Open sy00n opened 7 months ago
핵심: Training-free Adaptation(TFA), unified domain-aware contrastive state prompting template, Test-Time Adaptation(TTA)
Base Prompts
Contrastive-state
Domain-aware(DA) prompting
두 self-supervised discriminative tasks를 통해 w 최적화 1) original과 noise-corrupted toekn을 discriminate함 (CE loss)
2) anomaly localization 수행 : P를 pseudo lalel로 써서 encourage adaptor
최종 loss는 두 loss의 합임.
핵심: Training-free Adaptation(TFA), unified domain-aware contrastive state prompting template, Test-Time Adaptation(TTA)
Method
Training-free Adaptation(TFA)
Domain-aware State Prompting
Base Prompts
Contrastive-state
Domain-aware(DA) prompting
Test-time Adaptation(TTA)
두 self-supervised discriminative tasks를 통해 w 최적화 1) original과 noise-corrupted toekn을 discriminate함 (CE loss)![image](https://github.com/sy00n/DL_paper_review/assets/67910856/945d59dd-0eb5-466e-b751-9700cf165cd1)
2) anomaly localization 수행 : P를 pseudo lalel로 써서 encourage adaptor![image](https://github.com/sy00n/DL_paper_review/assets/67910856/c2011a0e-8ea5-40e4-89bf-982222a02a54)
최종 loss는 두 loss의 합임.![image](https://github.com/sy00n/DL_paper_review/assets/67910856/7f882015-a4fa-477c-be37-1e10aaf24896)