Think 2020 - Lab 4862: Build an Edge-based Object Detection System Using IBM Edge Computing for Devices and IBM PowerAI Vision

Speakers

Minsik Lee, Senior Software Engineer, IBM

Naeem Altaf, IBM Distinguished Engineer & CTO Space Tech IBM Cloud, IBM

Poojitha Bikki, Software Developer, IBM

Jose Pablo Monge Bolanos, IT specialist, IBM Cloud Arch. , IBM

Lab Guide

Introduction

In this lab, we are going to do real-time video analysis using intelligent edge devices. The machine learning model is centrally trained through Cloud's PowerAI Vision. The trained model is transferred to the edge device using IBM Edge Manager. On the edge device, live streaming data is transmitted in real-time and analyzed using an inference system and the machine learning model.

High Level Architecture

To analyze real-time sensor data at the edge, edge devices equipped with computing power and analysis software are deployed close to the sensors. In the center, usually cloud or data center, machine learning models can be trained with more computing power, and inference functions can be provided if necessary. The machine learning model created on the central computer is distributed to the edge devices, allowing the edge device to perform inferences using the up-to-dated machine learning model.

Figure. Edge computing architecture for realtime data analysis

IBM provides solutions to implement edge computing systems as the followings:

• PowerAI Vision(IBM Visual Insights): PowerAI Vision makes computer vision with deep learning more accessible to business users. It includes an intuitive toolset that empowers subject matter experts to label, train, and deploy deep learning vision models, without coding or deep learning expertise. It includes the most popular deep learning frameworks and their dependencies, and it is built for easy and rapid deployment and increased team productivity. By combining PowerAI Vision software with accelerated IBM® Power Systems™, enterprises can rapidly deploy a fully optimized and supported platform with blazing performance. PowerAI Vision can be deployed on IBM Power Systems for the model training and inference. Inference component can be deployed separately and it supports both of IBM Power Systems and X86 systems.
• IBM Edge Computing: IBM Edge Computing Manager for Devices is designed specifically for edge device management to minimize deployment risks and to manage the service software lifecycle on edge devices fully autonomously.IBM Edge Computing Manager for Devices manages service software compliance automatically on edge devices without any manual intervention. On each edge device, decentralized and fully autonomous agent processes run governed by the policies that are specified during the machine registration with IBM Edge Computing Manager for Devices.

Lab environment

Figure. Lab configuration

In the lab, you will access the systems as the followings:

• Jump Server: Jump Server uses a local network and provides Windows 10 operating system. Users can access systems in the local network and Internet from the jump server via Firefox or PuTTY.
• Power AI Vision: Power AI Vision is deployed on IBM Power System with GPU on IBM Cloud. It provides a web interface for model training, inference, and management. Users can access the server from the jump server via Firefox and train machine learning models.
• Edge Manager: In the lab, the edge manager is deployed in the local network and each user has his/her own edge manager. Users are able to access the edge manager via Firefox.
• Edge Device: Linux server is used as an edge device in the lab. Instead of the real sensor, a sample streaming file will be used as an input. The image capturing service will read the streaming and detect objects by sending images to the PowerAI Vision Inference. Users will log into the edge device and register the edge device using CLI.

Getting Started

In the lab, users will review the components of the edge architecture and perform end-to-end tasks to configure the edge device with the machine learning models.

TIP: retrieve your credentials before starting the lab.

1. Log into the jump server

Users can use the jump server to access the systems. The jump server is a windows server and users will use Firefox, PuTTY, and WinSCP.

Click Enter and login as Administrator

Figure. Jump Server

Tools in the jump server that will be used in the lab:

1. PuTTY: To log into the edge device via ssh.
2. WinSCP: To send files from the jump server to the edge device.
3. Firefox: To access the PowerAI Vision, Edge computing manager, and Image Capturing service. Users can use bookmarks in Firefox to access the systems.
4. Clipboard icon: If users need to copy and paste commands from local PC to the jump server, open a target terminal, editor, or application first, click the clipboard icon, and put the command. Then, users can paste the command in the jump server.
5. Display icon: To change the screen resolution of the jump server.

2. Building a machine learning model in the PowerAI Vision

In this task, users will build a machine learning model with a pre-defined dataset.

URL, userid, and password to access the Power AI Vision server are required to finish this section. Refer to the lab guide to get these details

Open Firefox and navigate to https://[IP]/powerai-vision.

To copy and paste the URL from local PC to the jump server, copy it to the jumpserver's clipboard. To do so, click the clipboard icon, and paste the URL here. Then, you can paste the content from the clipboard anywhere in the jump server with a right click.

Login to the PowerAI vision server dashboard. Ignore the warning message and continue, then you will see the following screen.

Figure. Power AI Vision page

Log into the web page with your username and password.

Click Data Sets menu. In the Data Sets menu, users can upload images to configure a dataset. In the lab, you can import a pre-defined dataset. Click [Import .zip file] and select car.zip from Desktop folder.

After the import, you will see the data set. Click the dataset.

Click one imported image and see the boundary box for the car. Users can edit the boundary definition. Because the dataset already contains the boundary definition, you can create a machine learning model with no edit.

Go back to the data set menu and click Train model.

In this menu, you will create a new machine learning model. PowerAI Vision supports multiple model types such as Faster R-CNN, YOLO, and Detectron. In this lab, we will use Faster R-CNN.

NOTE: Do not change the model and leave it to the default name car_model

Select Object detection and turn on Advanced settings.

We will use Faster R-CNN model for training

Scroll down and change Max iteration value to 1000. Then click Train. The training will take several minutes. If you need to change the option, you can stop the training task and start it with new options again.

You will see the training progress status. This task takes about five minutes.

[Note] Each Train model task consumes one dedicated GPU and each user can use 1 GPU. Please don't trigger Train Model multiple times. In that case, other users are not able to perform this task. If there is no available GPU, please skip the training task and continue the next step. You can come back to this task later.

Congratulations! You completed the Building a machine learning model task.

(Optional task) Copying the model to the edge device

The machine learning model can be deployed with external tools through the API or can be distributed using Edge Manager's training model management function. In this lab, users will export and upload the model manually. Users can skip this task because the machine learning model is already copied to the edge device.

1. Navigate to Models menu
2. select CAR_MODEL,
3. click Export Model button. Please wait until a popup is shown. It takes about 30 seconds.
4. Select Save File and click OK button.

Save the file in Desktop directory.

Click WinSCP and log into the edge device.

Select car_model.zip and click Upload. If necessary, overwrite the file.

2. Launching PowerAI Vision Inference with the machine learning model on the edge device

The next step is to launch a PowerAI Vision Inference with the machine learning model. Users can use the model to launch PowerAI Vision Inference server in the edge device.

From the jump server, double click PuTTY and select Edge Device. Then, click Load and Open button.

Login as ibmuser.

Deploy the PowerAI Vision inference server with the model car_model.zip.

/opt/powerai-vision/dnn-deploy-service/bin/deploy_zip_model.sh -m car -p 6001 -g -1 car_model.zip

The options used to deploy a model in the above command are explained below:

`-m` model-name - The docker container name for the deployed model.

`-p` port - The port to deploy the model to.

`-g` gpu - The GPU to deploy the model to. If specified as -1, the model will be deployed to a CPU.

zipped_model_file - The full path and file name of the trained model that was exported from PowerAI Vision. It can be an image classification model or an object detection model, but must be in zip format.

This will take about one to two minutes. If you get an error, make sure that you are in your home directory. You can move to your home directory by executing cd. The following is an example of the deployment result:

The PowerAI Vision Inference is deployed as a docker container. You can check the powerAI Vision Inference service with the following docker command:

docker ps

Let's check if the inference function works by sending a sample image using the REST API. Users can send a sample car image to the inference service using curl command:

curl -F "[email protected]" http://localhost:6001/inference | jq

With GPU, it takes less than a hundred milliseconds. In the lab environment, the edge device only has CPU and it takes several seconds. After several seconds, it returns detection confidence and the object boundary information. The following is an example of the result.

Congratulations! You completed the training mode deployment task.

3. Deploying the image capturing service

In this task, users will register the edge device to the edge manager. The image capturing service is a docker container built in python. This image capturing service will be used as simulation for an edge device (camera) that monitors traffic. We already have the docker image (detector) that is used to create this service on the edge device.

This container is used to create a service that runs on the edge device. The service is then referenced to create a deployment pattern. For simplicity, a service and a deployment pattern have been created on the edge manager.

In the following steps, we'll make use of the deployment pattern and register the edge device will automatically deploy the image capturing docker container in the edge device.

Figure. Image Capturing Service

The configuration will be done using the CLI tool hzn. hzn is a CLI tool that manages edge devices and applications on the edge devices. In the edge device console, execute the following command to check the registered nodes. Make sure that it returns empty array [].

hzn exchange node list 

Check an available pattern list which can be deployed on the edge device. The result will include icp/pattern-detector which provides streaming capturing service.

hzn exchange pattern list

Register the edge device with the pattern icp/pattern-detector.

hzn register -p icp/pattern-detector

You will see the following message.

After the device registration, check the node list again.

hzn exchange node list
[
  "icp/a79f61bea8e7b7e0cd9ccdff24e985f6922bc994"
]

The edge device is now registered

After the registration, the edge agent automatically deploys the image capturing service. It will take about one minute. Repeat the following command to monitor the image capturing service deployment status.

hzn agreement list

Repeat the above command until agreement_creation_time, agreement_accepted_time, agreement_finalized_time, and agreement_execution_start_time are updated.

About one minute later, check the status of detector service.

docker ps | grep detector

Open a Firefox and click the book mark Vehicle Monitoring or navigate to http://10.0.10.3:3000

[note] In this lab, the edge device only uses CPU and the detection takes several seconds. If the GPU is used, the detection time will be less than 100 milliseconds per frame.

If you have a problem with the edge node registration, you can manually launch the image capturing service.

docker run --rm -p 3000:3000 detector:1.0.0

Congratulations! You completed the image capturing service deployment task.

4. Explore Edge Manager Web menu(Optional)

Click the bookmark Edge Manager. If the page is redirected to the login page, log into the web page and click Edge Manager again.

In the Nodes tab, you can see the status of the registred nodes.

Navigate to Services tab. You can see the detector service that is deployed on the edge device.

Congratulations! You completed the end-to-end deployment tasks.

References

IBM Edge Computing for Devices

PowerAI Vision

PowerAI Vision Inference Server

IBM Edge Computing articles