Odahu provides a Trainer Extension for the popular MLflow framework.
This allows model Training in Python, and provides support for MLflow APIs. Trained models are
packaged using the General Python Prediction Interface.
- Odahu supports Python (v. 3) libraries (e.g. Keras, Sklearn, TensorFlow, etc.)
- MLeap is not supported
- Required packages (system and python) must be declared in a conda environment file
- Train must save only one model, using one MLproject entry point method. Otherwise an exception will occur
- Input and output columns should be mapped to the specially-named
head_output.pkl files to make it into the Packaged artifact
- Training code should avoid direct usage of MLflow client libraries
Source code is available on GitHub.
Low-level integration details are provided here.