Track and monitor experiments

Domino’s Experiment Manager uses MLflow Tracking to log experiment details, such as parameters, metrics, and artifacts. Domino stores all your experiment metadata, data, and results within the project, making it easy to reproduce your work. This allows you to:

  • Monitor experiments with both automatic and manual logging.

  • Compare and review experiment runs effectively.

  • Collaborate securely with colleagues.

Experiment Manager dashboard

MLflow runs as a service in your Domino cluster and works seamlessly with your workspace and jobs. You can use your current MLflow experiments without changing any code.

Domino provides better security than standalone MLflow by using role-based access control (RBAC). Users with view access to the project can see all materials, data, results, and logs for every experiment within the project.

Set up an experiment

The metadata, data, and results tracked in an experiment help evaluate various machine-learning approaches, track your team’s progress, and make informed decisions. The logs for each experiment will inform your approach.

Before you begin, verify that you have the MLflow package in your environment. By default, all Domino Standard Environments already have this package installed.

Step 1: Start an experiment

To get started with an experiment, simply navigate to a project and Before you start logging, set up an experiment. This helps MLflow keep your logs organized. MLflow will log to the project’s default experiment if you don’t specify an experiment.

  1. Click Experiments in Domino’s left navigation pane.

  2. Use the dropdown to create a Workspace using any environment based on the Domino Standard Environment.

  3. Create an MLFlow experiment in your notebook and start tracking it with logging functions.

    Make sure that you give your experiment a unique name.

# create a new experiment
import mlflow
import os
starting_domino_user = os.environ["DOMINO_STARTING_USERNAME"]
experiment_name = f"Domino_Experiment_{starting_domino_user}"
# Replace <your_experiment_name> with the name of your experiment
mlflow.set_experiment(experiment_name="<your_experiment_name>")

Option 1: Use auto-logging

MLflow has an auto-logging feature that makes it easy to log your experiments. It automatically records important information like parameters, metrics, and artifacts. However, remember that MLflow only works with specific libraries for auto-logging. MLflow Tracking has detailed information about which libraries support this feature and the metrics logged.

This snippet shows you how to auto-log a scikit-learn experiment.

# import MLflow library
import mlflow
import os

from sklearn.model_selection import train_test_split
from sklearn.datasets import load_diabetes
from sklearn.ensemble import RandomForestRegressor

# create and set a new experiment
starting_domino_user = os.environ["DOMINO_STARTING_USERNAME"]
experiment_name = f"Domino_Experiment_{starting_domino_user}"
mlflow.set_experiment(experiment_name=experiment_name)


# enable auto-logging
mlflow.autolog()

# start the run
with mlflow.start_run():
    db = load_diabetes()
    X_train, X_test, y_train, y_test = train_test_split(db.data, db.target)
    rf = RandomForestRegressor(n_estimators = 100, max_depth = 6, max_features = 3)
    rf.fit(X_train, y_train)
    rf.score(X_test, y_test)

# end the run
mlflow.end_run()

Option 2: Set up manual logging

You can manually log custom records in MLflow. The code below shows how to log run parameters and metrics. However, MLflow Tracking offers more details and examples.

# import MLflow library
import mlflow
import os

# create a new experiment
starting_domino_user = os.environ["DOMINO_STARTING_USERNAME"]
experiment_name = f"Domino_Experiment_{starting_domino_user}"
mlflow.set_experiment(experiment_name=experiment_name)
# start a run and log parameter,metric, and artifact
with mlflow.start_run():
    mlflow.log_param("batch_size", 32)
    mlflow.log_metric("accuracy", 0.75)
    with open("example.txt", "w") as f:
        f.write("This is a sample artifact.")

    mlflow.log_artifact("example.txt")

# end the run
mlflow.end_run()

Step 2: Monitor and evaluate experiments

In order to evaluate an experiment with its associated runs, make sure that you create and run the script with MLflow code via a job.

Once you log your runs in an experiment, use Domino to view and assess the results.

  1. Click Experiments in Domino’s left navigation pane.

  2. Find and click the experiment to evaluate and show its associated runs.

  3. Click a run to analyze the results in detail and find reproducibility information.

  4. Compare runs:

    1. Select up to four runs from the table view.

    2. Click Compare.

      Compare experiments

      You can compare runs to see how parameters affect important metrics like model accuracy or training speed. Save your visualizations and artifacts in the root folder for visual comparison of charts.

      Compare runs

Step 3: Export experiment results

You can export your single experiment results to a CSV or compare experiments and download to a PDF.

To download single experiment results:

  1. Click Experiments in Domino’s left navigation pane.

  2. Find and click the experiment with the results that you want to export.

  3. Select the three dots in the upper right of the experiment.

  4. Choose Download CSV.

To download a report of compared experiment results:

  1. Click Experiments in Domino’s left navigation pane.

  2. Find and click the experiment with the results that you want to export.

  3. Select up to four runs to compare.

  4. Click Compare.

  5. Select the three dots in the upper right of the experiment.

  6. Choose Export as PDF.

Step 4: Share your projects and experiments

You can register an experiment as a model to the model registry using the MLflow API or through the Domino web UI.

You can set permissions for project assets, including MLflow logs, on a project level. Use these methods on your projects to control access:

  • Choose the visibility for your project. This will help you control who can see your project.

    • Searchable: Discoverable by other Domino users.

    • Private: Only viewable or discoverable by your project collaborators.

  • Invite collaborators and set their permissions based on project roles. This gives you detailed control over what they can access.

Troubleshooting and limitations

  • Experiment names must be unique across all projects within the same Domino instance.

  • Child runs aren’t deleted when you delete a run that has children. Delete the child runs separately.

  • You can’t stop a run from the UI. To stop a run, execute mlflow.end_run() from your workspace or job.

  • When you trigger an experiment from a workspace, the manager shows the file name. But if you are running that and editing it, it doesn’t rename the experiment automatically. After completing an experiment in a workspace, trigger a job to manage version control.

  • Best practice is to give your runs a name - otherwise it makes one for you like gifted-slug-123.

Advanced functionality

Starting with Domino 5.11.1, you can upload large artifact files directly to blob storage without going through the MLflow proxy server.

This experimental feature must be enabled inside the user notebook code by setting the environment variable MLFLOW_ENABLE_PROXY_MULTIPART_UPLOAD to true.

import os
os.environ['MLFLOW_ENABLE_PROXY_MULTIPART_UPLOAD'] = "true"

This is helpful for both log_artifact calls and registering large language models. It is currently supported only in AWS and GCP environments. There are two additional settings available for configuration:

  • MLFLOW_MULTIPART_UPLOAD_MINIMUM_FILE_SIZE - the minimum file size required to initiate multipart uploads.

  • MLFLOW_MULTIPART_UPLOAD_CHUNK_SIZE - the size of each chunk of the multipart upload. Note that a file may be divided into a maximum of 1000 chunks.

Multipart upload for proxied artifact access in the MLflow documentation has more information on using this feature. Registering Hugging Face LLMs with MLflow has directions specific to Domino.

Next steps