Comment by buildbot

Why only MNIST and a Graph CNN? Those are small and somewhat odd choices. Scale these days should be at least 100 million param models and something like OpenWebText as a dataset in my opinion. Not sure what the SoTA is for visionm but same argument there.