Tracking

Contents

Tracking#

This page discusses the MLFlow component responsible for tracking - monitoring of the various attempts to train a model.

import mlflow
import numpy as np

import mlflow.models
import mlflow.sklearn
from mlflow.tracking import MlflowClient

from sklearn.linear_model import LinearRegression


client = MlflowClient()
mlflow.set_tracking_uri("file:///tmp/tracking_examples")

Tags#

Tags are key-value pairs used to systematize runs in MLflow. You can consider them as the properties of the run.

There are system tags that are defined by mlflow. These tags are listed in System Tags Reference section of the official documentation.

Technically, you can set or change tags using:

  • mlflow.set_tag(<name>, <value>): for setting a single tag.

  • mlflow.set_tags(<dict>): for setting a set of tags with a single call.

The following cell shows a run and how tags can be applied to it.

mlflow.set_experiment("tags_experiment")
with mlflow.start_run():
    mlflow.set_tag("single tag", "value for single tag")

    mlflow.set_tags({
        "tag1": "value1",
        "tag2": "value2"
    })

2025/08/10 16:12:37 INFO mlflow.tracking.fluent: Experiment with name 'tags_experiment' does not exist. Creating a new experiment.

🏃 View run likeable-sloth-308 at: http://localhost:5000/#/experiments/191345348810661561/runs/5c666b98b09349018ffa2d0c3061b080
🧪 View experiment at: http://localhost:5000/#/experiments/191345348810661561

The following cell searches for the previously created run. The specified tag1, tag2 and single_tag are in the same line with, for example, tags.mlflow.runName, because at the storage level this information has the same status.

id = mlflow.get_experiment_by_name("tags_experiment").experiment_id
mlflow.search_runs(experiment_ids=id)
run_id experiment_id status artifact_uri start_time end_time tags.mlflow.runName tags.mlflow.source.type tags.tag1 tags.mlflow.source.name tags.tag2 tags.single tag tags.mlflow.user
0 5c666b98b09349018ffa2d0c3061b080 191345348810661561 FINISHED mlflow-artifacts:/191345348810661561/5c666b98b... 2025-08-10 13:12:37.702000+00:00 2025-08-10 13:12:37.719000+00:00 likeable-sloth-308 LOCAL value1 /home/fedor/.virtualenvironments/python/lib/py... value2 value for single tag fedor

Experiment#

An experiment is a set of runs with different parameters for the same model. There are a few functions to manage the runs with very transparent names:

  • mlflow.create_experiment.

  • mlflow.set_experiment.

  • mlflow.get_experiment.

  • mlflow.get_experiment_by_name..

  • mlflow.delete_experiment.

For more details check Experiment page.

The following cell sets the “some experiment”.

experiment = mlflow.set_experiment("some experiment")
experiment

<Experiment: artifact_location='mlflow-artifacts:/316075063118674621', creation_time=1754732470778, experiment_id='316075063118674621', last_update_time=1754732470778, lifecycle_stage='active', name='some experiment', tags={}>

Based on the log message, you can see that “some experiment” was created during the set_experiment call.

Run#

Run in mlflow refers to the execution of a machine learning experiment or a piece of code that you want to track and record. It represents a specific execution instance with associated metadata and recorded metrics, parameters, artifacts, and tags.

The most usefull methods of mflow in roder to manipulate the runs are listted in the following table:

Method

Description

mlflow.start_run

To start run

mlflow.end_run

To end run

mlflow.delete_run

To delete run

mlflow.search_runs

To search for the runs

Note that one runs can be nested to the orhter - child nad parent runs.

For more details check out:

  • Run management section of the official mlflow tutorials - it describes functions to manipulate runs.

  • Runs page in this website.

Run can be started with mlflow.start_run() and end it with mlflow.end_run(). By passing the run_name parameter to the mlflow.start_run() function, you can define the name of the run. The following cell demonstrates how it may be:

import mlflow

mlflow.start_run(run_name="my_run_name")
mlflow.end_run()

🏃 View run my_run_name at: http://localhost:5000/#/experiments/0/runs/925288ded96849d38de4ab46b5f9e3e4
🧪 View experiment at: http://localhost:5000/#/experiments/0

Note it’s OK to have several runs with the same name - you can think of them as different versions of the same run.

The most popular approach to start a run is using mlflow.start_run as a context manager expression. The run begins upon entering the context and ends upon exiting it.

The follwoing cell shows launching an empty run using exactly context manager approach.

with mlflow.start_run(run_name="with run"):
    pass

🏃 View run with run at: http://localhost:5000/#/experiments/101046082978810346/runs/b96f254b78f14e7ca8e19b2ec23146f9
🧪 View experiment at: http://localhost:5000/#/experiments/101046082978810346

Use the mlflow.search_runs command to list all available runs.

mlflow.search_runs(search_all_experiments=True)
run_id experiment_id status artifact_uri start_time end_time tags.mlflow.source.name tags.mlflow.source.type tags.mlflow.runName tags.mlflow.user
0 7ac6c477dc724ea5ad07065cdae6794e 101046082978810346 FINISHED mlflow-artifacts:/101046082978810346/7ac6c477d... 2025-08-08 09:12:47.469000+00:00 2025-08-08 09:12:47.480000+00:00 /home/fedor/.virtualenvironments/python/lib/py... LOCAL with run fedor
1 925288ded96849d38de4ab46b5f9e3e4 0 FINISHED mlflow-artifacts:/0/925288ded96849d38de4ab46b5... 2025-08-08 08:46:34.027000+00:00 2025-08-08 08:46:34.041000+00:00 /home/fedor/.virtualenvironments/python/lib/py... LOCAL my_run_name fedor
2 c243c89b84fc4c62ad756d4512d6f4ee 0 FINISHED mlflow-artifacts:/0/c243c89b84fc4c62ad756d4512... 2025-08-08 08:41:22.880000+00:00 2025-08-08 08:44:31.008000+00:00 /home/fedor/.virtualenvironments/python/lib/py... LOCAL my_run_name fedor

Log#

The following table represents things you can log with mlflow and their corresponding functions.

Loggable Entity

MLflow Function(s)

Parameters (single)

mlflow.log_param(key, value)

Parameters (multiple)

mlflow.log_params(dict_of_params)

Metrics (single)

mlflow.log_metric(key, value, step=None)

Metrics (multiple)

Metric logging is usually done one at a time — there’s no built-in log_metrics(), but you can loop over a dict to call log_metric multiple times.

Artifacts (single file)

mlflow.log_artifact(local_path, artifact_path=None)

Artifacts (directory)

mlflow.log_artifacts(local_dir, artifact_path=None)

Models (generic)

mlflow.log_model(model, artifact_path, **kwargs)

Models (framework-specific)

- mlflow.sklearn.log_model(...)
- mlflow.pytorch.log_model(...)
- mlflow.tensorflow.log_model(...), etc.

The following code represents the procedure of model logging.

model = LinearRegression().fit(
    X=np.random.normal(0, 2, size=(100, 2)),
    y=np.random.normal(0, 20, size=100)
)
signature = mlflow.models.infer_signature(
    np.random.normal(0, 2, size=(100, 2)),
    np.random.normal(0, 20, size=100)
)

with mlflow.start_run(run_name="test_run"):
    mlflow.sklearn.log_model(model, signature=signature)

🏃 View run test_run at: http://localhost:5000/#/experiments/155825779609388391/runs/57a4997e6e7d4b37b4237f1254ffd2f7
🧪 View experiment at: http://localhost:5000/#/experiments/155825779609388391
Contents