Hi,
I think this section of the readme (fewshot imagenet experiments), the resnet models are actually BiT-M models not a vanilla (trained from scratch) resnet model.
I think it corresponds to BiT-M models mentioned in Fig. 5 of the paper.

Could you please clarify this?
All the best.