Daconμμ μ§νλ λμ νκ²½ λ³νμ λ°λ₯Έ μλ¬Ό λ³ν΄ μ§λ¨ AI κ²½μ§λν μ λν μ½λμ λλ€. (colab directoryμ μ½λκ° μ μ 리 λμ΄μμ΅λλ€.)
- python==3.7.12
- albumentations==1.1.0
- numpy==1.19.5
- pandas==1.3.5
- cv2==4.1.2
- sklearn==1.0.2
- json==2.0.9
- torch==1.10.0+cu111
- timm==0.5.4
- transformers==4.16.2
- Augmentations: Transpose, Flip, Rotate, RandomBrightnessContrast, Cutmix ...
- Augmentations : λ§μ½ κΈΈμ΄κ° 500μ΄μμ΄λΌλ©΄ randomμΌλ‘ sampling(2μ©)
- MinMax Scaling
- resnext50_32x4d
- K-fold Cross Validation(k=5)
- Use Cutmix augmentation until epoch 15
- Save model's weight when score is highest
- K-fold Model Ensemble(Soft Voting)