search

wgrathwohl / JEM

Project site for "Your Classifier is Secretly an Energy-Based Model and You Should Treat it Like One"
Apache License 2.0
415 stars 63 forks source link

Adapting JEM to high-resolution images #14

Closed Jun-Pu closed 1 year ago

Jun-Pu commented 1 year ago

Dear Will Grathwohl,

Thanks so much for the inspiring work!

Now I am trying to adapt your proposed joint energy-based model to more challenging dataset , where resized 256^2 ~ 512^2 images exist. As a result, the JEM does not converge well no matter what regularization methods (L2/L2 grad regularizations as described at https://github.com/wgrathwohl/JEM/issues/4 ) are used. Could you please provide with me some ideas for advancing JEM to high-resolution images?

Thanks in advance,

Best, Jun-Pu

chang-1 commented 4 months ago

Hi! Could you share some ideas on how to adapt the code to train on custom sizes? I tried increasing the self.last_dim parameter, but it takes too long to train on one epoch. Do you have any other suggestion? Thanks in advanced.