Model management in WeData is built on MLflow and Catalog, providing comprehensive lifecycle management capabilities for machine learning models, including:
Model registration: You can register trained models to the model repository.
Version control: It applies version control to models, facilitating the backtracking and comparison of different versions.
Lineage tracking: It displays the associations between a model and its upstream data and experiments, helping you trace the model's origins and dependencies.
Experiment tracking: It integrates with MLflow to automatically log model training parameters, evaluation metrics, runtime environments, and more, facilitating experiment comparison and performance monitoring.
Model deployment: It supports one-click deployment of models as online services, enabling real-time inference and application integration.
These features help teams collaborate efficiently, standardize the management and deployment of machine learning models, and enhance production efficiency and model governance capabilities.
Model Registration
Model interface
registration has two entry points:Method 1: Create a model directly on the Model Management page.
Method 2: Register a model based on a specific experiment run on the Experiment Run page.
Creating a Model on the Model Management Page
1. Go to the Model Management page from the left-side menu.
2. Click Create Model, select the model type "Machine Learning Model Registration", fill in the model name, select the catalog and schema path, and click Create. This step creates a model name and path in the catalog. (Note: Selecting a catalog address requires that the account already has authorization for the catalog data directory.)
3. Go to the Experiment Run page. Based on a specific model experiment run, click Model Registration, select "New Version of Existing Model" for the registration method, and select the model name you just created.
Creating a Model Based on an Experiment Run
1. Go to Experiment Management from the left-side menu, go to the best-performing experiment run, and click Model Registration.
2. Choose whether to register as a new model or as a new version of an existing model.
Register as a new model: Enter the model name, select the catalog and schema, and click Model Registration. (Note: Selecting a catalog address requires that the account already has authorization for the catalog data directory.)
Register as a new version of an existing model: Select the existing model and click Model Registration.
3. After successful registration, view the model in the Model Management list, click the model name, and go to the model details.
Version Management
1. Model registration is managed by version.
2. Click Associated Task to track and view the experiment runs for this model.
3. Click Model Version to view the basic model metadata, model signature, and lineage. You can also modify the model alias Tag and model description.
Lineage Tracking
Click Model Lineage to view the upstream and downstream lineage links of the model. Upstream nodes include raw data tables/feature tables/training data tables/experiment runs. Downstream nodes include model services/inference tables.
Experiment Tracking
1. Click Associated Task for the model version to track and view the experiment runs for this model.
2. During an experiment run, you can track the model's training metrics and parameters.
3. Track model artifacts and support exporting them.
Model Deployment
1. Go to the Model Version Details page and click Deploy.
2. Go to the model service creation process.
