USPS Classification


I’m trying to make USPS classification.
I can get 99% Accuracy but loss is high.
I use Adam optimizer with 0.0001 learning rate and (0.5,0.9) betas.
My loss range is [0.02 ~ 58982]… This is very strange.

What should I do in this situation?


Does this loss contain the l2 regularization term? If so it is expected not to go to 0.
If the accuracy is 99% then you’re happy anyway no?

I use cross entropy loss.
Because the loss range is wide, I think this training is not converged well…
so the loss is very big, I can’t believe the auccracy.