Creative-Adversarial-Networks icon indicating copy to clipboard operation
Creative-Adversarial-Networks copied to clipboard

Published 20 hours ago • verified publisher icon mlberkeley

Gradient Explode

Open HsiaoYingLu opened this issue 4 years ago • 0 comments

Hi all,

I wonder if anyone encountered gradient explosion after training for a few epochs? I am trying to use the CAN losses on a sequence generation model. But, the gradient of generator just kept becoming too large for that the sequence generated can still be classified too well. For now, I kept training even if the gradients are too large every 2 epochs and I don't see the result of my model being anywhere near creative. Any thought on this are appreciated!

Thank you

HsiaoYingLu avatar Feb 08 '21 03:02 HsiaoYingLu