Continue training from checkpoints reduced quality ?

[muhdlaziem]

im continuing my training from checkpoint. the quality of the output samples a bit reduced. Is it normal ?
This is last sample i stopped.

This is the first sample continue from the checkpoint

[ehsienmu]

Do you know how to solve it? I have the same problem. Thank you.

[abesmon]

bumping up! Having same issue here. It's not something super problematic, but makes me curious

[junisabot]

it is due to the fade in feature of stylegan, you need to keep track the latest alpha for resume training.
normally the alpha will be equal to 1 if you completed the training properly.
you may try changing the code of <models/GAN.py> <line 753>

alpha = ticker / fade_point if ticker <= fade_point else 1
to
alpha = 1