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