siim-isic-melanoma-2020

Pytorch lightning code for SIIM-ISIC Melanoma Classification https://www.kaggle.com/c/siim-isic-melanoma-classification/data

Requirements:

• python==3.7.7
• pytorch==1.5, torchvision (cuda==10.2), pytorch-lightining==0.7.6
• albumentations==0.4.5 (from sources)
• opencv==3.4.2, matplotlib==3.1.3, tensorboard
• pandas, jupyter, scikit-learn
• g++, apex (from sources)
• pip install efficientnet_pytorch

TODO:

• aug-mix
• AutoAugment (https://www.kaggle.com/nxrprime/siim-eda-augmentations-model-seresnet-unet/)
• other features & targets (MLP)
• • drop each feature is set to missing with this prob
• Kfold (stratified: https://www.kaggle.com/zzy990106/pytorch-5-fold-efficientnet-baseline/notebook?select=auto_augment.py)
• data echoing (https://arxiv.org/abs/1907.05550)
• try hyper-opt/optuna (focal loss weight, batch, emb-size)
• tta (multi_crop=5, flips, scale, )
• ensembling