Using MLflow with Databricks ③ --Model lifecycle management -

Introduction

I am writing about how to use MLflow, which is an open source for managing the life cycle of machine learning models, on the Databricks UI.

Using MLflow with Databricks ① --Experiment tracking on notebook- Using MLflow with Databricks ② --Visualization of experimental parameters and metrics-

This is a continuation of this article. Databricks managed MLflow also allows staging of trained models on the UI. This time I will write about this.

Model registration

This is a comparison screen for each experiment in the previous article. Suppose you are about to move the leftmost model from here to production. Go to the individual page from run id. image.png

Go to the individual page and scroll down. If you select the artifact "model", "Register Model" will be displayed. Click it. image.png

Enter the model and model name and register. When creating a model for the first time, only "+ Create New Model" can be selected, but if there is a registered model, that model can be selected. In that case, it will be registered as a different version of the existing model. image.png

Once registered, you will be able to go to the individual pages of your model. You can move it from the part where the model name and version are written on the right side of the screen, or from "Models" on the left sidebar. This time we will move from the sidebar. image.png

The registered model and version are displayed. Click the version that manages the life cycle. image.png

Lifecycle management

You can move to the individual page of the corresponding run id from "Source Run" of the individual page. I would like to perform staging immediately. image.png

When you click the Stage part, you can select the changed stage state from the pull-down menu. There are three types of stages: "Staging", "Production", and "Archived". You can choose to request each request and then change it, or change it directly. image.png

This time, select "Request transition to → Staging". Add a comment and ok. image.png

The request is displayed in the "Pending Requests" section. Members who have access to this model will be able to Approve or Reject. (The person who made the request can also cancel.) Try Approve. image.png

Add a comment and confirm. image.png

It has disappeared from Pending Requests and Stage is moving to "Staging". image.png

Scroll down to see the history. image.png

in conclusion

This time I wrote about model lifecycle management. Next time, I would like to write about calling the model when the model to be put into production is created.

Recommended Posts

Using MLflow with Databricks ③ --Model lifecycle management -
Use MLflow with Databricks ④ --Call model -
Using MLflow with Databricks ① --Experimental tracking on notebook -
Using MLflow with Databricks ② --Visualization of experimental parameters and metrics -
Using Python and MeCab with Azure Databricks
Machine Learning with Caffe -1-Category images using reference model
[Summary] Object detection model "End-to-End Object Detection with Transformers" using Transformer
Routine management with calendar
Regression with linear model
Create a python machine learning model relearning mechanism with mlflow