Prior to initiating the training process, and now that you’re familiar with the monitoring view, let’s rewind a little and remind ourselves what processes are involved in the Data Acquisition phase.

The illustration below shows how, during the Data Acquisition phase, training data is generated from devices and pushed to the system.

Each image, captured by a worker and sent over the network, is received and pushed to local S3 storage on the Edge. This phase may take a certain period of time until a large number of images is collected. To maximise accuracy you ideally want to train the model with vast amounts of training data.

To initiate the training process, click the button on the upper-left side of the window:

After you click Train Data, you’ll see in the monitoring view a series of live animations illustrating the actions actually taking place in the platform. The following enumeration describes the process:

All the steps above form part of the Data Preparation and Modelling phase (described in the Architecture chapter) and are well illustrated in the diagram below:

The end-to-end process is not done yet. It then enters into the Delivery phase. The new model has now been pushed to an S3 bucket edge1-ready that is being monitored by an integration point on the Edge (Manager)

When the Tekton pipeline uploads the new model to the S3 bucket, the Edge Manager notices the artifacts and initiates the download of the model and hot deploys it in the TensorFlow model server, as shown in the picture below:

The AI/ML engine, powered by the TensorFlow Model Server, reacts to the new version (v2), now available in its local S3 bucket, and initiates a hot-deployment. It loads the new version and discards the old one that was held in memory. This process happens without service interruption. Clients sending inference requests inadvertently start obtaining results computed with the new hot-deployed version (v2).

The platform keeps running its live services at all times. Customers (shoppers) and workers interact with the platform while, in the background, new models are continuously being trained, delivered and deployed.

The demo’s inferencing phase is illustrated in the picture below:

You should already be familiar with the flows above. You had the chance to perform some positive/negative tests via HTTP/MQTT. The Shopper application (Camel) first sends an inference request against the AI/ML engine to identify the product, then:

You’ve already interacted with the application using the demo App. Let’s use it again to try out the newly trained version.

Go back to the Smart Application in your browser and this time, with the newly trained model (v2), send the Bali Green Tea image that v1 didn’t know about.

This time, the product should be identified, and you should obtain a price tag as follows:

Bravo !! you have completed the full cycle.

First things first, head to your OpenShift console and find the following ConfigMap:

Edit the catalogue to include a new product.
Follow the steps below to find your way:

The JSON data configures the products. You’ll find in the definition all the products you have been playing with.

Now include in the configuration a new product. If one does not come to mind, use the JSON data below to configure a Computer Mouse product:

Copy the JSON node above and paste it in the ConfigMap.

Click Save.

Now you need to restart the Price Engine. You can simply kill the pod and Kubernetes will restart a new one that will read the new ConfigMap value.

You can use the web Console to do so, or execute the oc command below:

You can validate, by inspecting the logs, the new started pod has loaded the new product catalogue. Again, you can use the web console, or execute from your terminal the command below:

In the output logs from the command above, you should find the value "Computer Mouse". Your logs should look similar to:

Next, you need to configure the Smart Application to allow selecting the new product for training.

You need to perform a similar operation. From your OpenShift console, find the following ConfigMap:

Follow the same steps as previously done to find the ConfigMap and open the Edit window.

Replace the old product by the new one.
If you used the previous sample configuration for the "Computer Mouse", copy the value below and replace the old product in your text area:

Click Save.

Now you need to restart the Shopper application. You can simply kill the pod and Kubernetes will restart a new one that will read the new ConfigMap value.

You can use the Web Console to do so, or execute the oc command below:

You’re now ready to generate and ingest training data.

In this section you will use the web based Smart App to capture images and push them to the platform. The easiest way to capture images is to use your smart phone.

Share the URL’s address with your phone and open it with the device’s browser.

Follow the steps illustrate below to capture and send images:

Above, in step 2 you should find the option that you configured earlier. If you used the sample JSON snippets you should find on your screen the option:

Now, take various pictures and push the as indicated. Change angles, rotate the object, flip the object. The more data you push, the more accurate your model will become.

When you’re done with the Ingestion phase, you should end up with a collection similar to the picture below:

As you did earlier, from the administrative page, trigger the training process using the UI button.

Click Train Data.

Verify the pipeline is running and wait for its completion.
Use the command:

You should obtain something similar to:

where you can see the previous PipelineRun in status Succeeded, and the new one Running.

When the pipeline completes, go back to your App’s Detection Mode and try the new product out. If you trained your computer mouse, take a picture and send it.

You should obtain a price tag for your mouse:

Well done! You have iterated the Product Catalogue to include new articles on offer.