Closed Kingthegarden closed 2 years ago
오 weight 초기화 방식으로 성능 향상 이끌어내신거 놀랍네요. 기대되네요.
현재 score는 5CV 평균 기준인가요? CAWR의 장점이 generalization도 있는걸로 알고있어서요^^
score는 LB 기준이고 CV에서도 높은 score가 나오는 걸로 확인했어요. CosineAnnealingLR, CosineAnnealingWarmRestarts 두개 같은 CFG로 실험해봤는데 CV, LB 모두 성적이 떨어지네요... scheduler를 좀 더 튜닝해서 실험해봐야겠습니다 !
LB 뿐아니라 CV에서도 좋은 결과가 있었다면 너무 좋은 결과인 것 같습니다! 수고하셨습니다 :)
모델 fine-tuning 결과
1. fc-layer
2. scheduler 변경
3. weight 초기화 방식 변경
해당 내용을 추가하여 pr 보내도록 하겠습니다.