All results based on leave-one-out-validation instead of training-domain-validation?
BierOne opened this issue · comments
Hi, thanks for this outstanding work. I really appreciate this public code.
In your paper, you mentioned both the leave-one-out selection and training-domain-validation selection. However, as suggested by Domainbed paper (page 18), these are two distinct strategies for model selection.
Could you clarify your employed selection strategy? I am quite confused in this regard.
Thank you so much.
We used training-domain validation
model selection. leave-one-out cross-validation
in the paper indicates the evaluation scheme, not the model selection. We agree that this is easy to confuse, and honestly speaking, I think it would have been better if DomainBed had used a different term (the term leave-one-out cross-validation
have exist before DomainBed).
Got it. Thanks!