Model Catalog

Registering Models and Editing Metadata

Notes:

1. This feature requires PostgreSQL to be installed.

2. This demo expects that there is no existing model with the same URI or the same combination of Model Name and Version. The model catalog enforces uniqueness on these fields. Requests that result in a conflict will be rejected.

Register New Model

1. From the Deployments Overview page, select Model Catalog at the top of the page.

2. Click Register A New Model to open the registration modal.

3. Enter the following parameters:

• Model Name: cifar10

• URI: gs://seldon-models/tfserving/cifar10/resnet32

• Artifact Type: Tensorflow

• Version: v1.0

• Task Type: classification

1. Append tags and metric parameters with the following values. To add more parameters, use the + button. To delete a parameter, use the x button:

• Tags:

  • key: author, value: Seldon,

  • key: training_set, value: Resnet32,

• Metrics:

  • key: p1, value: 0.8,

  • key: p2: value: 0.6

1. Click Register Model

Edit Model Metadata

From The Model Catalog

Save metadata from the model catalog

1. From the Deployments Overview page, select Model Catalog at the top of the page.

2. Select a model. For the purposes of this demonstration, the model registered in the Register New Model step will be used.

3. On the side drawer that opens, click EDIT METADATA.

4. Add a new tag with the following values:

• key: stage, value: production

1. Click SAVE METADATA at the top right hand side of the side drawer to save your edit.

From A Deployment’s Dashboard

Note: Please follow the instructions on creating an Outlier Detector.

1. From the Deployments Overview page, select the deployment to append metadata on the running model. For the purposes of this demonstration, the deployment created from the Outlier Detector example will be used.

2. From your selected deployment’s dashboard, click the model being used to open a drawer on the right.

3. Click EDIT METADATA.

4. Under the Tags section, press the + button to add a new tag.

5. For the new tag, enter the following parameters:

• key: stage, value: production

1. Click SAVE METADATA on the top right of the side drawer.

Deploying Models From The Model Catalog

Note: This demonstration requires the cifar10 model to have been registered already.

1. From the overview page, navigate to the model catalog using the header tabs.

2. Select the more icon for the model named cifar10.

3. In the dropdown menu that appears, select Deploy.

4. In the wizard which appears, name the deployment cifar10, and set the protocol to Tensorflow. Choose your deployment type. Click next when done.

5. On the Default Predictor Parameters screen, the model URI should automatically be appended. For Triton models, the model name will also be appended automatically. Click next to continue.

6. Skip the remaining steps and click Launch.

Select Deployments header tab to verify that your model has been deployed.