Need help in understanding #1 and #16

soumith / ganhacks

starter from "How to Train a GAN?" at NIPS2016

11.47k stars 1.67k forks source link

Open miqbal23 opened 5 years ago

miqbal23 commented 5 years ago

Hi, I have question regarding some points in there

regarding point 1, is there a reason behind normalizing the input into (-1,1) scale? Is it because the usage of tanh as generator output?
regarding point 16, I see lots of GANs still use method similar to concatenation eg. CGAN use pure concat while StarGAN expands it further to depth-concat. What makes using Embedding layer recommended that other conditioning methods?

shaurov2253 commented 4 years ago

On the NIPS 2016, while discussing trick#16, Soumith mentions the name of Michel Matthew. I am looking for what he referred to. Can anyone help?

miqbal23 commented 4 years ago

On the NIPS 2016, while discussing trick#16, Soumith mentions the name of Michel Matthew. I am looking for what he referred to. Can anyone help?

That's Michael Mathieu, from Deepmind