you have a total of 7864*64 identities. Even if I add all the identities of all the datasets (VGGFace2 + Ethnicity) mentioned in the DECA paper, it amounts to a total of roughly 24k identities.
@yfeng95 could you please shed some light on this? Am I overlooking something?