프로젝트 활동

  • LR Scheduler 실험

    • Base model: FPN efficientNet b1 pretrained
      • Optimizer
        • Lookahead (AdamP base)
      • Transform
        • HorizontalFlip
        • ShiftScaleRotate
        • RandomBrightnessContrast
        • RandomResizedCrop(512,512,scale = (0.5,0.8))
        • VerticalFlip
    • Cosine Annealing
      • T_max=50, eta_min=1e-8
      • epoch 23
        • mIoU: 0.5532
      • epoch 40
        • mIoU: 0.5785 (best)
    • Cosine Annealing Warm Restarts
      • T_max=50, eta_min=0
      • epoch 44
        • mIoU: 0.5097
      • epoch 50
        • mIoU: 0.5190
    • 결론
      • Cosine Annealing이 더 적합함.