Faced with new and different data during testing, a model must adapt itself. We consider the setting of fully test-time adaptation, in which a supervised model confronts unlabeled test data from a different distribution, without the help of its labeled training data. We propose an entr