This repository is the 3rd place solution the sinusitis classifcation of KHD2020 based on sample code.
NOTE : It is unable to run this code on your local machine because the challenge was based on NSML infrastructure and the dataset was private. You have to customize the data_loader.py to run training code with your own dataset and specify the directory on DATASET_PATH argument.
Requirements : run pip install -r requirements.txt
Pre-processing : Check notebook to see specific pre-processing process.
- Zero-padding to 300x600
- Windowing
- Background reduction
- RoI crop to 224x224
- Min-Max scaling
We trained network using SGD optimizer with a momentum of 0.9 and decay of 0.3. However, we didn't apply weight decay on bias term. We used a decaying learning rate with a cosine annealing warm-up start method setting an initial learning rate to 0.0005 and a minimum rate as 5e-6. We used a small batch size of 8 to increase training stability and trained the network for 60 epochs. The sinusitis dataset has a severe class imbalance, therefore, we adopted class weights for loss function with the ratio 1:4:6:9. To avoid overfitting, we added dropout on the fully connected layer with 0.5 probability. We applied simple data augmentation techniques such as random rotation (-15~15 degress) and scaling (x0.85~1.15).
model/: network architecture definitions and training/validation core functionsutils/: image processing code, config argument parser, and data loading codemain_pytorch.py: code for model training and validation
- Main command :
nsml run -d 2020KHD_PNS -e main_pytorch.py -c 8 -a '--batch_size 64' - Submit command :
nsml submit NSML_ID/2020KHD_PNS/SESSION_NUM CHECKPOINT_NUM
- Main command :
python main_pytorch.py --DATASET_PATH YOUR_DATASET_DIR --batch_size 64