APTOS2019

43th place (top2%) Solution for Kaggle APTOS 2019 Blindness Detection

• inference code
• detailed summary

General

This is a not bad solution to get top2% without TTA or coefficient optimization.

Our Solution

Data Augumentation

• Introduce 2015 Diabetic Retinopathy competition data
• Conduct regular transformations that create less black padding
  • do_flip
  • flip_vert
  • max_zoom

Preprocessing

• Thanks to the @Neuron Engineer, we refer to his APTOS [UpdatedV14] Preprocessing- Ben's & Cropping, and set sigmaX=10

Pretrained Model

• We choose EfficientNet-PyTorch as our base model, this series model are quite accurate and fast to train.

Training

• Because this is a ordinal classification task, we train it as regression problem.
• We first pretrain model on 2015 data, then finetune on 2019 data

Ensemble

Stage 1

• Train efficientnet-b3, efficientnet-b4, efficientnet-b5 models on splitted 5-fold data resulting in 15 base models.

Stage 2

• Train xgboost, svr, catboost models on logits output of stage 1 base model.

Stage 3

• Bagging from stage 2 models