Using batch size > 2 (half size > 1)

HsinYingLee / DRIT

Learning diverse image-to-image translation from unpaired data

844 stars 155 forks source link

Using batch size > 2 (half size > 1) #54

Closed chinglamchoi closed 3 years ago

chinglamchoi commented 3 years ago

Does DRIT and MSGAN work with batch size > 2, half size > 1? I ran some preliminary tests and there is zero diversity, full mode collapse when the aforementioned settings are used.

Can DRIT and MSGAN support / be modified to support these settings? Thanks.