Combining inherent knowledge of vision-language models with unsupervised domain adaptation through self-knowledge distillation.