Skip to content

mlflow-experience

A simple, configuration-driven MLOps pipeline built with Scikit-Learn, Pydantic, and MLflow.

Welcome to mlflow-experience! This project lets you build, train, evaluate, and track machine learning experiments without writing boilerplate code. Instead of editing Python scripts every time you want to try a new feature, a different scaler, or a new model, you define your experiment in a YAML configuration file.

Why Use This Project?

  1. No Manual Glue Code: Define your data source, preprocessing steps, model type, and hyperparameters in YAML.
  2. No Data Leakage: Transformations (like scalers, encoders, and imputers) learn their parameters only on the training set and apply them safely to the test set.
  3. DataFrames Stay DataFrames: We use Scikit-Learn's transform_output="pandas" so column names are never lost during preprocessing.
  4. Automatic MLflow Tracking: Your hyperparameters, evaluation metrics, model files, and console logs are automatically saved to MLflow.
  5. Type Safety with Pydantic: Your YAML configuration is checked before running. If a column name is missing or a strategy typo exists, it fails immediately with a clear error.

Quick Example

Here is how simple an experiment looks in YAML:

# configs/mvp.yaml (snippet)
dataset:
  train:
    type: csv
    path: data/raw/churn_train.csv
    target_column: Churn
    id_column: CustomerID

preprocessing:
  - name: handle_missing
    type: missing
    params:
      columns:
        - name: Age
          strategy: mean

  - name: scale_features
    type: scaling
    params:
      columns:
        - name: Age
          strategy: standard

model:
  type: logistic_regression
  task: 'binary:classification'
  hyperparameters:
    max_iter: 250

Run this in your terminal:

uv run python -m main --config configs/mvp.yaml

The pipeline loads the data, cleans missing values, scales features, fits a logistic regression model, evaluates accuracy and F1 score, and logs the run to MLflow.

How It Works

graph LR
    Config[YAML Config] --> Validator[Pydantic Validation]
    Validator --> Loader[Data Loader & Splitter]
    Loader --> Preprocessor[Preprocessing Pipeline]
    Preprocessor --> Model[Model Estimator]
    Model --> Evaluator[Evaluation Metrics]
    Evaluator --> MLflow[MLflow Tracking & Artifacts]

Documentation Guide