Diabetic Retinopathy is a very common eye disease in people having diabetes. This disease can lead to blindness if not taken care of in early stages. Machine learning helps in automated diagnose but its lack of interpretability prevents people from fully trusing it. This project aims at obtaining a more robust and more interpretable model through adversarial training.
Kaggle provides a very hefty and diverse dataset that contains round about 30,000 images. You can download it from here.
-
Report_jiahuaWU_final.pdf: Report summarizing research methods and results.
-
resize_data.py: Contains code for image preprocessing using Graham's algorithm
-
training_experiment.py: Contains training script using sacred libarary for automatic metrics logging and parameters recording.
-
adversarial.py Pytorch implementation of adversarial examples generation using FGSM, PGD-attack and Boundary attack
-
data.py: Codes for data loading and dataset balancing using pytorch.
-
train.py: Codes for model training scheme.