exp : LOSS

[quasar529]

Description

Default loss 인 CrossEntropyLoss외 Dice Loss, Focal Loss 등 과 2개 이상의 Loss 조합 실험

Details

• CrossEntropyLoss를 최적화하는 것이 iou를 maximize하는 것과 방향이 다르다고 생각.
• Semantic Segmentation Task 에서 자주 사용하는 Dice Loss, Focal Loss, Lovász Loss 사용해서 실험.
• Kaggle Solution에서 대다수의 경우 Loss를 2개 이상 사용하는 경우가 많아, 해당 방법 적용한다.
  • EX. CE+DICE, CE+Focal, Focal+DICE

Reference / Image

https://www.kaggle.com/code/sungjunghwan/loss-function-of-image-segmentation

---

HuBMAP - Hacking the Kidney 1st : https://www.kaggle.com/competitions/hubmap-kidney-segmentation/discussion/238198 SIIM-ACR Pneumothorax Segmentation 1st : https://www.kaggle.com/competitions/siim-acr-pneumothorax-segmentation/discussion/107824 Understanding Clouds from Satellite Images 1st :https://www.kaggle.com/competitions/understanding_cloud_organization/discussion/118080

HuBMAP + HPA - Hacking the Human Body(2022) 1st : https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/356201 7th : https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/354859 11th : https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/354701

UW-Madison GI Tract Image Segmentation(2022) 2nd : https://www.kaggle.com/competitions/uw-madison-gi-tract-image-segmentation/discussion/337400

[quasar529]

Default : CE

1. CE + DICE

• Paper Pack 에 대한 mIoU 상대적으로 높음

  2. CE + Focal

3. Lovasz + DICE

• 20 epochs 진행 후 valid mIoU 0.2로 낮음. 구체적인 조정이 필요해보임.