search

autonomousvision / stylegan-t

[ICML'23] StyleGAN-T: Unlocking the Power of GANs for Fast Large-Scale Text-to-Image Synthesis
Other
1.15k stars 53 forks source link

Hyperparameters not matched between code and paper #12

Open keyu-tian opened 1 year ago

keyu-tian commented 1 year ago

Thanks for your great work. In Table 4 from the paper EMA is 0.9978, but in the code generator.py line504 it's 0.995. Which is the actual value?

Also the batch size is 16 times smaller (128) than that in paper (2048).

ParthaEth commented 11 months ago

Agreed

ParthaEth commented 11 months ago

Specifically - clip_weight default set to 0 paper has different for light and full, Also number of residual blocks is reported in paper 3 but in code it's set to 2