For ages timm (https://github.com/rwightman/pytorch-image-models) had 2 models that were at or past 87% top-1 on #ImageNet. Today there's ~18 of them. Prepping a big update while finishing off some experiments and cleanup, I'm sitting on another 20+ (9 > 88%) weights trained / fine-tuned in timm or ported from elswhere. ImageNet-1k is heading in the direction of #CIFAR -- a side benchmark for adaptation and eval.
@rwightman timm is such a breath of fresh air! It not only works, but it is very pleasant to read through and explore.