MediaBrain-SJTU / FACT

MIT License
165 stars 22 forks source link

On the model selection #3

Open YBZh opened 3 years ago

YBZh commented 3 years ago

In your paper, you write: "we conduct the leave-one-domain-out evaluation. We train our model on the training splits and select the best model on the validation splits of all source domains"

But, in my opinion and according to the description in "https://github.com/facebookresearch/DomainBed", "select the best model on the validation splits of all source domains" is the "Training-domain validation" or "IIDAccuracySelectionMethod", rather than the " leave-one-domain-out evaluation".

What's your opiniton? Thanks for your reply.