Performance on Vanilla GAN ?

[debasmitdas]

Did you trying using the Vanilla GAN before using W-GAN ? I tried your approach using Vanilla GAN on toy 2D data and it does not work and wanted to ask whether you faced similar problems.

[AntreasAntoniou]

I did try it. It did work for my toy problem but did not perform well on VGG Faces for example. LS-GAN and standard GANs worked very well on Omniglot.

On Fri, 2 Nov 2018 at 20:36, Debasmit Das [email protected] wrote:

Did you trying using the Vanilla GAN before using W-GAN ? I tried your approach using Vanilla GAN on toy 2D data and it does not work and wanted to ask whether you faced similar problems.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/AntreasAntoniou/DAGAN/issues/10, or mute the thread https://github.com/notifications/unsubscribe-auth/AKSuNo9oRWV62B7jN3TUjTsCWY_Fz0Dwks5urKyzgaJpZM4YMSwm .

[debasmitdas]

Thanks for the quick response. I am trying on 2D data with gaussian classes. I am trying to learn from 3 classes and trying generate samples for the 4th class for which only a single sample exists. The results look something like this - Here the 3 training classes are blue, red and black centered at (-3,3), (3,3,) and (0,0) and I am trying to generate samples from the 4th class using a sample centered at (-3,-3) but it would not generate the correct distribution

Any reason this toy problem won't work ? Less Data or Less no. of classes or anything.