Available at https://retinopathy-classifier.herokuapp.com
The following were used for model training (see requirements.txt):
- fastai: version 1.0.52
- PyTorch: version 1.0.0
- Python: version 3.6
A SqueezeNet pretrained on the ImageNet dataset was used to train the classifier.
Training was done with Kaggle Kernels. Training history is provided in history.csv
The dataset came from the Diabetic Retinopathy Kaggle Competition, with the files cropped to remove any black space, and resized to a width of 1024 (and maintaining the aspect ratio), before being loaded in fastai.
The following were used for model deployment:
- Heroku (Free Dyno)
- Flask: version 1.0.2
- gunicorn
The dataset was hosted in Kaggle Datasets. Model progress (monitored by CSVLogger Callback in fastai) and saved models (saved by the SaveModelCallback in fastai) were outputted by the kernel.
The web app was based on work by Nidhin Pattaniyil and Reshama Shaikh