- Data topology
- MLflow compatibility layer
Model registry is the component in the MLOps lifecycle responsible for managing model versions. Beyond tracking just artifacts, it is responsible to track the metadata associated to each model, like:
- Data lineage
All entities belong to a project, and only users with access to the project can interact with the entities.
- Holds general information about a model, like name and description.
- Each model as a default
Ml::Experimentwith the same name where candidates are logged to.
- Has many
- Is a version of the model.
- Links to a
Packages::Packagewith the same project, name, and version.
- Version must use semantic versioning.
- Collection of comparable
- A candidate to a model version.
- Can have many parameters (
Ml::CandidateParams), which are usually configuration variables passed to the training code.
- Can have many performance indicators (
- Can have many user defined metadata (
To make it easier for Data Scientists with GitLab Model registry, we provided a compatibility layer to MLflow client. We do not provide an MLflow instance with GitLab. Instead, GitLab itself acts as an instance of MLflow. This method stores data on the GitLab database, which improves user reliability and functionality. See the user documentation about the compatibility layer.
Some terms on MLflow are named differently in GitLab:
- An MLflow
Runis a GitLab
- An MLflow
Registered modelis a GitLab
To test the an script with MLflow with GitLab as the backend:
mkdir mlflow-compatibility cd mlflow-compatibility pip install mlflow jupyterlab
In the directory, create a Python file named
mlflow_test.pywith the following code:
import mlflow import os from mlflow.tracking import MlflowClient os.environ["MLFLOW_TRACKING_TOKEN"]='<TOKEN>' os.environ["MLFLOW_TRACKING_URI"]='<your gitlab endpoint>/api/v4/projects/<your project id>/ml/mlflow' client = MlflowClient() client.create_experiment("My first experiment")
Run the script:
Go to the project
/-/ml/experiments. An experiment should have been created.
You can edit the script to call the client methods we are trying to implement. See GitLab Model experiments example for a more complete example.