dcgan.torch icon indicating copy to clipboard operation
dcgan.torch copied to clipboard

Published 20 hours ago • verified publisher icon soumith

Original hyperparameters

Open mescarra opened this issue 7 years ago • 1 comments

Hi!

When I try to use the original hyperparameters in the paper (ngf=ndf=128), the generator won't catch up with the discriminator and its error will get stuck forever. Is the Adam algorithm missing regularization?

Epoch: [1][      53 /      910]	 Time: 0.321  DataTime: 0.000    Err_G: 27.6310  Err_D: 0.0039	
Epoch: [1][      54 /      910]	 Time: 0.321  DataTime: 0.000    Err_G: 27.6310  Err_D: 0.0000	
Epoch: [1][      55 /      910]	 Time: 0.321  DataTime: 0.000    Err_G: 27.6310  Err_D: 0.0000	
Epoch: [1][      56 /      910]	 Time: 0.326  DataTime: 0.000    Err_G: 27.6310  Err_D: 0.0000	
Epoch: [1][      57 /      910]	 Time: 0.322  DataTime: 0.000    Err_G: 27.6310  Err_D: 0.0000	
Epoch: [1][      58 /      910]	 Time: 0.322  DataTime: 0.000    Err_G: 27.6310  Err_D: 0.0002	
Epoch: [1][      59 /      910]	 Time: 0.515  DataTime: 0.000    Err_G: 27.6310  Err_D: 0.0000	
Epoch: [1][      60 /      910]	 Time: 0.279  DataTime: 0.005    Err_G: 27.6310  Err_D: 0.0000	
Epoch: [1][      61 /      910]	 Time: 0.323  DataTime: 0.000    Err_G: 27.6310  Err_D: 0.0000

mescarra avatar Jul 25 '17 12:07 mescarra

This is a common problem in GANs. Try setting ngf = ndf * 4. As an example, if you have enough memory, try ngf = 256; ndf = 64

adam-hanna avatar Dec 31 '17 06:12 adam-hanna