Effective Robustness against Natural Distribution Shifts for Models with Different Training Data