u-net-model-with-submission

[dg4271]

https://www.kaggle.com/hmendonca/u-net-model-with-submission

[dg4271]

상위권 discussion

• https://www.kaggle.com/c/airbus-ship-detection/discussion/71782 (6th)

  • Performance by models
  • Ensemble of 4 models + TTA for ResNet34 + U-Net. Private LB: 0.846 -> 0.848 with Classification
  • (deleting masks for confident empty images)
  • Ensemble of 3 models + TTA for ResNet152 + U-Net. Private LB: 0.775 -> 0.848 with Classification
  • Ensemble of 3 models + TTA for Mask R-CNN. Private LB: 0.843 -> 0.850 with Classification
  • Final Ensemble (As a final ensemble we used:)
  • Geometric mean of 7 U-Net models including one model with pseudolabels and one 2nd level model on OOF predictions. Denote predictions of this ensemble as unet_mask
  • Ensemble of 3 Mask R-CNN models

• https://www.kaggle.com/c/airbus-ship-detection/discussion/71595 (9th)

  • 2018 dsb 1위 코드를 사용했다고 함(https://www.kaggle.com/c/data-science-bowl-2018/discussion/54741) 그리고 이 사람이 본 컴피티션의 2위

[dg4271]

Test Time Augmentataion

• Data augmentation

  • train에서 모델에게 데이터를 변형한(rotation, zoom, shift, flip..) 이미지를 계속해서 보여줌으로써 모델을 generalize 하는 방법

• TTA (Test Time Augmentation)

  • Test에서 모델의 성능을 향상시키는 방법
  • 원본 test 이미지 뿐만 아니라 이를 변형한 이미지들에 대해서도 prediction을 진행하고, 여러 번의 prediction 결과를 평균내 최종 prediction을 결정하는 방식
  • prediction ensemble

• 참고: https://towardsdatascience.com/test-time-augmentation-tta-and-how-to-perform-it-with-keras-4ac19b67fb4d

  • https://machinelearningmastery.com/how-to-use-test-time-augmentation-to-improve-model-performance-for-image-classification/

• TTA 적용 전

• TTA 적용 후

  • TTA 사용해서 private LB 점수를 10% 가까이 올릴 수 있었음

[dg4271]

Classification and Segmentation

• https://www.kaggle.com/hmendonca/classification-and-segmentation-fp