Code and pre-trained models can reproduce results according
to the paper: https://arxiv.org/abs/2006.06666v1