MI Pipelines Overview¶
The production-type machine learning models generally require constant iterations due to data update and algorithm optimization. The MI Pipelines runs through the entire machine learning lifecycle, covering data preprocessing, model training, model registration, model deployment, model services, and operation monitoring.
The Enterprise Analytics Platform provides the MI Pipelines product for users to design and orchestrate pipelines, publish them to production and schedule them, as well as express them through DAG to achieve the automatic iterative publish of models. A wealth of operators are integrated in the offline design to support logical expressions such as conditional judgments and loop unfolding, and allow to complete offline experimental simulation. In the online production, it provides automatic scheduling, and can record, track and archive each running instance. It helps to achieve the end-to-end full-lifecycle management of production-type machine learning and the iterative optimization of models.
The composition and architecture of the MI Pipelines are shown in the figure below:
