I found use frozen_stages&norm eval may improve transfer learning performance.
Currently，the timm wrapper not support these tricks.