Domain Generalization by Mutual-Information Regularization with Pre-trained Models