Expected performance for the DomainBed implementation about fishr HOT 2 CLOSED

Thank you for your interest.
Yes, these numbers were obtained with a similar implementation, with default hyperparameters from DomainBed and - most importantly - with the default architecture, a ResNet-50. In this case, the overhead (in the last layer) is relatively very cheap compared to the computation in the features extractor - in memory and time with a GPU. In your case, your relatively-larger overhead is certainly due to your smaller architecture - and perhaps also a little by your larger batch size (64 rather than 32 in my computation). I hope this answer helps you.

from fishr.

Thank you for taking the time to respond. That provides me with some more certainty that I'm doing the right thing here! Much appreciated.

from fishr.