Support | official |
---|---|
Language | Python 3.6+ |
Legion provides officially supported :term:`Toolchain Train Integration` with MLflow - Legion MLflow.
Installation instructions are available on official GitHub page.
This integration allows to train models, written in Python with usage of MLflow API, and to convert them into :term:`General Python Prediction Interface` (that is :term:`Trained Model Binary Format`).
Warning
To use train models, written using these integration, please ensure, that appropriate :term:`Toolchain Train Integration` has been installed on a platform (for cloud usage).
Installation instructions are available on official GitHub page.
- Legion supports all model flavours, that have Python flavor (e.g. keras, sklearn and etc.). MLeap flavour is not supported;
- Only Python programming language version 3 is supported;
- MLproject has to use conda environment management, all required packages (python and system) has to be declared in conda environment file;
- Training code may save only one model in one MLproject entry point, otherwise exception will be raised;
- To allow using names of input / output columns, artifacts named
head_input.pkl
andhead_output.pkl
have to be saved in artifact's folder; - Legion executes exact one entry point during :term:`Model Training process`;
- Direct usage of MLflow client inside model training code is undesirable;