Question - Why save with avg_param?

HYOJINPARK / MC_GAN

MC-GAN: Multi-conditional Generative Adversarial Network for Image (BMVC 2018)

47 stars 4 forks source link

Question - Why save with avg_param? #3

Open gagandaroach opened 4 years ago

gagandaroach commented 4 years ago

Hello. Thank you for your great work.

Why do you save the model with the average G params?

https://github.com/HYOJINPARK/MC_GAN/blob/master/Model1/trainer.py#L449

Why not use the actual generator weights?

gagandaroach commented 4 years ago

I traced the process through StackGAN++ and subsequently this VaE design. The move is coined a "reparameterization trick" in sec 2.4 of the paper. It assists when the incoming noise is not entirely deterministic.

HYOJINPARK commented 4 years ago

Hi @gagandaroach Thanks for your asking Yes, I also follow StackGAN instruction and their code. As you mentioned it has a purpose for stable training. This is also written StackGAN++ paper in page4 line4