Model state completely collapses after a resume

AlexiaJM / RelativisticGAN

Code for replication of the paper "The relativistic discriminator: a key element missing from standard GAN"

719 stars 103 forks source link

Open FeepingCreature opened 6 years ago

FeepingCreature commented 6 years ago

For some reason, model quality completely falls apart after a resume. G/D_err stays stable, so idk. Something is still going wrong. :(

DanyangYue commented 4 years ago

Me too，the result is bad after being trained some iteration

fc392 commented 4 years ago

I have the same problem with RGAN without GP. . (RGAN-GP did not tested yet!)