Get started with MLflow experiments

This article gives an overview of how to use MLflow in Databricks to automatically log training runs and track parameters, metrics, and models. For more details about using MLflow to track model development, see Track ML and deep learning training runs.

MLflow is an open source platform for managing the end-to-end machine learning lifecycle. MLflow provides simple APIs for logging metrics (for example, model loss), parameters (for example, learning rate), and fitted models, making it easy to analyze training results or deploy models later on.

Install MLflow

If you’re using Databricks Runtime for Machine Learning, MLflow is already installed. Otherwise, install the MLflow package from PyPI.

Automatically log training runs to MLflow

With Databricks Runtime 10.4 LTS ML and above, Databricks Autologging is enabled by default and automatically captures model parameters, metrics, files, and lineage information when you train models from a variety of popular machine learning libraries.

With Databricks Runtime 9.1 LTS ML, MLflow provides mlflow.<framework>.autolog() APIs to automatically log training code written in many ML frameworks. You can call this API before running training code to log model-specific metrics, parameters, and model artifacts.

Note

Keras models are also supported in mlflow.tensorflow.autolog().

# Also autoinstruments tf.keras
import mlflow.tensorflow
mlflow.tensorflow.autolog()
import mlflow.xgboost
mlflow.xgboost.autolog()
import mlflow.lightgbm
mlflow.lightgbm.autolog()
import mlflow.sklearn
mlflow.sklearn.autolog()

If performing tuning with pyspark.ml, metrics and models are automatically logged to MLflow. See Apache Spark MLlib and automated MLflow tracking.

View results

After executing your machine learning code, you can view results using the Experiment Runs sidebar. See View notebook experiment for instructions on how to view the experiment, run, and notebook revision used in the quickstart.

Track additional metrics, parameters, and models

You can log additional information by directly invoking the MLflow Tracking logging APIs.

Numerical metrics

  import mlflow
  mlflow.log_metric("accuracy", 0.9)

Training parameters

  import mlflow
  mlflow.log_param("learning_rate", 0.001)

Models

import mlflow.sklearn
mlflow.sklearn.log_model(model, "myModel")
  import mlflow.spark
  mlflow.spark.log_model(model, "myModel")
import mlflow.xgboost
mlflow.xgboost.log_model(model, "myModel")
import mlflow.tensorflow
mlflow.tensorflow.log_model(model, "myModel")
import mlflow.keras
mlflow.keras.log_model(model, "myModel")
import mlflow.pytorch
mlflow.pytorch.log_model(model, "myModel")
import mlflow.spacy
mlflow.spacy.log_model(model, "myModel")

Other artifacts (files)

   import mlflow
   mlflow.log_artifact("/tmp/my-file", "myArtifactPath")

Example notebooks

Note

With Databricks Runtime 10.4 LTS ML and above, Databricks Autologging is enabled by default, and the code in these example notebooks is not required. The example notebooks in this section are designed for use with Databricks Runtime 9.1 LTS ML.

The recommended way to get started using MLflow tracking with Python is to use the MLflow autolog() API. With MLflow’s autologging capabilities, a single line of code automatically logs the resulting model, the parameters used to create the model, and a model score. The following notebook shows you how to set up a run using autologging.

MLflow autologging quickstart Python notebook

Open notebook in new tab

If you need more control over the metrics logged for each training run, or want to log additional artifacts such as tables or plots, you can use the MLflow logging API functions demonstrated in the following notebook.

MLflow logging API quickstart Python notebook

Open notebook in new tab