Skip to content

Latest commit

 

History

History
77 lines (51 loc) · 5.16 KB

README.md

File metadata and controls

77 lines (51 loc) · 5.16 KB

KAN API overview

KubeAI Application Nucleus for edge (KAN) API defines a common object model that describes the full stack of intelligent Edge solutions, from AI models to solutions to devices and sensors. Because these objects are defined as standard Kubernetes custom resources, you can use popular Kubernetes tools like kubectl to manipulate these objects.

KAN API object models

Mapping between KAN API object models and KAN portal concepts

The KAN portal experience aims to provide a streamlined experience of creating and managing intelligent Edge solutions leveraging cameras. Hence, we’ve hidden some KAN API concepts and renamed a few objects to make the UX more intuitive for target scenarios. The following table summarizes how portal concepts are mapped to API concepts:

API Object Portal Concept
device.fabric.kan Camera
instance.solution.kan Deployment
model.ai.kan AI Model
skill.ai.kan AI Skill
solution.solution.kan There are no solutions surfaced on portal. Essentially, a portal operates on a single, system-maintained solution object on behalf of the user.
target.fabric.kan Compute device

Typical workflows

Depending on your focus, you can start with the AI workflow, the device workflow, or the solution workflow as described below. At the end, you can create Instance objects which represent running deployments of your intelligent Edge solution.

AI workflow

  1. Create your AI model using tools of your choice.
  2. Once you have the AI model file, register your AI model with KAN as a Model object.
  3. Then define AI Skill objects that define processing pipelines. A processing pipeline reads data from a data source, applies one or more AI models (and other transformations), and sends inference results to designated outputs.

Device workflow

  1. Register your computational devices with KAN as Target objects. You can also specify desired runtime components, such as a KAN agent, in your target definition.
  2. Manually register your non-computational devices, such as sensors and actuators, as Device objects. You can also leverage projects like Akri to auto-discover devices.

Solution workflow

  1. Define your intelligent solution as a Solution object, which consists of a number of Component elements. A component is usually a container, and it may refer to AI Skill objects in its properties.
  2. Define a Instance object that maps a Solution to one or multiple Target objects. Once the instance object is created, KAN ensures the impacted targets are updated according to the desired solution state and target state.

Note

The current version of KAN portal doesn't explicitly expose the Solution object. Behind the scenes, you always work with a single Solution object that is automatically managed. However, you can use KAN API to examine and update the object.

A sample workflow

Assume that you are creating an intelligent edge solution that uses a website to show number of cars passing an intersection each hour. The following workflow describes how to create and deploy such a solution using KAN API.

  1. Create or select a car detection model. KAN comes with a model zoo, which contains a car detection model you can use.

  2. Register the model as a KAN Model object.

  3. Define a KAN Skill object that defines a pipeline that:

    • Takes input from a camera;
    • Sends frames to the car detection model;
    • Collects inference results and sends detection events to an output (such as IoT Hub or an HTTP endpoint).

  4. Define a KAN Solution object that will create a Docker container Component that takes the Skill as input and drives the inference process.

    Since KAN provides some containers out-of-the-box, you don't have to create these containers yousrelf.

  5. Create your website container and add it as a Component of your Solution.

  6. Define a Target that represents a computer to which you want to deploy your solution.

  7. Define Device objects for cameras you want to use. These Device objects are associated with your Target object through labeling.

  8. Create a Instance object that deploys your Solution to your Target.

Getting started

Visit the KAN API Quickstart to try out the tutorials.

Next steps