Last updated 10th January, 2020.
Deploying models is the main feature of the ML Serving. This guide provides step by step instructions on how to deploy a serialized model.
- Having previously created a ML Serving namespace. Steps are described in section Initializing namespace
- Having previously exported your machine learning model into a supported format. Steps to export TensorFlow models are described here, steps to export Scikit-Learn models are described here.
Step 1 - Upload the serialized model into the public cloud object storage
You need to upload your serialized model into the object storage
container which is linked with your namespace. On the
main page of your wanted namespace you can click on the
name of your linked container under the
Container Object Storage section.
This step is just a redirection on
Public Cloud >
Object Storage >
You can then click the
Add object button and select your exported
model from your local file storage.
After clicking the
Import button you should see your exported model in
the list of containers object.
Step 2 - Reach model page
Inside your ML Serving namespace page there is a dedicated tab for
managing models :
You can start the deployment of a new model by clicking the
Deploy Model button.
Step 3 - Select the custom model
In this guide, we explain how to deploy a serialized model also
called custom model. Just select custom model and click
Step 4 - Select your exported file
Here you will click on file and search on the dropdown list your model
Step 5 - Choose a name for your model
That name identifies your model among others on your namespace.
After you filled your wanted name, click the
Step 6 - Choose your auto-scaling options
A model is composed of one or several running instance(s). These instances are automaticaly scaled by the ML Serving depending on the input load of your model. That step allows you to configure the minimum and maximum number of instances that you want to run.
During the beta phase, the auto-scaling feature is not customizable and we reserve the right to remove unused models
Step 7 - Choose your instances configuration
Each model instance is related to a CPU & RAM flavor. You can choose the wanted flavor among a list of existing ones.
During the beta phase only one type of instance is available and is free of charges. Additional flavors will be created to fit specific needs.
Step 8 - Wait for model deployment
The ML Serving will sequentially perform the following tasks :
- The building of the model : you see the build status as
- The deployment task : you see the API Status as
When everything is up and running you see the build status as
Deployed and the api status as
Running. The URL where to reach
your model is also displayed so you can start requesting it.
Step 9 - Check the model web interface
When you follow the model URL in your web browser, you should arrive on the web page that describes the available API.
On your first visit, you will need to provide credentials to access the web page. Just log in with no username and your model-evaluation token as password.
This page should look like this :
There are two endpoints available on that kind of model :
- An endpoint to describe the model's inputs and outputs
- An endpoint to query the model
The web interface allows you to interactively execute queries on the different endpoints of your model if you want to test quickly.
- You can check the official documentation of the open source product ML Serving about Models
- You can check the supported compatibilities for custom models