Model Management (xvc model)

[iesahin]

Model management involves file versioning the model files, and associating the data and code that's used to train them. It can also automate deployment. It can associate models with arbitrary JSON documents for various annotation, like adding features or hyperparameters used to train the model.

We already have output file handling in pipelines, but these are files in general. Models build on top of this.

Some possible features

• Selection of models by their data (e.g. when xvc pipeline run produces a model, it can associate the model with the inputs of the pipeline.)
• Selection of models by their hyperparameter values (e.g. when xvc pipeline run produces a model, it can find params in the hyperparameter files and these can be added as annotations to model.)
• Adding annotations to models from code files. (e.g. In Python API, we can have facilities to record metadata about models and these can be tracked along with the models.)
• Selection of models by their performance and metrics, via metrics facility of the pipeline.