A Practical Machine Learning Workshop.

Building a scalable model available via secured API on public cloud.

You have data streaming in from sensors. You want to train a model on that data and apply it to the data in real time, in the real world.

Introduction

The term "Machine Learning" or "AI" is fairly well understood by a lot of people but when it comes to you actually building a machine learning model and applying it to real data, I, for one, had a lot of unanswered questions, for example:

I know you create these models by training them but what do I train? What is the thing that sits there waiting to be given knowledge?
What do I end up with after a model has been created and trained? Is it just some code? What does it need to run on?
Assuming it is some code, what does that look like? How do I call it? What about authentication and management of that "thing"?
Then there's a whole question about how do you maintain it, change it, if new information comes along that needs to be incorporated eg in the summer it becomes clear that temperature is more or less of a factor that you thought? Or in autumn, who knew leaves falling on something would be a thing?

You can find the solution to all these questions above by experimenting and doing it yourself, with various technologies and techniques, but the value of using a cloud service like Watson Studio really becomes evident in situations like this, I'd suggest. Really shortening the time to implement a working solution and really reducing the learning time that might otherwise be involved.

Summary

You will create a simple spreadsheet containing data you might get from a sensor. This allows you to get familiar with how a model is created from real data. You will create a Machine Learning model from the data with no code (using the powerful IBM Auto AI service). We will see how IBM Auto AI pretty much performs the complex role of a data scientist for you in most situations. Critically we will then be able to see what this abstract idea of a Machine Learning model looks like, and how it can be practically used. Here, “practically used” means put somewhere where it can run, can be tested with some manually input data and can be called from code/your app/system.

Steps

The first thing we need in order to train something, is knowledge or data. In the real world this might mean a LOT of data. It's the job of a data scientist to prepare and make sense of that data so the correct algorithm (from a finite but large list) can be used to create a model. Finding that best fit of data and algorithm is the art of data science and can be a long and iterative process that may even result in a new algorithm being invented or created. It's why data science is such a skilled discipline. However, as with any area, there is a scale of complexity. This is where the IBM Auto AI feature of Watson Studio comes in. The reality is that for a lot of situations the data science process, although long and iterative, doesn't need the skill of a human, it just needs processing power to run that iterative process and compare results. IBM Auto AI does just that.

Given some data it will perform the basic role of a data scientist for us ... to prepare the data and iteratively run it through a selection of algorithms that it thinks have a good chance of success, compare the results and build a model that we can test, based on those results.

I've created a file we can use for this workshop here. Essentially the data in this file shows one thing ... once "y deflection" reaches 1.2 we get a failure. Although temperature and x deflection change throughout the file, I've crafted the data so that these should have no baring on the likelihood of failure and we should end up with a model that simply predicts failure as y deflection nears 1.2.
No we can set up our Data Science environment on IBM Cloud. This is all contained within the Watson Studio service, which you can set up for free (and it will remain available if you use it at least once every 30 days). Create this service and then click the Get Started button.
Slightly confusingly, this then opens "IBM Cloud Pak for Data" but don't worry, Watson Studio is integral to IBM Cloud Pak for Data so this Cloud Pak is what Watson Studio runs within. An IBM Cloud Pak is essentially a curated collection of related services that you can pick and choose from, and swap and change around, with a unified management layer on top and a unified monitoring and logging layer underneath. If you pay for a pak, then you can also efficiently use licences with any of the services in the pak using them on any combination of services and changing them around whenever you want.
Underneath the welcome message you'll see three options. Select the Create a project link under Work with data then click Create an empty project. You can also create a project by going to the main menu on the left and clicking Projects
Give the project a name. Every project will have files and data associated with it and they need to be stored somewhere. You may be asked to create some Cloud Object Storage at this point. If you are, then follow that process then click Create to create the project. If you already have Object Storage configured in your IBM Cloud account this will automatically be selected (as it is below).
You'll then be presented with a page where you can add and manage all the components of the project. One thing to note here is the Environments page. This is important because it allows you to track how much of your free credit is left, and if you sign up for the paid service, how much resource it is using. You have access to vast amounts of computing power from IBM here, capable of running the most demanding scientific processes. Paying attention to what you are provisioning is important. You can also edit/delete environments and add new environment configurations that will then be available to you and other users you add to your project.
Click on the Assets tab and note the file upload section on the right had side (Click the New data asset link if the upload box doesn't appear). Add the data file discussed above to your project.
Click on the Settings tab and locate the Associated services section. For specific tasks, Watson Studio needs a service to do that work. For example, no machine learning is going to to happen or be available to use in your system without a machine learning service being connected to Watson Studio. Scroll down to the Associated services section and click on the Add service drop down on the right hand side and select Watson. If you have a Watson Machine Learning service already set up on your cloud account you will be able to select it. If not, you can create one here by selecting the New service blue button and then add that service.
Next click on the Add to project blue button (top menu bar to the right) and select the Auto AI Experiment option. You'll need to give that service a name. Note that this Auto AI Experiment service required the Machine Learning service we just associated with the project. It also has some resource requirements that are quite high (20 capacity units per hour) but we will only be using it for a couple of minutes so it's nothing to worry about in this case. Click the Create button.
Now we are ready to get Watson to build our machine learning model automatically via the Auto AI service. This is where it does the job of a Data Scientist by crunching the data and trying to fit it to a number of likely algorithms and comparing the result/performance of each to get the best one. Select the Select from project button and add the data we imported earlier. Watson will look at the data and ask what particular thing you want to be able to predict using your model that will be created. Select "failed" from the list because we want to know how likely failure is based on what we input to the model. Note that you can view the data and look at distribution etc by clicking the slider icon next to the imported data file on the left. Note also that you can fine tune the Auto AI experiment by clicking on the Experiment settings button next to the Run experiment button.
Click the Run experiment button and note the progress being displayed, which can be quite interesting to watch. Try switching the view around and note the "leaderboard".
When the experiment is complete, you can view the leaderboard and save the top (or any other) result either as a machine learning model that you can then use, or as a notebook that allows you to view the code. How do you then use your model? We can deploy it to the Machine Learning service from where it can be called securely by any external code/system!
Select the result from the leaderboard that you want to convert to a model, and click the Save as button on the right hand side.
Select As a model then, when that's done, a pop up will appear that allows you to go directly to the model in your project. You can also view the model by going to the Assets tab of your project.
Your model is just an XML file asset at the moment. In order to turn this into something useful that can run and be called by your code, app, system etc you need to deploy it to the IBM Machine Learning service. This service contains run times to allow the model to run when called, and all the security and API management required of a cloud based service. Your model will effectively be available as a secured, cloud based API that can be called however you want, from wherever you want. If you want to deploy your model somewhere else on your own infrastructure you can do that too. Cloud Pak for Data also contains a catalogue feature that allows you to publish your model to your organisation, but that is not covered by this workshop. While viewing the model, click the Promote to deployment space button.
A "Space" is effectively a container to hold your models in. It has controllable access and run time resources within it. This allows you to manage all aspects of your deployment for a set of models. In the Promote to space options page select a space you already have to deploy to (or create a space using the dropdown). Select Go to the model in the space after deploying it so you can easily do the next step.
The next page allows you to create an actual deployment in the space. Click on the Create deployment button
In the page that appears, select the the deployment method. Online in this case, and provide a name for the deployment.
The next page shows the list of deployments you have in the selected space, and their status. After a few seconds your deployment will be "Deployed" which means it is available to be used. Note the three dot menu on the right hand side of the deployment that allows you to Replace asset (if you ever update the model) and Edit configuration (assign more resources to running it, if it is heavily utilised).
Click on the model to open the page that allows you to test the model with manually entered data and provides the details you need to use the model via Curl, Java, Javascript, Python and Scala. Note that you need the API endpoint credentials (key) to use any of these methods of course. See Using your model below on how to obtain this API key. Let's test the model! Click on the Test tab and enter the 4 sets of data shown below.
Then click the Predict button. You should get these results...

{
    "predictions": [
        {
            "fields": [
                "prediction",
                "probability"
            ],
            "values": [
                [
                    "no",
                    [
                        1,
                        0
                    ]
                ],
                [
                    "no",
                    [
                        1,
                        0
                    ]
                ],
                [
                    "yes",
                    [
                        0,
                        1
                    ]
                ],
                [
                    "yes",
                    [
                        0,
                        1
                    ]
                ]
            ]
        }
    ]
}

As we can see from the result, the 3rd and 4th data sets result in the model predicting failure, as we would expect!

Using Your Model

In step 20 above we could briefly see how to call the API of our model. That API was created when we deployed our model. Within Watson Studio our model sits within a deployment. At this point the model is separate from project. That project can be deleted or added to etc with no impact to the deployed model. Note on the main menu of Watson Studio (click the burger icon at the top left of the Watson Studio page) that there is a section for Projects and a separate section for Deployments that contains your deployment Spaces. To locate your "running" model (essentially get back to the page you see in step 20 above) use the main menu and go to the deployment space that contains your model. Slightly confusingly the Space contains Assets and Deployments. We have a Deployment/Space/Deployment structure which seems strange, but that is the way it has been built.

Within this Space select the Deployments tab and click on your model

This page shows the unique public API endpoint for your deployed model. This endpoint is unique to your deployment and will remain the same if you replace the model with a different one at any time. Note also the sample code provided that can be used to call/use your model from various common languages as well using the basic/universal CURL command. You will note that authentication is required to call the API of course.

To obtain the API key discussed on this page, click here or go to your IBM Cloud Account and select Manage then Access (IAM) from the Manage menu on the top menu bar then API Keys from the menu on the left of the IAM page that appears. From that page you can create a new API key.

See here for more information on the authentication process you'll need to know about before deploying your application.

Testing with the Curl command

Save the API key you obtained above by running the following command

on Mac and Linux export API_KEY=[your key]

on Windows set API_KEY=[your key]

Then run the following command from a terminal/command prompt

curl --insecure -X POST --header "Content-Type: application/x-www-form-urlencoded" --header "Accept: application/json" --data-urlencode "grant_type=urn:ibm:params:oauth:grant-type:apikey" --data-urlencode "apikey=$API_KEY" "https://iam.cloud.ibm.com/identity/token"

JSON data is returned and you will see an access_token value. Copy this value and set the environment variable IAM_TOKEN to this value using the same method as above.

on Mac and Linux export IAM_TOKEN=[your token]

on Windows set IAM_TOKEN=[your token]

Next we will call the endpoint via curl as shown in the API Reference tab of the page shown below. We will call it and provide the test data we want to use.

The curl test data can be obtained in the correct format from the Test tab of the page above. If you enter some values to test and then click the Provide input data as JSON icon you can copy and paste the test data into your curl command (after edtiing it to remove all the indentation etc).

This is an example you can use, but remember to replace the endpoint with your own.

curl -X POST --header "Content-Type: application/json" --header "Accept: application/json" --header "Authorization: Bearer $IAM_TOKEN" -d '{"input_data": [{"fields": ["Sensor","x defection","y defelection","temperature"],"values": [["Sensor1",0.3,1.3,21]]}]}' "https://eu-gb.ml.cloud.ibm.com/ml/v4/deployments/2cf41df0-bd88-4545-b318-ff118a7a8abd/predictions?version=2021-07-22"

and this JSON should be returned

{
  "predictions": [{
    "fields": ["prediction", "probability"],
    "values": [["yes", [0.0, 1.0]]]
  }]
}

Next Steps

There are many ways streaming (or flowing or "in motion") data can be analysed with a Machine Learning model. They range from something like a serverless function monitoring a ftp folder, an API (like the one you created above) being called directly by an application, using something like the "low code" NodeRED open source application available on pretty much any platform including Raspberry Pi, right up to building an enterprise AI pipeline with IBM Data Stage and Apache Kafka (or the IBM managed service version of Kafka... IBM Event Streams) and Watson Studio to manage the whole process end to end.

If you build services or solutions for you customers using IBM technology, IBM can help you with expertise and resources tailored to your situation. Contact xxxx.

IBMDeveloperUK / Practical-Machine-Learning-Part-1