Aravind Eye Hospital in India provided a large set of retina images taken using fundus photography under a variety of imaging conditions. A clinician has rated each image for the severity of diabetic retinopathy on a scale of 0 to 4. The task of this competition was to train a network that can detect the level of severity.

I treated the task as a regression problem. I experimented with several Resnet, SE-Resnet, Resnet, ResNext and SE-Resnext models until I found out that Efficientnet model performs better than most of the pretrained models in this competition. Our final silver winning solution was an ensemble of 5 different Efficientnet models. Here is an overview:

• Preprocessing: Only Circle Cropping from this kernel.

• Augmentations: Horizontal and Vertical Flip, 360 rotation, zoom 20%-25%, lighting 50%. Used Fast.ai transformations.

• Training: Mostly followed this method. Did 5+10 training using 2015 data as train and 2019 as validation set. Then I split the 2019 data into 80:20 ratio using sklearn's Stratified Split function and finetuned the previously trained model over it for 15 epochs. I took the best model based on validation loss which didn't seem to be overfitted and used it for submission. I used a fixed seed so that I can compare different model performances over the same val set. Didn't use CV or TTA. I wrote a pytorch and a fastai version of my codes for the competition.

I mostly used my fastai codes throughout the competition. It requires a lot of work to cleanup my codes and add a detailed explanation of my workflow. I guess I will keep updating this repo.

Classification and Ordinal Regression

Here is the summary of my final models and their ensemble scores:

• Competition Link
• Pytorch Sample Code
• Weighted Kappa and Metric Optimizer Link
• Data augmentation library Albumentation example
• Preprocessing Technique (borrowed from last competition winner)
• Discussion on our strategy.
• All gold medal solutions and discussions

• torch
• torchvision
• fastai
• sklearn
• numpy
• tqdm

• Steps
  • Create a directory data/new_data and keep the 2019 APTOS competition data there.
  • Run cyclic_lr.py and find a suitable learning rate.
  • Run cd pytorch_version and then train.py. It will run a 5 fold cross-validation with 10 epochs for each fold.
  • You can experiment with albumentations augmentations by using DRDatasetAlbumentation as your Dataset class.
  • The code monitors both val_loss and kappa scores and saves model based on them. In my experience, kappa score is unstable and often doesn't seem to be correlated with val_loss. The safe option here is to choose and save your model based on the val_loss only.

The notebook fastai_version/Blindness_Detection.ipynb was written for training a model over both 2015 and 2019 datasets. This notebook contains a complete set of instruction and codes from downloading the datasets to train a model in 2 phases.

• Steps

Will be added soon