can we avoid having to train two separate models for both the mobile apps and the website? instead we could share a feature extractor, with one head factorized & therefore smaller for mobile, and another head unfactorized and therefore slightly more accurate. this should make it easier to keep web and mobile models in sync and reduce overall training time.
can we avoid having to train two separate models for both the mobile apps and the website? instead we could share a feature extractor, with one head factorized & therefore smaller for mobile, and another head unfactorized and therefore slightly more accurate. this should make it easier to keep web and mobile models in sync and reduce overall training time.