Improving realistic semi-supervised learning with doubly robust estimation