limcheekin / mlflow-examples

Basic MLflow examples

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

MLflow Examples

MLflow examples - basic and advanced.


Python examples

  • hello_world - Hello World - no training or scoring.
  • sklearn - Scikit-learn model - train and score.
    • Canonical example that shows multiple ways to train and score.
    • Options to log ONNX model, autolog and save model signature.
    • Train locally or against a Databricks cluster.
    • Score real-time against a local web server or Docker container.
    • Score batch with mlflow.load_model or Spark UDF>
  • sparkml - Spark ML model - train and score. ONNX too.
  • Keras/Tensorflow - train and score. ONNX working too.
  • xgboost - XGBoost (sklearn wrapper) model - train and score.
  • catboost - Catboost (using sklearn) model - train and score. ONNX working too.
  • pytorch - Pytorch - train and score. ONNX too.
  • onnx - Convert sklearn model to ONNX flavor - train and score.
  • h2o - H2O model - train and score - with AutoML. ONNX too.
  • model_registry - Jupyter notebook sampling the Model Registry API.
  • e2e-ml-pipeline - End-to-end ML pipeline - training to real-time scoring.
  • reproduce - Reproduce an existing run.
  • scoring_server_benchmarks - Scoring server performance benchmarks.

The sklearn and Spark ML examples also demonstrate:

  • Different ways to run a project with the mlflow CLI
  • Real-time server scoring with docker containers
  • Running a project against a Databricks cluster

Scala examples - uses the MLflow Java client

  • hello_world - Hello World - no training or scoring.
  • sparkml - Scala train and score - Spark ML and XGBoost4j
  • mleap - Score an MLeap model with MLeap runtime (no Spark dependencies).
  • onnx - Score an ONNX model (that was created in Scikit-learn) in Java.




Use Python 3.7.5

  • For Python environment use either:
    • Miniconda with conda.yaml.
    • Virtual environment with PyPi.
  • Install Spark 2.4.2 on your machine.
  • For ONNX examples also install:
    • onnx==1.6.0
    • onnxmltools==1.6.0
    • skl2onnx==1.6.0
    • onnxruntime==1.1.0


  • Install miniconda3:
  • Create the environment: conda env create --file conda.yaml
  • Source the environment: source activate mlflow-examples

Virtual Environment

Create a virtual environment.

python -m venv mlflow-examples
source mlflow-examples/bin/activate

pip install the libraries in conda.yaml.

MLflow Server

You can either run the MLflow tracking server directly on your laptop or with Docker.


See docker/docker-server/README.

Laptop Tracking Server

You can either use the local file store or a database-backed store. See MLflow Storage documentation.

Note that new MLflow 1.4.0 Model Registry functionality seems only to work with the database-backed store.

First activate the virtual environment.

cd $HOME/mlflow-server
source $HOME/virtualenvs/mlflow-examples/bin/activate

File Store

Start the MLflow tracking server.

mlflow server --host --port 5000 --backend-store-uri $PWD/mlruns --default-artifact-root $PWD/mlruns

Database-backed store - MySQL

  • Install MySQL
  • Create an mlflow user with password.
  • Create a database mlflow

Start the MLflow Tracking Server

mlflow server --host --port 5000 \
  --backend-store-uri mysql://MLFLOW_USER:MLFLOW_PASSWORD@localhost:3306/mlflow \
  --default-artifact-root $PWD/mlruns  

Database-backed store - SQLite

mlflow server --host --port 5000 \
  --backend-store-uri sqlite:///mlflow.db \
  --default-artifact-root $PWD/mlruns  


Most of the examples use a DecisionTreeRegressor model with the wine quality data set.

As such, the python/sparkml and scala/sparkml are isomorphic as they are simply language variants of the same Spark ML algorithm.


Before running an experiment

export MLFLOW_TRACKING_URI=http://localhost:5000


Data is in the data folder.

wine-quality-white.csv contains the training data.

Real-time scoring prediction data

  • The prediction files contain the first three records of wine-quality-white.csv.
  • The format is standard MLflow JSON-serialized Pandas DataFrames split orientation format described here.
  • Data in predict-wine-quality.json is directly derived from wine-quality-white.csv.
    • The values are a mix of integers and doubles.
  • Apparently if you score predict-wine-quality.json against an MLeap SageMaker container, you will get errors as the server is unable to handle integers (bug).
  • Hence predict-wine-quality-float.json whose data is all doubles.
ezoic increase your site revenue


Basic MLflow examples


Language:HTML 59.7%Language:Python 29.7%Language:Scala 6.7%Language:Jupyter Notebook 3.8%Language:Shell 0.1%Language:Dockerfile 0.1%