Closed eps696 closed 4 years ago
Hi @eps696,
once the training completed, the run_train.py
script generates interpolation charts for each direction and three randomly sampled latents:
https://github.com/anvoynov/GANLatentDiscovery/blob/master/run_train.py#L94
it saves images to <out_dir>/charts_s*
with two different scale amplitudes. Apparently, this is the easiest way to inspect the learned directions.
The jupyter notebook is quite universal, there is no hand-picked constant. It only shows how to load and inspect the provided pretrained models. Once you want to use it for your run, you should update the passes in load_from_dir
call in the first block. The second block visualizes interpolations along a direction inspection_dim
for different latents. Just replace inspection_dim = annotated[0]
with an assign to a desired direction index.
@anvoynov thank you very much for prompt response. so, i load all found vectors/directions as
for i in range(512):
interpolation_deformed = interpolate(G, zs[i].unsqueeze(0), .. dim=i, deformator=deformator)
and look for those directions which make visual disentagled sense, right?
Yes, though (imho) it seems better to fix the latents zs
for all directions.
ah indeed, thanks for advice.
btw, is it max_latent_ind
that defines the amount of searched directions (K from the paper)?
(i've tested the code on StyleGAN2 model, and it explored 512 directions by default)
pardon me, seems i've been looking at some obsolete code; found directions_count
param on new updates
Yes, after the recent commit max_latent_ind
is directions_count
once other parameters are default.
i've completed the training on my stylegan2 model, and would like to explore the results - could you describe how to do that? how many directions have been found, how to check them explicitly, etc.?
provided jupyter notebook gives very little clues about that, it shows only your hand-picked results.