Santiago
Santiago

@svpino

7 Tweets 1 reads Jan 13, 2023
90% of machine learning models never make it to production.
That's depressing.
And while most focus on making better predictions, here is something different you can do:
1 of 7
To deploy a model, you mainly need two components:
• An interface to interact with the model
• An inference pipeline
It could get more complex, but this is a good start.
2 of 7
You will use the interface to send production data to your model and receive answers. This is usually a RESTful API.
The inference pipeline transforms the data, runs it through the model, and processes the output before sending it back.
3 of 7
Here is the catch:
Building this for every model is a nightmare!
I've been there. I've done it. I've had to check myself in a clinic after that.
Okay, I'm exaggerating, but I am trying to make a point here.
4 of 7
So let me introduce you to Hopsworks, and you'll never have to deal with this problem:
hopsworks.ai
Hopsworks is a serverless platform that makes it dead simple to deploy a model.
But that's not all!
5 of 7
Hopsworks combines a feature store with a prediction service.
That means they take care of the whole ordeal for you! You can go from training to a live service in no time!
Here is the best part:
6 of 7
Hopsworks has a generous free tier!
This means you could build an end-to-end machine learning prediction service for free using:
• Python
• Hopsworks
• GitHub Actions
Here is an entire video showing how to build a whole prediction system:
youtube.com
7 of 7

Loading suggestions...