GlossaryΒΆ

Code Repository

A code repository is a link to a GitLab/GitHub/Bitbucket project. This is the place where your machine learning code is stored and developed. The algorithm determines how data is processed and what predictions or classifications the system can make. We also support DVC. Hence, data can be versioned in a git project as well.

Connection

A Connection is a set of credentials to access some third party service, e.g. the JUWELS supercomputer.

Data Repository

A data repository is a reference to data that can be used for ML training and/or validation.

Experiment

An experiment is a collection of runs. It is a crucial aspect of MLOps, experiments involve tweaking parameters, hyperparameters, and data to optimize the performance of the ML application.

firecREST

firecREST is a software that allows access to an HPC site via a REST API. It is developed and maintained by the CSCS Swiss National Supercomputing Centre.

MLflow

MLflow is an open-source software for the Machine Learning lifecycle, e.g. experiment tracking and model versioning. For further details, see [](./mlflow/introduction.md).

MLflow Experiment Id

The MLflow experiment Id is the Id of that specific experiment on the Mlflow database. It is an integer and it is necessary for submitting a run to HPC.

MLproject

An MLflow Project is a format targeted to make ML projects reusable and reproducible. For further details, see here.

MLproject file

The MLproject file defines how an application can be executed via MLflow. For further details, see here.

Model

A model is the final output of an ML application and results from the experiments performed. The model is what makes predictions or decisions based on the input data.

Project

Mantik projects are the main organizational unit for your ML applications. A project is a collection that includes everything needed to reproduce your results, including code and data repositories, runs, experiments, and models.

Run

A run is the execution of code for a given data set on an external infrastructure. Usually it refers to training of a ML application.

UNICORE

UNICORE is a software that allows access to an HPC site via a REST API. It is developed and maintained by the JΓΌlich Supercomputing Centre (JSC).