Skip to main content

Logging Experiments

Updated May 18, 2023 ·

Logging Experiments with MLflow​

MLflow is a tool for managing the machine learning lifecycle. It helps track experiments, compare models, and ensure reproducibility.

  • Log parameters, metrics, and results.
  • Makes the process more organized and manageable.
  • Use a web UI to compare different experiment outcomes.

Create an Experiment​

To start, set the experiment name using mlflow.set_experiment(). This defines a workspace to store all experiments for a specific project.

import mlflow 

# Set experiment name 
mlflow.set_experiment("Health Stats Classification")

Run an Experiment​

Once the experiment is set, we can start a new run using mlflow.start_run(). This represents a single execution of the code. We can log parameters and metrics during this run.

  • Log Parameters – Use mlflow.log_param() to log any model parameter.
  • Log Metrics – Use mlflow.log_metric() to track model performance.

Each run is linked to the active experiment. We can call mlflow methods multiple times during a run to log various metrics or record the same metric more than once.

# Start new run 
with mlflow.start_run():
  logistical_model = LogisticRegression()

  # Log paramenter and model performance
  mlflow.log_param("n_estimators", logistic_model.n_estimators)
  mlflow.log_metric("accuracy", logistic_model.accuracy)

  # Print metrics 
  print("Model accuracy: %.3f" % accuracy)

Sample output:

Model accuracy: 0.94

Retrieve Experiments​

Fetch experiment data by using mlflow.get_run(run_id) or get data for multiple runs with mlflow.search_runs(). This returns all parameters, metrics, and tags in a pandas DataFrame.

# Fetch run data 
run_data = mlflow.get_run(run_id)
print(run_data.data.params)
print(run_data.data.metrics)

# Search all runs 
exp_id = run_data.info.experiment_id
runs_df = mlflow.search_runs(exp_id)

Sample output:

{'epochs': '30', 'accuracy': 0.94}

MLflow UI​

MLflow’s web UI allows you to compare experiments, sort and filter runs, and view detailed information. This helps identify the best-performing models.

  • Compare experiments by viewing multiple runs side by side
  • View detailed metrics, parameters, and other run information
  • Manage all experiments in one centralized location
  • Streamline your workflow and identify successful models

Feedback