Skip to content

DevIwon/Computer_Vision_Anomaly_Detection_Algorithm-1

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 

Repository files navigation

Computer_Vision_Anomaly_Detection_Algorithm

Dacon 2022 Computer Vision 이상치 탐지 알고리즘 경진대회

Team 'remember' : 이승리, 서원진, 최재홍

Private 66th, Score 0.8232 (66/481, 13.7%)


Development Environment

  • Google Colab pro

Train (Train.ipynb)

  • data split : 0.8/0.2(train/validation, stratify=label)

  • nomalization : mean=0.5, std=0.5

    ↳ Dataset으로 계산한 값으로 학습한 것보다 0.5로 학습한 것이 성능 더 높게 나옴

  • augmentation : ShiftScaleRotate, Rotate, VerticalFlip, HorizontalFlip (filp은 metal_nuts에서 제외)

    ↳ Mixup, CutMix, Sharpness, MedianBlur, IAAEmboss, CLAHE, RandomBrightness 등 시도

  • epoch : 100

  • lr : 0.001

  • loss : CrossEntropyLoss

  • optimizer : Adam

  • scheduler : ReduceLROnPlateau(patience=4)

  • model(public) :

    • swin_tiny_patch4_window7_224 (0.76627)
    • mixnet_s (0.72670)
    • efficientnetB2 (0.73043)
    • efficientnetB0 (0.75146)

    ↳ model마다 class weight 유무 실험 후 성능 높은 모델로 채택 → efficientnetB2는 class weight X

    ↳ 비교적 parameter 수가 작은 것에 성능이 잘 나옴

    ↳ pretrain model 사용

  • input size : 300(mixnet_s, efficientnetB2, efficientnetB0), 224(swin_tiny_patch4_window7_224

Inference (Inference.ipynb)

  • Ensemble(soft-voting)
  • Test Time Augmentation(TTA)
    • Rotate90(angles=[0, 90, 180, 270])
    • Multiply(factors=[0.9, 1, 1.1])

About

Dacon/Dacon/2022 Computer Vision 이상치 탐지 알고리즘 경진대회

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 100.0%