Nice work ! based on the benchmarks in the paper and also this repository, it seems that all ImageNet-1K benchmarks use self-supervised pre-trained checkpoints. If this is the case, I would like to ask if authors can provide ImageNet-1K benchmarks for models that are trained from scratch. This enables direct comparison to other models which don't use self-supervised pre-training which constitute a large fraction of the literature.
Nice work ! based on the benchmarks in the paper and also this repository, it seems that all ImageNet-1K benchmarks use self-supervised pre-trained checkpoints. If this is the case, I would like to ask if authors can provide ImageNet-1K benchmarks for models that are trained from scratch. This enables direct comparison to other models which don't use self-supervised pre-training which constitute a large fraction of the literature.