search

RobotPsychologist / bg_control

Improving short-term prandial blood glucose outcomes for people with type 1 diabetes, a complex disease that affects nearly 10 million people worldwide. We aim to leverage semi-supervised learning to identify unlabelled meals in time-series blood glucose data, develop meal-scoring functions, and explore causal machine-learning techniques.
https://blood-glucose-control.streamlit.app/
18 stars 42 forks source link

Dissemination - MLOps - MLFlow Integration into sktime annotators #98

Open RobotPsychologist opened 3 weeks ago

RobotPsychologist commented 2 weeks ago

Conditional on base class rework completion. Flavours in https://github.com/ml-toolkits/mlflavors https://www.sktime.net/en/latest/api_reference/deployment.html

fkiraly commented 2 weeks ago

For context, sktime has mlflow integration, originally developed by Benjamin Bluhm. He later included this in his mlflavours package which has mlflow integration for various time series packages, but it is no longer maintained.

There are a few "entry level" issues which might help getting into mlflow integration with sktime in general that might help understanding, I would suggest in this sequence - and these are not conditional on the base class consolidation:

  1. There is a tutorial in this wokrshop on the integration with forecasting: https://github.com/sktime/sktime-tutorial-europython-2023 - it would be great to check if it still works
  2. as a second step, I would suggest to (a) merge the tutorial into sktime, then (b) merge the sktime flavour in mlflavours back into sktime. Finally, (c) update the tutorial with the sktime flavour.

These two should give provide some user and developer experience with writing mlflow flavours, which then can be used to integrate the detection module.