Closed johntiger1 closed 4 years ago
This is not included in BigGAN-PyTorch. As that section indicates, BigGAN collapse is not successfully mitigated by any one of a variety of normalization, clipping, or clamping strategies, even though they successfully treat the symptoms of spectral explosion. (Spectral Explosion is also probably an attack from a ghost anime). If you want to try SV clamping for your own experiments, all the tools are provided if you subclass an SN
layer and implement eq (4).
Wow, thanks for the quick reply! Sure, but computing the singular values at every single update is the main challenge I would say. Was hoping to find the Arnoldi iteration method in Pytorch somewhere
EDIT: seems like your code in https://github.com/ajbrock/BigGAN-PyTorch/blob/a5557079924c3070b39e67f2eaea3a52c0fb72ab/layers.py
is most of the way there! Thanks!
I am looking for the code that does the clamping of the singular values from the weight matrices:
(i.e. page 6 of the arxiv paper)
but can't seem to find it in the training loop. Does anyone know where it is?
Thanks