Model Versions

After you create a model, you can create new model versions or import existing ones using the AnalyticOps UI. A model can contain multiple versions. The Model Versions list allows you to access all of the versions that belong to a model in one place. You can import models, create new models, initiate the model version lifecycle and compare models, view and download artefacts and manage partitioned models within this workspace.

This chapter covers the following details related to model versions:

Create and Import Model Versions

AnalyticOps UI provides support to create (train) versions for a Git model, and import an existing pre-trained version for a BYOM model.

Train a Git Model Version

To create (train) a Git model version:

  1. From the List of Models, select a Git model and click the View Model Versions icon.

    The Model Versions page displays for the selected model showing no trained models are available.

  2. Click the Train Model button.
    The Train Model Version dialog displays.

    https://docs.tdaoa.com/images/ug_train_version_dialog.png
  3. In the Basic tab, set the properties:

    Property

    Description

    Model

    Specifies the model name in read-only format.

    Dataset Template

    Specifies the required dataset template.

    Dataset

    Specifies the dataset to be used to train the model.

    Connection

    Specifies the connection settings to be used for the training job.

    Hyper Parameters

    Lets you set the training variables manually with a pre-determined value before starting the training job.

    https://docs.tdaoa.com/images/train_basic.png
  4. In the Advanced tab, set the properties:

    Property

    Description

    Engine

    Docker Image

    Resource Template

    https://docs.tdaoa.com/images/train_advanced.png
  5. Click Train.
    The Training in Progress page displays the model version training progress.

  6. Click Done when the training progress completes.
    The trained model version adds to the list of model versions.

    The List of model versions displays the following properties for each version.

    Property

    Description

    Trained Model ID

    Specifies the auto-generated trained model ID.

    Status

    Specifies the version status as Trained, Evaluated, Approved, Rejected, Deployed, Retired.

    Dataset

    Specifies the dataset used to train the version.

    Created By

    Shows the username who has created the version.

    Tags

    Comment: cannot see any Tags in training properties

    Champion

    Shows whether the version is champion or not.

    View Events

    Allows you to view the model version lifecycle events. For details, see Model Lifecycle.

Import a BYOM Model Version

To import a model version for a BYOM model:

  1. From the List of Models, select a BYOM model and click the View Model Versions icon.

    The BYOM model versions page displays for the selected model showing no imported models are available.

  2. Click the Import Model button.
    The Import Model Version dialog displays.

    https://docs.tdaoa.com/images/byom_import_version_dialog.png
  3. Set the properties:

    Property

    Description

    External ID

    Lets you assign an ID to the model version.

    Link to BYOM Model Version

    Optionally allows you to specify the external link to BYOM version.

    Format

    Specifies the model format in read-only format.

    Tags

    Lets you add tags with the model versions to make the search easier.

    Enable Model Evaluation

    AnalyticOps provides you the support to enable or skip the evaluation of the imported model version. For details, see Enable Evaluation for BYOM.

    Upload Files

    Allows you to upload the required files for the selected format type.

    https://docs.tdaoa.com/images/byom_import_version_properties.png

    Note: Deselecting the Enable Model Evaluation option does not allow the model version evaluation. For details, see Model Lifecycle.

  4. Click Import.
    The import process progress displays in the dialog and the imported model displays in the model versions list.

Enable Evaluations for BYOM Models

The evaluation step of the model version lifecycle is optional for BYOM models.

If the user does not select the Enable Model Evaluation option in the Import Model Version dialog, the model version approval steps will be performed without evaluation and the model version lifecycle screen does not display the Evaluate action for the user.

https://docs.tdaoa.com/images/ug_enable_evaluation.png

If the user selects the option, the version lifecycle will proceed like normal Git models.

For details, see the Model Lifecycle.

Model Version Lifecycle

AnalyticOps UI provides a single screen view for the complete model version lifecycle where you can train, evaluate, approve, deploy and retire a model version and see the details of each step.

It is important to note that the lifecycle does not necessarily mean to always move forward to the next step, you can also move back to the previous step if required. For details of each step, see Model Lifecycle.

To view the model version lifecycle page:

  1. From the Model versions page, click the View Events icon for a trained model version.

    The Model Version Lifecycle page displays.

For details of model version lifecycle, see Model Lifecycle.

Model Auditability

AnalyticOps provides the complete model auditability over the lifetime of a model training, evaluation and deployment.

The Model versions are managed inside the Models module and the system keeps track and shows all events applied to a model including when a model was trained, evaluated and deployed, what training and evaluation datasets are used, in which environment the deployment was done and who made all these events.

The Models list, Model Versions list and Model Version Lifecycle pages show all the information related to the model auditability.

View and Download Model Artefacts

Model Artefacts are the output created by the training or evaluation processes. The output could be a fully trained model, a model checkpoint (for resuming training later), or simply a file created during the training process such as an image generated while training a model. AnalyticOps allows you to view and download training and evaluation artefacts of a model version.

To view artefacts of a model version:

  1. From the Model Versions list, select a model version.
    The Actions button enables.

  2. Click the Actions button.
    The Actions menu displays.

    https://docs.tdaoa.com/images/ug_artefact_action.png
  3. Select View Artefacts.
    The Model Artefacts page displays listing all the artefacts for the selected model.

    The Model Artefacts list displays the following details:

    Property

    Description

    Title

    Specifies the title of the model artefact.

    Job ID

    Specifies the related training or evaluation Job ID.

    Scope

    Specifies the scope of the job as Training or Evaluation.

    Download

    Lets you download the artefact.

    Note: The Model artefacts can also be accessed from within Training and Evaluation job details on the model version lifecycle page. For details, see Model Lifecycle.

View Micro Models

AnalyticOps provides you the support of micro modelling where an individual model is trained on every partition of the data for each model version. For example, if you have a dataset of product sales and you want to create a predictive model for future sales for every product, you can use micro modelling.

A model can have thousands or even millions of model partitions, though it is represented as one model version in the UI. A micro-model is recognized by the fact that it generates a partitions artefact after the training.

To view model partitioned for a micro-model:

  1. From the Model versions list, select a micro-model version and click on the View Events icon.

    The Model Version Lifecycle page displays.

  2. Expand the Training Details section.
    The training details section expands showing all details of the model version training.

    Note: For details of the model version lifecycle, see Model Lifecycle.

  3. Select the View Training Details for Model Partitions option at the bottom.
    The Model Partitions page displays listing all the partitions of the selected model.

    The Partitions list displays the following details:

    Property

    Description

    Partition ID

    Specifies the partition ID.

    Partition Records

    Specifies the number of records in the partition.

    Total Rows

    Displays the total number of records in all partitions.

    Total Partitions

    Displays the total number of partitions of the model version.

  4. Select the Select Hyper Parameters button to select hyper parameters to display in the Partitions list.
    The Select Hyper Parameters dialog displays.

    https://docs.tdaoa.com/images/ug_model_artefacts_4.png
  5. Hover on a field in the Available Fields list and click on the Add icon.

    https://docs.tdaoa.com/images/ug_model_artefacts_5.png

    The field adds to the Selected Fields list.

    https://docs.tdaoa.com/images/ug_model_artefacts_6.png
  6. Add all the required fields to the Selected Fields list and click Select.
    The values of the selected parameters display in the Partitions list for all the partitions.

Search Partitions List

You can search the partitions list by Partition ID and the list displays the related partitions only.

  1. Enter a Partition ID in the Search bar.
    The partitions list displays the result by listing the related partitions only.

  2. To remove the search, click the Close icon on the right of the Search bar.
    The search results remove and the list displays all the partitions.

Filter Partitions List

Filtering allows you to filter the Partitions list based on defined criteria.

Comment: Filtering not working at the moment

Sort Partitions List

Sorting allows you to sort the Partitions list by any of the available columns.

  1. Hover over a column name and click the arrow icon.
    The Partitions list sorts the content accordingly in ascending or descending order.

Archive a Model Version

The archiving feature allows you to hide a model version from the list to better organize your models. You can view the archived model versions and unarchive them if required.

Note: The Archiving feature is provided with multiple modules including projects, models, model versions, dataset templates, datasets, and connections.

To archive a model version:

  1. Select a model version in the list.
    The Actions button enables.

  2. Click the Actions button.
    The Actions menu displays.

    https://docs.tdaoa.com/images/ug_archive_version_action.png
  3. Click Archive Model.
    The model version archives and hides from the current List of Models. A confirmation message displays on the top.

  4. To view an Archived model version, click the Show Archived option on the top.
    The archived model version displays in the list along with an Archived label.

To Un-archive a model version:

  1. Select an archived model version in the list.
    The Actions button enables.

  2. Click the Actions button.
    The Actions menu displays.

    https://docs.tdaoa.com/images/ug_unarchive_version_action.png
  3. Click Un-Archive Model.
    The model version un-archives and the label Archived removes. A confirmation message displays on the top.