티스토리 뷰
contrastive loss와 같이 loss function 에 의한 발전에 속한다고 한다. (특히 val_loss에 해당)
- Anchor(a), Positive(p),Negative(n) 세가지 input을 사용
L = max( d(a,p)-d(a,n) + margin,0)
- positive는 가깝고 negative는 멀어지게 되는데 margin이 더해져 있어서(?) negative를 멀리 보낸다고 한다.
Triplet mining
- Triplet loss에 따라 dataset의 카테고리를 3종류로 (Easy, Hard, Semi-hard) 나눌 수 있다.
- loss=0이면 easy triplet이다. *) d(a,p)+margin < d(a,n)
- negative가 positive보다 anchor에 가까우면 hard다. *) d(a,n) < d(a,p)
- negative가 positive 보다 anchor에서 멀리떨어져 있긴 한데 margin 보다는 아닌, loss > 0인상태면 semi-hard다.
- d(a,p) < d(a,n) < d(a,p) + margin
그림으로 표현하면 이렇다고 한다.
data를 이런 식으로 triplet을 만들어 학습하면, 차별화를 명확하게 한다고 한다. 단점은 학습과정이 느리고 트리플셋이 균형잡혀있어야 해서 데이터가 많이 필요하다고 한다.
출처 :
https://soobarkbar.tistory.com/43
+chatGPT4.0
'AI > 주워들은 것들' 카테고리의 다른 글
LLM의 RAG란 무엇인고 (0) | 2024.11.16 |
---|---|
torch)메모리 부족 이슈로 Data loader가 뻑사리 날 때 insufficient shared memory (1) | 2024.10.23 |
CI-UDA 읽은 후기 (0) | 2024.03.30 |
soft label vs pseudo label (0) | 2024.03.25 |
Domain Gereralization과 Domain Adaptation 차이 (0) | 2024.03.23 |
공지사항
최근에 올라온 글
최근에 달린 댓글
- Total
- Today
- Yesterday
링크
TAG
- textual guidance
- 첫논문
- max-margin_loss
- /var/log/tallylog
- Deformable Part Model
- ctfloader
- 샴네트워크
- 도메인 어댑테이션
- /var/log/fontconfig.log
- Group Acitivty Recognition
- alternatives.log
- Deformable Convolutional Networks
- /var/log/alternatives.log
- StyleCLIP
- 딥페이크탐지
- yolov3 #tensorflow_convert
- 1-stage detection
- /var/log/faillog
- Chatgpt4
- Source-free unsupervised domain adaptation
- batch란
- neural network applications
- torchmodel
- 전이학습
- TransferLearning
- tf2to1
- Arbitrary image stylization
- DetectoRS
- tf version porting
- tensorflow convert
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 | 31 |
글 보관함