search

anvoynov / GANLatentDiscovery

The authors official implementation of Unsupervised Discovery of Interpretable Directions in the GAN Latent Space
416 stars 52 forks source link

Arguments used for experiments reported in the paper #16

Closed chi0tzp closed 3 years ago

chi0tzp commented 3 years ago

Hi, could you provide the arguments that you have used for the experiments reported in the paper? For instance, a set of args.json and command.sh files. That would help in reproducing your results.

Thanks!

anvoynov commented 3 years ago

Hi @chi0tzp please check the args.json in */deformator/args.json in the pretrained models archives from https://github.com/anvoynov/GANLatentDiscovery/blob/master/download.py all the other parameters should be default so you can omit them. Hope this is helpful.

chi0tzp commented 3 years ago

Hi @anvoynov, thanks for your response. Just a quick question, is max_latent_ind the same as max_latent_dim? Because I cannot find any instance of max_latent_ind in the current version of the code (only in the args.json files).

anvoynov commented 3 years ago

Ah, sorry, indeed, these parameters were renamed after the refactoring. max_latent_ind now is directions_count

chi0tzp commented 3 years ago

Hi @anvoynov thanks for your response. Keep in mind though that max_latent_dim is still being used in various parts of the code, e.g. here and here.

anvoynov commented 3 years ago

max_latent_dim aims to bring new functionality that allows you to restrict the directions into a latent coordinates subspace. This could be useful if the latents are graduated. For instance, once you want to find only geometrical transformations in BigGAN, it could be reasonable to set max_latent_dim = 20 and optimize the directions that modify only the first 20 entries of z. max_latent_dim == G.dim_z is the default option that brings no additional effects.

chi0tzp commented 3 years ago

@anvoynov Thanks :)