Tests for reaching close to referenced (blog, papers) accuracy on efficient net B0 specifically on cifar100 and imagenet tiny, with imagenet pretrained weights.
Collected hyperparams leading to performance within couple % of reported larger model variants on each.
See code comments for ref to sources.
Up to (k=1 early stopping, epoch 0-indexed, acc @1):
CIFAR100 - e5 train: 0.8904 val: 0.835 test: 0.8354
ImagenetTiny - e14 train: 0.8629 val: 0.7914 test: 0.797