Python Code adapted from https://github.com/sveitser/kaggle_diabetic, the Team o_O solution for the Kaggle Diabetic Retinopathy Detection Challenge.

Substract mean and divide by std of all images. No augmentation at first, since data is quite large enough.

Use 20% of data to tune parameters for CNN, since it takes more than 1 hours per epochs for full data , customer the F1 score. Approach the multi-label problem as regression problem

Training a big CNN from scratch is time-consuming, 1.fine tuning with pre-trained model and extract features from the layers fc6, fc7 with Alexnet 2. average the features for data with same label. 3. train a linear svm to do the classification

This method achieves 0.81 F1 score on private LB.