DACON Monthly Computer Vision Anomaly Detection Algorithm Competition

안녕하세요. Team remember 입니다.
좋은 성적을 받진 못했지만 코드 공유를 해보고 싶어 업로드 합니다!

colab pro 환경이라 런타임이 자주 끊기는 탓에 k-fold는 시도해 보지 못해 아쉽습니다..

원래 Train 코드와 Inference코드가 따로 존재했는데 업로드가 파일 하나밖에 안되길래 그냥 붙여서 업로드했습니다.
원래 코드와 세부사항을 보고 싶으시다면 https://github.com/tmdrn9/Computer_Vision_Anomaly_Detection_Algorithm  을 참고해주시기 바랍니다.

다들 고생하셨습니다🤗

---

Dataset
- "metal_nut" class 중 filp class의 존재로 "metal_nut" 일 경우 filp이 적용되지 않게 데이터 셋을 커스터마이징하였습니다. 

Augmentation
-nomalize는 계산한 값으로 했을 때보다 0.5로 통일했을 때 성능이 더 높아 0.5로 실행하였습니다.
-albumentation 라이브러리 사용
-H/V Flip, Rotate, ShiftScaleRotate

Model
-swin_tiny_patch4_window7_224/input size:224/class weight 0
-efficientnet-b2/input size:300/class weight x
-efficientnet-b0/input size:300/class weight 0
-mixnet_s/input size:300/class weight 0