I have a model that after 60 epochs the train error is larger than the train loss, is that normal:
epoch=59, train_loss_epoch=0.03931455929022359, train_error_epoch=0.9796516262755102, test_loss_epoch=1.836074846982956,test_error_epoch=0.8577952665441175
I didn’t think it was possible cuz I thought Cross Entropy (CE) was a upper bound on the error (0-1 loss). I think they call it convex relaxation. If this is true then shouldn’t the train error I am observing impossible? (I think it should also hold for the test loss vs test train).