Deploy a ModelVersion

At this stage, you are supposed to have at least one Model and ModelVersion in your Private Model Registry, either created from an Experiment or manually imported.

It is now high time to deploy thisModelVersion and make it perform inferences. To do so we will instantiate from the ModelVersion that needs to be deployed a Deployment.

A Deployment is a Picsellia object, linked to a ModelVersion from the Private Model Registry , that allows users to track and monitor the ModelVersion behavior. It has three main features:

• Monitor the ModelVersion behavior by displaying performance metrics.
• Visualize the predicted images and the associated predictions made by the ModelVersion.
• Define a ModelVersion retraining pipeline.

1. Deploy button

To create a Deployment from a ModelVersion, you need to access it in your Private Model Registry.

From the ModelVersion overview, you just need to click on Deploy.

You first need to decide where your ModelVersion will be hosted and served, meaning the infrastructure that operates the ModelVersion and lets it make predictions.

As for the training, you can either use the infrastructure provided by Picsellia or use your own infrastructure.

In both cases, to leverage the Deployment features of Picsellia, the infrastructure serving the ModelVersion must be connected to the related Deployment, so that you can visualize the predictions made by the ModelVersion and access the performance dashboard.

A. Picsellia Serving

Select Picsellia Serving to use the Picsellia infrastructure to host and serve your ModelVersion. In this case, Picsellia handles the whole deployment workflow: your ModelVersion is deployed on a ready-to-go serving engine with ad-hoc GPUs and an automated pipeline.

In case you want to use the Picsellia infrastructure to host and serve your ModelVersion, you need to select Deploy on Picsellia Serving. In this case, Picsellia is handling the whole deployment workflow, yourModelVersion will be deployed on the Picsellia Serving Engine hosted on AWS Cloud and powered by T4 GPUs.

This workflow is fully transparent for the user, meaning that aDeployment will be instantiated for you, and the API connection between your ModelVersion hosted on the Picsellia Serving Engine and the Deployment will be done automatically.

This way all the Predictions done by the ModelVersion, will be logged seamlessly in the Deployment and, as a consequence, the performance metrics are computed and displayed in the Dashboard. More details about Deployment are available here.

In the Deployment modal, when selecting the Deploy on Picsellia Serving option, you will be asked to input a Threshold. Only Shape predicted on an image by the ModelVersion with a Confidence score over this Threshold will be considered. A best practice is to set a low Threshold at first to see how the ModelVersion is behaving and modify it later in the Settings of the Deployment.

Please note that no Threshold is requested to deploy a ModelVersion that is doing Classification as Detection Type.

You will also be asked to select a Datalake. Indeed, any predicted image visualized in a Deployment has to be physically stored through a Storage Connector that is linked to a Datalake. So in case the current Organization has several Datalake instances, you can choose among them where the predicted images will be stored and as a consequence in which Datalake they will be visualizable.

Once the Threshold and Datalake have been defined, you just need to click on Deploy to seamlessly deploy your ModelVersion on the Picsellia Serving Engine and initialize a Deployment.

To use Picsellia Serving, make sure you still have MPU available for the current Organization. As making predictions requires GPU usage, this usage is contractually tracked and billed through MPU. You can track your MPU consumption in the Plan and usage dashboard, as detailed here.

The Picsellia Serving Engine is able to serve properly any ModelVersion that is inherited from a Picsellia Experiment using as Base Architecture a ModelVersion from the Public Model Registry.

However, if you have manually imported theModelVersion to your Private Model Registry, there is no guarantee that this ModelVersion can be properly served by the Picsellia Serving Engine. In this case, it is strongly advised to contact the Picsellia team before deploying this ModelVersion on the Picsellia Serving Engine.

B. Monitoring Only

In case you want to deploy your ModelVersion on your own serving infrastructure, you can download all the model files from the ModelVersion overview as explained here and deploy it on your infrastructure.

First, download the Model files from the ModelVersion overview, as explained here, and deploy them on your infrastructure. Then, to leverage the Deployment features of Picsellia, click on Deploy from the ModelVersion overview and select Monitoring Only in the modal.

In the Deployment modal, when selecting the Monitoring Only option, you will be asked to input a Threshold. Only Shape predicted on an image by the ModelVersion with a Confidence score over this Threshold will be considered. A best practice is to set a low Threshold at first to see how the ModelVersion is behaving and modify it later in the Settings of the Deployment. Please note that no Threshold is requested to deploy a ModelVersion that is doing Classification as Detection Type.

You will also be asked to select a Datalake. Indeed, any predicted image visualized in a Deployment has to be physically stored through a Storage Connector that is linked to a Datalake. So in case the current Organization has several Datalake instances, you can choose among them where the predicted images will be stored and as a consequence in which Datalake they will be visualizable.

After having clicked on Deploy, a Deployment related to this ModelVersion will be created.

Now that the Deployment is created on Picsellia and attached to a ModelVersion from your Private Model Registry, you need to perform the interconnection between your infrastructure that is serving your ModelVersion and the related Deployment created on Picsellia.

Basically, you need to add from your serving infrastructure some API calls (using the Picsellia Python SDK) after each Prediction performed by the ModelVersion on your infrastructure to send back to the related Picsellia Deployment, the PredictedAsset and the associated Prediction. This way, from the Deployment, you will be able to visualize the Predictions, access performance metrics, and review the Prediction for future usage.

Here is the procedure to follow the perform the interconnection between your serving infrastructure and your Picsellia Deployment.

Obviously, by using your own infrastructure to serve your ModelVersion, no MPU quota will be consumed.

Regardless of the deployment type you chose, your ModelVersion is now deployed, and the associated Deployment has been created in your Organization. The next pages detail the different features of Picsellia Deployment.