Comment by preetum

This is the exact comparison we make in the paper. We have subsampled ImageNet experiments as well; see Figure 3 in the full paper: https://arxiv.org/abs/2010.08127