From Edge Impulse

About Edge Impulse

"Edge Impulse is the leading development platform for machine learning on edge devices, free for developers and trusted by enterprises." You can find the Edge Impulse training platform here.

Preliminaries

This guide covers how to train a model on the Edge Impulse platform and deploy it to an edge device using the AI Cloud.

Before you start, make sure you have:

• An edge device with the AI Manager installed.

• A camera connected as an input source in Nx Meta.

• An Edge Impulse account. Sign up at https://studio.edgeimpulse.com/.

Due to policy changes, Edge Impulse has removed the built-in YOLOv5 block. Your existing model will still work. To train a new model, re-upload the YOLOv5 block following the instructions at https://github.com/edgeimpulse/yolov5. This takes about 5–10 minutes.

Quick overview

This guide walks through four steps:

1. Train a model on the Edge Impulse platform. Full platform documentation is at https://docs.edgeimpulse.com/docs/.

2. Import your Edge Impulse model into AI Cloud. Sync the trained model to your model catalogue.

3. Deploy and test the model on your edge device. Assign the model to a device via the AI Cloud.

4. Retrain your model. Collect new training images in the field and repeat the cycle to improve accuracy over time.

Step 1: Model training using the Edge Impulse platform

Start by creating a new project on the Edge Impulse platform:

Edge Impulse project overview — start here to build and train your model

Upload and annotate your training examples, then train an object detection model. A getting started guide is available at https://docs.edgeimpulse.com/docs/tutorials/detect-objects-using-fomo.

Work through the data acquisition and impulse creation steps to reach the model selection screen:

Select FOMO MobileNetV2 or YOLOv5 — these are the models supported for import into AI Cloud

Select FOMO MobileNetV2 (0.1 or 0.35) or YOLOv5. Click "Start training" and wait for training to finish.

Only FOMO MobileNetV2 and YOLOv5 are supported for import from Edge Impulse.

Step 2: Coupling your Edge Impulse model with Nx

Open https://admin.sclbl.nxvms.com/ and log in to AI Cloud. Click the + icon in the sidebar and choose Link an Edge Impulse model. The dedicated Edge Impulse wizard opens at step 1: enter your API key and project ID.

Step 1: enter your Edge Impulse API key and project ID to link your trained model.

Your Project ID is shown in the project info panel, or as the last number in the URL when you open the project. Your API key is found in the Edge Impulse dashboard under API Keys:

Edge Impulse API key: found on the dashboard under API Keys.

Paste the API key and project ID into the form, then click Next.

API key and project ID entered: click Next to continue to owner selection.

The wizard continues through owner selection, catalogue assignment, and hardware support steps. These work the same as the generic upload wizard. On the final step, click Upload to link the model.

Step 4: choose hardware conversions, then click Upload to import the model.

When the import completes, you are taken to the model detail page with a confirmation notice.

Model link successful: the Edge Impulse model is now in your library and ready to deploy.

You are now ready to deploy your model to your selected edge device.

Step 3: Deploying and testing your model on your edge device

In Nx Meta, connect to your system and open the plugin page.

Click Manage device to open the model pipeline selector

Click "Manage device" and select the model you created.

Select your Edge Impulse model and click Add to pipeline

Click "Add to pipeline". The model is assigned and you return to the plugin settings.

Activate the Objects tab in the camera view. Detection boxes appear on the video feed:

Detection boxes appear on the camera feed once the Objects tab is active

Your Edge Impulse model is now running on your edge device.

Step 4: Retraining your model

After deploying your model, you can improve it over time by collecting new training images in the field.

Set up a postprocessor from the integration SDK for image uploads. You can set the postprocessor up to send images every N seconds or when the result is below a certain P value.

Let the system run with the postprocessor for a while to collect images.

Navigate back to your Edge Impulse project, label the uploaded images, retrain the model, then re-deploy it. Repeat this cycle to improve accuracy over time.