search

lucabergamini / VAEGAN-PYTORCH

VAEGAN from "Autoencoding beyond pixels using a learned similarity metric" implemented in Pytorch. Clean, clear and with comments.
182 stars 34 forks source link

About trend of reconstruct loss #2

Open yasuobinggan opened 4 years ago

yasuobinggan commented 4 years ago

Thank you so much for your code! I have a question about reconstruct loss, I try to according to your code, and paper to write code, but with you because of my own to build the network structure cannot get the same initialization, through the experiment discovered the decrease in the reconstruction loss is the basic of my code, and your reconstruct loss is rising, I believe you generate results better, I want to ask, reconstruct loss decreasing is the phenomenon of right?

yzcv commented 3 years ago

and your reconstruct loss is rising, I believe you generate results better, I want to ask, reconstruct loss decreasing is the phenomenon of right?

Hi,

I am studying this kind of problem too. I think for your case, the decrease in reconstruction error is fine as long as the total loss is decreasing. It is probably a matter of optimization order.

Best