How to evaluate out-of-distribution performance

[enor2017]

Thank you for your interesting work! I noticed the current scripts only support in-domain training and test set inference, such as the same object but unseen colors. May I know how I could evaluate out-of-distribution performance, as mentioned in your paper? Such as unseen objects and unseen attributes,

Thank you so much!