Hello, We are sending data from a smartwatch -> IoT Central -> Event Hubs -> Data Explorer -> Blob Storage. We are then using the blob storage as a datastore in Machine Learning, which we make a dataset of. We deployed a model we trained locally to Azure Machine Learning. We now want to send new data from the watch to the model to make predictions on. We are wondering how we can do this? Do we just update the dataset the same way we are currently sending the data? and if so, how can we then auto send it to the model? Or is there another way to send this new data? Can we still send through blob storage? Or should we send the data directly from the watch to the webservice made by the model? Thanks so much!

Hello @yjay , so you get new data from devices and you want to predict using that data? Check out Azure Stream Analytics which can ingest messages from the Event Hub. Then, you can Azure ML as a function on make decisions based on the incoming data using ML.

Send new data to Deployed model

Accepted answer

Sander van de Velde 29,281 Reputation points MVP

2021-04-27T22:16:06.013+00:00

Hello @yjay ,

so you get new data from devices and you want to predict using that data?

Check out Azure Stream Analytics which can ingest messages from the Event Hub.

Then, you can Azure ML as a function on make decisions based on the incoming data using ML.
Please sign in to rate this answer.

1 person found this answer helpful.
yjay 256 Reputation points

2021-04-28T14:06:58.377+00:00

Hi @Sander van de Velde ,

Thanks so much for your reply, would calling the webservice (that was made after deploying the model) directly from the watch work as well?

Sander van de Velde 29,281 Reputation points MVP

2021-04-28T14:46:28.297+00:00

Hello @yjay-43,

yes, technically this could work. Keep in mind you pay twice for the same data (a call to the service, an answer comes back, send the answer to the IoT Hub).

you need to secure access to prevent paying for other accessing your model. You need to store the security token on your device.

As an alternative to a Stream Analytics job, you can use an Azure Function in the cloud instead (route messages from IoT Hub to your azure function which calls the ML). Consumption plan pricing includes a monthly free grant of 1 million requests and 400,000 GB-s of resource consumption per month per subscription in pay-as-you-go pricing across all function apps in that subscription

You can also try to run the ML model on your device...

Please accept an answer if correct. Original posters help the community find answers faster by identifying the correct answer.

yjay 256 Reputation points

2021-04-28T17:54:51.887+00:00

Hi @Sander van de Velde

Thanks so much for these solutions, very helpful!

What do you mean by:

a call to the service, an answer comes back, send the answer to the IoT Hub

Why does it send back to IoT hub?

Sander van de Velde 29,281 Reputation points MVP

2021-04-28T19:06:48.593+00:00

(sorry for the cryptic answer. There is a 1000 character limitation).

I mean that if the device first calls the ML service to get an answer (the outcome of the prediction) and then sends that same answer to the cloud using IoT HUb DeviceToCloud communication, this is a slow process and you 'pay twice' for the same information.

It seems in your case you are only interested to make the ML call and that's it.

Regards,

Sander

Yael 26 Reputation points

2021-04-29T13:48:08.21+00:00

(No problem!)

Yes in our case we are looking to display the outcome of the answer on the device itself or to send it to our own platforms but not back to azure.

In this case would you recommend just calling the webservice directly from the watch?

Thanks so much!

Sander van de Velde 29,281 Reputation points MVP

2021-04-29T15:03:04.147+00:00

Hello @Yael ,

Nice use case, and with each solution, it depends.

Adding new functionality in an agile way is the best way to figure out if this works for you.

So try out this online solution and learn.

I advise you to research running the ML logic locally (which is hard for a limited device).

Using the ML model online (either directly or via a public (azure function?) API) is low-hanging fruit but can make you depending on a good internet connection which is perhaps annoying your customers due to the data consumption and perhaps the roundtrip which takes considerable time.

Please accept an answer if correct. Original posters help the community find answers faster by identifying the correct answer.

Yael 26 Reputation points

2021-04-29T15:53:31.55+00:00

Hi @ SandervandeVelde42,

Thank you for the suggestions, sounds like a good plan!

Thanks so much for all your help!
Sign in to comment

Send new data to Deployed model

0 additional answers