- Move keras prediction to API (separate django or flask app)
- Setup GCP serving pipeline for API
- Create sequence for adding classes to ImageNet without disrupting decode_predictions for ResNet
- Make use of protobuff to speed up training
- Seamless integration of misclassified images to trigger training