Rethinking Two Consensuses of the Transferability in Deep Learning