About inflated groupnorm usage.
chenerg opened this issue · comments
In the first training stage, no inflated groupnorm
is used, but in the second training stage you set use_inflated_groupnorm = true
. Why use two different settings? Should I set use_inflated_groupnorm = true
during training stage1? Also, can you give me any clue to learn more about the inflated groupnorm
?
AniPortrait/src/models/resnet.py
Line 21 in 8b435af
According to this function, it rearranges the batch dimension and frame dimension before applying group normalization. However, in stage 1, the frame dimension is consistently equal to 1. It would be not necessary to do inflated group norm here.