yabk / lumen Goto Github PK

Replace dummy predict function on the backend with a real model. We should be able to load the model when starting the backend container via environment variable.

Experiment with training last 2 layers of ResNet instead of only the last one.

Last year winners also used transfer learning. Only for the first 25 epochs they trained the last layer and the rest of training they trained the whole model. Here is their documentation.

Deploy app and set up CI/CD pipeline for automatic deployments.

Write project documentation

Write technical documentation

Write README for the deployment of both the frontend and the backend. Maybe include an example docker-compose.yml.

Cleanup and upload notebooks used for data processing and model learning to the repository.

Do a grid search of all the necessary model and learning hyperparameters. EarlyStopping could be useful for this.