Overview

From the Deployments list, click on the Name of a Deployment to land on its Deployment overview. This is where a Computer Vision team checks, day after day, that a ModelVersion in production keeps performing as expected, and where it configures how the Deployment is operated.

The header of the Deployment overview displays the Name of the Deployment and its identifier, which is useful when you work with the SDK. Next to the Name, colored badges tell what is set up on the Deployment:

• A green Campaign badge, when a Prediction Review Campaign exists. Click on it to open the Task list of this campaign.
• One badge per step of the Pipeline that is configured: Training Data, Feedback Loop, Continuous Training, and Continuous Deployment, as detailed in Pipeline.

The overview is composed of four tabs:

• Dashboard, where you access the metrics related to the performances of the ModelVersion in the frame of the current Deployment.
• Predictions, which lets you visualize the predicted images and the associated Prediction. The number next to its name is the number of PredictedAsset logged.
• Charts, where you can build your own charts about the Deployment.
• Settings.

1. Dashboard

The Dashboard tab gives you a clear and comprehensive idea of the health status of your Deployment. It is composed of contextual information and metrics.

The contextual information is displayed on top of the Dashboard:

• Champion Model: the Model and ModelVersion deployed on the current Deployment, as a link to the ModelVersion, and how long ago it was deployed.
• Shadow Model: the Model and ModelVersion potentially deployed as a Shadow Model, also as a link. When there is none, a Set shadow model button is displayed instead. When there is one, a Promote button turns it into the Champion Model, as detailed in Shadow Deployment.
• Overview: the number of inferences already performed on the Deployment (Total Inferences) and their Mean Latency.
• Alerts: the list of the Alerts triggered on the Deployment, or No alert triggered. Alerts are set up as detailed in Alerts.

Below, the Dashboard displays the metrics that any user can use to assess the behavior of the ModelVersion: the Latency chart (which you can display over the last hour, the last 24 hours, the last 7 days, the last month, or the last year), the Heatmaps, the Last Predictions, and then the charts of the metrics, such as the Outlier Score, the KS Drift, the Label Distribution, or the Average Precision and Average Recall.

These metrics are divided into two categories:

• Unsupervised metrics, computed from the inferences only.
• Supervised metrics, which also rely on reviewed Prediction.

Details about each metric are available in Dashboard metrics.

2. Predictions

The Predictions tab displays an overview of all the PredictedAsset and Prediction logged on the current Deployment.

The Predictions overview follows the same philosophy as the other image overviews of Picsellia, such as the Assets overview. It lets you visualize all the PredictedAsset with their Metadata, and access the Prediction Review tool. It is detailed in Predictions overview.

3. Charts

The Charts tab lets you create your own charts about the Deployment, with the + Create chart button, as you can already do on a Datalake or a DatasetVersion. It displays No charts found as long as you have not created any.

4. Settings

The Settings tab gathers everything that configures the Deployment. A menu on the left lists its sections:

• General.
• Access, with Members.
• Configuration, with Labels.
• Embeddings, with Data.
• Pipelines, with Training Data, Feedback Loop, Continuous Training, and Continuous Deployment.
• Notification, with Alerts.
• Campaign, with Workflow.

A. General

From the General section, you can rename your Deployment or modify its Confidence Threshold. Picsellia only stores and monitors the predictions above this threshold, so any Shape with a confidence score below it is ignored. A low threshold lets you see how the ModelVersion really behaves, whereas a higher one keeps the Deployment focused on the predictions you would act on. After any modification, click on Update to save it.

The Danger zone lets you delete the current Deployment, with the Delete this deployment button. For security reasons, you will be asked to type the name of the Deployment.

However, each PredictedAsset logged on a Deployment is physically stored through the proper Storage Connector, on the storage associated with the target Datalake selected when deploying the ModelVersion. Each time a PredictedAsset is logged, a new Data is created in this Datalake, and, as Asset does, PredictedAsset inherits from Data.

This means that when you delete a Data, any related PredictedAsset is removed from its Deployment. But when you delete a whole Deployment, all the logged PredictedAsset are deleted while their related Data remain in the target Datalake.

B. Members

The Members section manages who can access the Deployment, the same way as on a Model:

• Admins of the Organization can always access every Deployment.
• Organization access shows how many Members and Teams can already access it through their own Organization Access.
• Direct access lets you grant a Member or a Team a role on this specific Deployment, on top of what their Organization Access already provides. Use the + Add Member or + Add Team button, and the Role and Type filters to find an access.

The avatars displayed in the Members column of the Deployments list are the ones of the members who can access the Deployment. The rights are detailed in Users management system.

C. Labels

The Labels section lists the Label of the Deployment, each with its index (#0, #1...). From here, you can search a Label, and add an Attribute or a Skeleton to it, as on a DatasetVersion.

D. Embeddings

The Data section of Embeddings manages the embeddings of the PredictedAsset, which let you search for similar images among the predictions of the Deployment. A status tells whether they are Active, and the section displays the number of indexed Data with a chart of their status (Success, Error, Unsynced, Unprocessable, or Queued), which helps you check that every image is searchable.

The Delete button, in the Delete image embeddings area, deactivates the visual search and deletes all the embeddings.

E. Pipelines

The four Pipelines sections set up the Continuous Learning mechanism of the Deployment: Training Data, Feedback Loop, Continuous Training, and Continuous Deployment. Each of them is detailed in Pipeline.

F. Alerts

The Alerts section of Notification lets you create and edit the Alerts of the Deployment, as detailed in Alerts.

G. Workflow

The Workflow link of Campaign opens the Prediction Review Workflow of the Deployment, where you create and manage its Prediction Review Campaign, as detailed in Prediction Review Campaign.