I use the edl loss to train in mini-imagenet dataset with 64 classes, but the loss can't converge and the accuracy is very low.