SIIM-ISIC Melanoma Classification (6th rank approch)

Identify melanoma in lesion images

Team Members

• Pratik Bedre
• Gyanendra Das
• Saksham Aggarwal

Our Approach

We started this competition after 2 months after the start. For a baseline, we used @cdeotte's Triple Stratified Kfold With Tfrecords for tensorflow and @shonenkov 's Training CV Melanoma Starter for PyTorch. Thanks to these amazing kernels.

We used EfficientNet [B0-B6], Resnest,Resnext, with Sizes 192x192 256x256 384x384 512x512 768x768 384x512[HxW]

Summary

What Worked for Us

• Heavy TTA (X20)
• Cutmix
• Coarse dropout
• SWA(Stochastic Weight Averaging)
• Loss-Label Smoothing, BCE
• Optimizers - AdamW, Adam
• 2018, 2020 and malignant datasets
• 5 checkpoints' prediction averaging(stabalised our model's predictions)
• some models were trained with different height width ratios

What didn't Work for Us

• Loss functions-Focal loss, dice loss
• Optimizer- Ranger
• Hair removal/addition
• Pseudo labelling
• 2019 dataset
• Preprocessing techniques from Aptos Competition
• Progressive learning

Ensembling techniques

• Weighted average
• Power Average
• Minmax ensemble(didn't help)
  
  3hr before end of competition we came across rank ensembling and and we did this ensemble and got 0.9697 for our last submission

Our 3 final Submission

We new the shakeup was coming, so we tried to select different approaches.

1. All 15+ pytorch gpu solution models(with context) - 0.9530 (public LB) 0.9380 (private LB) 0.9541 (CV)
2. 15+ pytorch model (with context) and 15+ tf models (without context) - 0.9627 (public LB) 0.9470 (private LB) 0.9618 (CV)
3. Blend of public submission with 2nd submission with post proccessing technique - 0.9697 (public LB) 0.9126 (private LB) (overfitted) all the above were also ensembled with the meta only submission.

We wanted to give a shot to public lb overfitted submission but obviosly didn't work out well. I guess we were lucky enough to select the best private lb submission from our arsenel. We found the discussions and public kernels really fruitful and learnt a lot from this competition.