Same exact model in keras but i just changed the loss from mse to mae. First one works, but the second just gives nans as predictions. Huber also gives predictions, but mae does not.
1 Answer
Removing clipnorm=1 from compile fixed it for me. Not sure why.