SIIM-OSIC Melanoma Classification: 2nd Place

Environment:

• Python 3.7.7
• Anaconda
• PyTorch 1.6
• 4 NVIDIA Quadro RTX 6000 24GB

Setup Python environment

conda create -n melanoma python=3.7 pip
pip install -r requirements.txt

Download data

cd data/isic2019
bash download_isic2019.sh
unzip ISIC_2019_Training_Input.zip
cd ..
kaggle competitions download -c siim-isic-melanoma-classification
unzip siim-isic-melanoma-classification.zip 

[Optional] Download trained checkpoints

cd checkpoints
kaggle datasets download -d vaillant/melanoma-checkpoints
unzip melanoma-checkpoints.zip

Train ISIC 2019 model

cd src/etl
python 0_create_isic2019_splits.py
cd ..
python run.py configs/isic2019/mk001.yaml train --gpu 0,1,2,3 --num-workers 4

Run ISIC 2019 model on ISIC 2020 data

Please change model_checkpoints in src/configs/predict_isic2019.yaml if using a checkpoint with a different name.

cd src
python run.py configs/predict/predict_isic2019.yaml

Assign nevus labels to ISIC 2020 data

cd src/eval
python nevi.py
cd ../etl
python 10_combine_cdeotte_nevi_with_isic2019.py

Train ISIC 2019+2020 model

cd src
bash train_kfold.sh

Create pseudolabels

cd src
python run.py configs/predict/predict_bee_nometa.yaml predict_kfold \
    --gpu 0 --backbone tf_efficientnet_b6_ns \
    --model-config configs/bee/bee508.yaml \
    --checkpoint-dir ../checkpoints/bee508/tf_efficientnet_b6_ns/ \
    --save-file ../lb-predictions/bee508_5fold.pkl --num-workers 4
cd etl
python 12_make_pseudo_nometa.py

Train on pseudolabeled data

cd src
bash train_kfold_pseudolabel.sh

Inference on test set

cd src
bash inference.sh

Generate submission

cd src/eval
python generate_sub.py

Final submission CSV will be saved within eval/ directory as final_submission.csv.