search

adamian98 / pulse

PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models
7.91k stars 1.5k forks source link

Train pulse on our own dataset. #56

Open 0xinit opened 4 years ago

0xinit commented 4 years ago

Hi,

Thanks for great work. Can you explain me a bit how I can train pulse on my own dataset? Which one of the files is training code? for some reason, I couldn't figure out. Thanks.

xjw00654 commented 4 years ago

this model is training when you do inference. If you want to get an idea result (maybe not such style), you need to train stylegan or stlegan2 with your dataset and replace pretrianed FFHQ model with your model.

0xinit commented 4 years ago

Hi @xjw00654 , yeah I understand that I may need to train stylegan. But what do you mean by "training when we do inference?"

xjw00654 commented 4 years ago

Hi @Gauravajariwal11. PULSE fixed the parameters of StygleGAN's generator but make the noise and the latent code the trainable variables. The "inference" phase is actually the "training" phase of the noise and latent code.

This solution is much like what stylegan2 does in projecting image to latent code

Maybe, I have some understanding mistakes : )

0xinit commented 4 years ago

@xjw00654 okay yeah. I think I understand now what you mean. Thanks a lot. I am just trying to train a dataset on pulse, and couldn't understand where the training code is.

James0329 commented 3 years ago

hey buddy . I got same problem. the pretrained model basicly from white/euro people. I put a Asian into realpics, output a European. Its too weird. the result not work on other human race. Need a way to train our own datasets.