Notice
Recent Posts
Recent Comments
Link
| 일 | 월 | 화 | 수 | 목 | 금 | 토 |
|---|---|---|---|---|---|---|
| 1 | 2 | 3 | 4 | 5 | 6 | |
| 7 | 8 | 9 | 10 | 11 | 12 | 13 |
| 14 | 15 | 16 | 17 | 18 | 19 | 20 |
| 21 | 22 | 23 | 24 | 25 | 26 | 27 |
| 28 | 29 | 30 |
Tags
- nlp
- Computer Vision
- cp949에러
- 운영체제
- Error
- annoy에러
- clip코드
- nlp논문
- self-Attention
- 에어비앤비 추천인
- 파이썬 warnings
- 에어비앤비 호스트등록
- 파이썬 ignore
- 음성 모델
- BERT논문리뷰
- 텍스트 생성
- opencv
- 파이썬 NLTK
- huggingface clip
- 파이썬 BLEU
- label smoothing
- 에어비앤비 호스트
- open ai whisper
- clip text encoder
- 숙박업 추천인
- 단어 예측 방법
- 언어모델 디코딩 전략
- Attention
- 자연어처리 논문
- transformer
Archives
- Today
- Total
yeon's
Label smoothing 본문
Attention Is All You Nedd(2017) 5.training 파트에 label smoothing 에 관한 내용이 두 줄 나온다.
나는 처음보는 내용이고, 언제나 처럼 어정쩡하게 그냥 넘어갈 뻔 했지만 박사님께서 집어주셔서 알고 넘어가야 할 내용같아 기록해둔다.

Label이 A와 B인 binary classification task의 경우에 아래와 같이 0또는1이 아니라 smooth하게 부여하는 방법이다.
| Label | A | B |
| predict | 1 | 0 |
| Label | A | B |
| predict | 0.9 | 0.1 |
장점은
1. label을 0 또는 1이 아니라 smooth하게 부여해서 잘못된 loss의 영향을 줄일 수 있고,
2. mislabeling 된 데이터를 고려하기 위해 사용되는 것에 추가적으로 regularization에 도움을 주며 model generalization과 calibration에 도움이 된다고 한다.
'Life hack' 카테고리의 다른 글
| Warning: Permanently added 'node000' (000) to the list of known hosts. 오류 해결 (0) | 2024.09.30 |
|---|---|
| Python 텍스트 파일 read error (cp949) (0) | 2023.02.08 |
| annoy 설치 (0) | 2023.02.08 |
| 파이썬 경고 메시지 숨기기, 보이기 (Jupyter notebook, ipython) (0) | 2023.02.05 |
| [PyTorch] Class nn.Module (0) | 2023.01.18 |
Comments