tip

• project 준비 방식
  • model(paperswithcode)
    1. EfficientNet →
    2. nfnet →
    3. timm (https://github.com/rwightman/pytorch-image-models/tree/master/timm) →
    4. ViT →
  • data → 데이터 불균형 문제를 어떻게 처리해야할지, mask dataset훨씬 많아서
    1. https://aihub.or.kr/aidata/27716
    2. TTA : https://inspaceai.github.io/2019/12/20/Test_Time_Augmentation_Review/
    3. data augmentation (https://arxiv.org/pdf/2009.08369.pdf)
    4. Fast augmentation (https://hoya012.github.io/blog/albumentation_tutorial/)

💡 idea

1. GPU도 좋으니 image size를 크게 가져가자 → 512x512(요즘 대세)

• Model 🦆