Model selection
This page describes the Machine Learning Manager, a service of the ForePaaS Legacy Platform that is not available on OVHcloud Data Platform. See the current documentation.
Model selection is the step in which you can compare all trained models in the pipeline, especially over time, and choose which one you want to deploy to production. The word model always refers to the combo of two elements: an estimator and its parameters fitted on training data.
Think of this step as a control center where you can come back at any time to monitor how well your deployed model is doing over time. If needed, you can switch it for another more accurate or efficient model.
A pipeline can store several trained models (they are all displayed on the Model Selection page) but only one can be deployed for predictions.
Here, you can:
Manage trained models
The main focus of this page is the Generated models board. It lists all models that have been fitted since the creation of the pipeline.
Remember that not all combinations from the hyper-parameter tuning step are saved but only the one with the best validation score.
Manage versions
For the moment, models' versions are identical to the version of the pipeline whose configuration was used to generate it. This ensures you have traceability over how every one of your predictions was made.
All model versions that were saved over time are listed in the table. You can choose to download a specific model version by clicking the 📥 icon.
Deploy model
The model at the top with the green atom ⚛️ icon is the model that is currently deployed in your pipeline. This means that it is the model used to make inferences, using the specifications (i.e. consumers) listed in the Deployment options step.
You can choose to manually deploy another model of your choice by clicking on the Play ▶️ button at the end of its line.
Scores
The Generated models list also display both the validation score (evaluated on the Validation set) and the testing score (evaluated on the Testing set). By default, the testing score isn't calculated in order to minimize the resources used. You can choose to compute it manually by clicking on the Refresh 🔄 button in the Testing Score column.
There is a cup 🏆 symbol next to the model with the best validation score to help you spot it more easily!
The default scoring function used in both columns is the Highlight score: it is the scoring function you highlighted ⭐ in the Training step. To display the results with another scoring function, click on the dropdown menu at the top-right of the table and select a function among those you added in the Training step.
Finally, you can also compare the scores of several functions side by side for a specific model on its details page. Simply click on the Stats 📊 icon at the end of the model's row to open it.
Set up automatic deployment of models
A key feature of the Model Selection step is the ability to configure the auto-deployment of models based on certain criteria.
Learn how to auto-deploy models
Go further
If you need training or technical assistance to implement our solutions, contact your sales representative or click on this link to get a quote and ask our Professional Services experts for a custom analysis of your project.
Ask questions, give your feedback and interact directly with the team building the Data Platform on the dedicated Discord channel.
If you need support with your OVHcloud services, create a request in our Help Centre.
Join our community of users.