Caching a stylegan2-ada model?

🎈 Using Streamlit
1

Hello,

Been poking around Streamlit for the first time with a GAN. I tried to use the Demo GAN app in the galley as a template. However, I am not experienced with caching, hashmaps, etc


This is what I have tried. I assume I need to move the call of ‘tflib.init_tf()’ to this hash_funcs. But nothing has worked.

TL_GAN_HASH_FUNCS = {
tflib.init_tf():id
}
@st.cache(allow_output_mutation=True, hash_funcs=TL_GAN_HASH_FUNCS)
def load_model(network_pkl):
print(‘Loading networks from “%s”
’ % network_pkl)
with dnnlib.util.open_url(network_pkl) as fp:
_G, _D, Gs = pickle.load(fp)
noise_vars = [var for name, var in Gs.components.synthesis.vars.items() if name.startswith(‘noise’)] Gs_kwargs = dnnlib.EasyDict()
Gs_kwargs.output_transform = dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=True)
Gs_kwargs.randomize_noise = False
return Gs, noise_vars, Gs_kwargs

Has anyone had luck caching their stylegan GANs?

image
image1176×504 51.2 KB

2

Hello Frewil,

I’m facing the same issue. Were you able to find any solution?

3

No, unfortunately. Keep me posted if you get it :pensive:

4

I have spent good quality time on this but I wasn’t able to find a solution. So, I changed the codebase to PyTorch, and is working well.

For your issue, u can use
hash_funcs={’_thread.RLock’ : lambda _: None,
‘_thread.lock’ : lambda _: None,
‘builtins.PyCapsule’: lambda _: None,
‘_io.TextIOWrapper’ : lambda _: None,
‘builtins.weakref’: lambda _: None,
‘builtins.dict’ : lambda _:None,
}
as the hash function which can solve but I’m getting errors after that. Hope this fixes ur code.

1 Like
5

thanks for the tip!

I think I may be too lazy to convert to pytorch. I’ll wait it out for an update