Your 1,000-class image classifier drives training cross-entropy close to zero, and the gap between the correct class's logit and every other logit keeps growing.
Validation accuracy stopped improving a while ago, and most validation predictions, wrong ones included, get over 99% probability.