Shouldn't the upsample blocks of the Generator in models.py be added per powers of 2? The current implementation adds them per multiples of 2, which will only work for upsample=2 and upsample=4.
For example, to move from an image of size 128 to an image of size 1024 you need upsample=8 and three upsample blocks: 256,512,1024. which is lg(8)
The current implementation creates four blocks: 256,512,1024,2048 which is 8/2
Shouldn't the upsample blocks of the
Generator
inmodels.py
be added per powers of 2? The current implementation adds them per multiples of 2, which will only work for upsample=2 and upsample=4.For example, to move from an image of size
128
to an image of size1024
you needupsample=8
and three upsample blocks: 256,512,1024. which is lg(8)The current implementation creates four blocks: 256,512,1024,2048 which is 8/2