search

tamarott / SinGAN

Official pytorch implementation of the paper: "SinGAN: Learning a Generative Model from a Single Natural Image"
https://tamarott.github.io/SinGAN.htm
Other
3.31k stars 611 forks source link

Questions regarding the loss trends #57

Open junkwhinger opened 4 years ago

junkwhinger commented 4 years ago

Hi, first of all, thank you for sharing your wonderful project!

I'm running an experiment with my own natural images atm. I got curious with the generator and discriminator loss trends, so I logged them with tensorboardX.

Here are some of the results: image

As far as I know, if GAN is considered to be trained well, the generator loss goes down when the discriminator loss goes up and converge. Some errG and errD show such trend, but others seem to fail.

The output results such as random sampling and editing are not of bad quality. Do you thinkI can improve the quality of the generated outputs by improving those training failures? (i don't know how tho)

Thank you!

tamarott commented 4 years ago

If you wish to, you can play with all the hyper parameters (e.g. number of channels, number of layers, etc.) and monitor the learning curves.

junkwhinger commented 4 years ago

Thank you :) Can I safely assume that my model is not properly trained from the loss trends? Or maybe it's not the best idea to evaluate SinGAN performance based on the loss trends..