Fake it till you make it: Learning transferable representations from synthetic ImageNet clones