Silver Medal winning solution for SIIM-ISIC Melanoma Classification Challenge 2020

• Training a single model. See notebook here
  • Different variants of efficient net backbones pretrained on imagenet was finetuned. Training was done with the help of Colab TPUs.
  • Datasets are loaded from the Google Cloud Storage Buckets with the help of address to the dataset.
  • Stratified KFold Cross Validation was used in order to see th performance of the datasets.
  • Augmentations used - colour constancy, rotae/shear/soom, random masks
  • Different combinations of datasets were used for better variation/ generalization.
• Some hyperparameter searches done.
  • Dataset - ISIC2020, ISIC2020+ISIC2019
  • Efnet variants - B0 to B7
  • Optimizers - Nadam, Adam, SGD
  • Loss - BinaryCrossEntropyLoss with/without label smoothing, Focal loss, Sigmoid focal loss
  • 5Fold/ 15 Fold Cross Validation
  • Epochs - 10 to 30
  • Retrain - with/without pseudo labels
• Ensembling the base models.
  • Light Gradient Boosting with Bayesian optimization. See notebook here