glow-pytorch icon indicating copy to clipboard operation
glow-pytorch copied to clipboard

Published 20 hours ago • verified publisher icon rosinality

loss NAN

Open ledashuai opened this issue 3 years ago • 1 comments

In my experiment, when the data set is small, the experiment can run normally. When the number of data sets is large, the loss value always changes to Nan when there are thousands of epochs

ledashuai avatar Jan 04 '22 07:01 ledashuai

Glow training is not very stable. You can adjust training or model hyperparameters like batch sizes.

rosinality avatar Jan 04 '22 12:01 rosinality