Can't get anything close to good results on a 1400 image dataset very similar to the Pokémon model

[soupmixer]

The model seems to not improve beyond FID~75 no matter what settings I try

what I tried so far :

cmap=65536 , cmax=1024 , batch=128 , lr=0.0002
cmap=65536 , cmax=1024 , batch=8 , lr=0.00003
cmap=32768 , cmax=512 , batch=10 , lr=0.00003
cmap=49152 , cmax=768 , batch=32 , lr=0.00008

all on fastgan config

fastgan_lite config is giving me errors about separable discs
stylegan2 config is giving me errors about missing dependencies (even after I install them)

[soupmixer]

In contrast, when training in Stylegan2-ADA , cmap=65536 improves FID end result , just makes training much much slower.

On Projected_Gan , it seemingly just makes things collapse

example :

this dataset on Stylegan2-ADA

cmap = 16384 -> FID = 63 , converge in 15 hours
cmap = 32768 -> FID = 49 , converge in 100 hours
cmap = 65536 -> FID = 38 , converge in 500 hours

In contrast , Projected-GAN stops improving after the first 2 hours of training and slowly collapses , with the best FID around 75 at all the settings I have tried

[soupmixer]

I want complex and diverse generations from the AI so I think for my use-case higher cmap is important

Any tips if my goal is to go for very high diversity and low FID , if I don't really care about the time it takes to train? (ideally sub 200 hours though , but I don't necessarily need the blazing fast speed of projected gan , just the diversity and quality improvements of it)

[soupmixer]

yet even with now attempts at cmap = 32768 and 16384 , I can't go below FID=75 and the results look horrible even after 1000kimg

[soupmixer]

@xl-sr
seeing you helped out others for improving their config for a specific dataset could you help me? (sorry if tagging you is an annoyance)

some samples of the real dataset :
https://drive.google.com/file/d/1-6tKFb7oMTj7UxDx5eID82qffGsXqSjv/view?usp=sharing

the results are basically random color spatters with some pattern , sometimes eyes or claws , but they look really bad even after 1000kimg and FID never goes below 75

[soupmixer]

so to sum up the issue :

dataset works fine with standard Stylegan2-ADA but I wanna try to improve it with Projected-Gan ,

fastgan config has very very poor results (FID 75+ and constant collapses) no matter what setting I tweak

fastgan_lite and stylegan2 config of Projected_GAN both give me errors I haven't yet managed to fix.

[soupmixer]

Okay after messing around with fastgan_lite config for almost 4 hours of trial and error

I managed to fix the original error but then it was saying "act1 layer is not defined"

I simply deleted act1 layer from the projector.py code and now it starts up fine

I fear now that because I deleted a crucial part of the fastagan_lite code , the result I'm gonna get will be bad.

Any fix for act1 attribute not defined in EfficientNet error message?

[soupmixer]

another update :
fastgan_lite is seemingly making genuine progress with FID = 68 at 200kimg
altough diversity is very low, for now, will report back later

but the root of the issue is still there , so I think my issue report should be interpreted as these problems :

• fastgan config doesn't work and collapses for everything early on @nom57 also mentioned this in his opened issue
• fastgan_lite is better, but it needs a lot of tinkering to even get running