About network parameter setting about pro_gan_pytorch

[akanimax]

Hi @deartonym,

I have made the number channels (i.e. the network size) a function of latent_size and depth. If you are encountering same number of channels for all layers, then perhaps your depth <= 4. If you wish to have an architecture same as that mentioned in the paper, please use the values: depth=9 and latent_size=512. You could just change the config values in the config_file in the published examples for this package.

BTW, I created an architecture which bypasses the whole complicated progressive training (thereby has much less hyperparameters) called MSG-GAN. MSG-GAN is also able to generate images at [1024 x 1024] resolution. Do check it out 😄.

I hope this helps

Cheers 🍻!
@akanimax

[deartonym]

Hi @deartonym,

I have made the number channels (i.e. the network size) a function of latent_size and depth. If you are encountering same number of channels for all layers, then perhaps your depth <= 4. If you wish to have an architecture same as that mentioned in the paper, please use the values: depth=9 and latent_size=512. You could just change the config values in the config_file in the published examples for this package.

BTW, I created an architecture which bypasses the whole complicated progressive training (thereby has much less hyperparameters) called MSG-GAN. MSG-GAN is also able to generate images at [1024 x 1024] resolution. Do check it out smile.

I hope this helps

Cheers beers!
@akanimax

Thanks for your explanation! I will go and check MSG-GAN.

[deartonym]

@akanimax
Thanks again!
The layers and parameters are exactly as described in the paper!