MLIP Lab7 - Manage and Track Machine Learning Project with MLFlow

In Lab7, you will explore how to manage and track Machine Learning Projects using MLFlow. Although MLFlow is helpful to team project M3, We recommend doing this lab on your personal machine to avoid multiple installations and environmental conflict.

MLFLow is a powerful tool that enables practices aligned with MLOps principles. MLFlow can track experiment results during development. It can perform version control on dataset and model and store versioned model objects to make your project reproducible and manageable.

To receive credit for this lab, please show your all deliverables to the TA during recitation.

Deliverables

• Show the webpage MLFlow Tracking Server. Explain to TA your process of setting up the environment for this lab.
• Complete the train_model.py and load_model.py. Show TA your uploaded runs in Tracking Server webpage and the console output of running load_model.py
• Deploy the MLFlow model as a docker container. Show TA the console output of running ./test_inference.sh.

Environment Setup

In this step, we create a virtual environment, install MLFlow, and set up a tracking server. To begin this section, please clone Lab7 repo to your machine. Run cd MLIP_Lab7 to enter into the work directory. Then we begin the virtual environment setup"

Setup Virtual Environment

1. Run:

   python -m venv lab7

   to use python venv to create a virtual envrionment.
2. Run:

   source lab7/bin/activate
   pip install mlflow databricks-sdk
   

   to activate virtual environment and install mlflow package

Setup MLFLow Tracking Server

Option 1(Please use this option in lab7): Run MLFLow Tracking Server on Localhost

1. Run mlflow server --host 127.0.0.1 --port 6001 to launch tracking server on port 6001. Show the logs in the terminal to TA for deliverable 1.
2. Visit http://127.0.0.1:6001 to verify your MLFlow Tracking Server is running. Show the webpage in browser to TA for deliverable 1.

Option 2: Use Databricks free MLFlow Server

This option does not provide model registry. This is provided because a cloud server is better at team collaboration than local server.

1. Go to the login page of Databricks CE
2. Click on ==Sign Up== at the right bottom of the login box
3. Fill out all the necessary information. Remeber to choose community edition instead of any cloud services. When you set tracking server, instead of running mlflow.set_tracking_uri("<your tracking server uri>") in the python script, you should run mlflow.login and provide:

• Databricks Host: https://community.cloud.databricks.com/
• Username: Your Databricks CE email address.
• Password: Your Databricks CE password. For more details, please visit Additional Reference

Complete the Machine Learning Pipeline

In this step, we try to build a simple machine learning project to simulate real-world scenario. In WORK_DIR/utility/pipeline.py, we defined 3 utility functions: data_preprocessing function that generates train and test dataset, train_logistic_regression that generates trained sklearn logistic regression model, and evaluation function that returns accuracy of trained model on test dataset.

Now, we use these 3 funtions to build a training pipeline in WORK_DIR/train_model.py and a inference pipeline in WORK_DIR/load_model.py. In this process, please check documentation in WORK_DIR/utility/pipeline.py carefully. They are very helpful.

Complete the Training Pipeline

In this step, we modify WORK_DIR/train_model.py only.

1. Complete the TODO in Line 22 to extract the train and test dataset from pipeline.data_preprocessing. This function accept now arguments and returns X_train, X_test, y_train, y_test
2. Complete the TODO in Line 33 to obtain trained model from pipeline.train_logistic_regression. This function accepts X_train, y_train, params (sklearn logistic regression parameters) and output a fitted regressor.
3. Complete the TODO in Line 36 to obtain accuracy score from pipeline.evaluation. This function accepts X_test, y_test, model and output a float type accuracy score. Now, we have a complete machine learning training pipeline without MLFlow component.

Complete the Inference Pipeline

In this step, we modify WORK_DIR/load_model.py only.

1. Complete the TODO in Line 20 to predict the numpy array datapoint. You need to convert the numpy array to pandas Dataframe for inference due to the constrain of MLFlow.

Complete MLFlow Components

Complete the tracking and model registering process

In this step, we complete the MLFLow tracking and model registering components in WORK_DIR/train_model.py and WORK_DIR/load_model.py. After the completion of WORK_DIR/train_model.py, we need to run it before the next step because next step requires the tracking server to have a registered model already.

For WORK_DIR/train_model.py:

1. Provide the tracking server uri in Line 10.
2. Provide your own email as experiment name in Line 15. Now, please run python train_model.py to train the model and upload the model metrics and register the current model to MLFlow Tracking Server.

Complete the loading process

For WORK_DIR/load_model.py:

1. Provide the tracking server uri in Line 6.
2. Provide the uri of registered model in Line 10. To obtain the uri, you need to visit your tracking server webpage. Clike the Experiment section on the top. Click your experiment on the left side. Click the runs under run name column. Click on the artifacts on the top. On the screen's right side, under Make Prediction section, you can see something like logged_model = 'runs:/69c93a9c4bd14210871e7ee78483f30e/iris_model'. This string is the uri you need. Show TA this page for deliverable 2 Now, please run python load_model.py This should print out a prediction on your screen. Please show this to TA for deliverable 2.

Deploy MLFlow Registered Model as Docker Container.

Now, we use MLFlow to help us with deployment. MLFlow can pack a registered model into a docker container server. It also provides inference protocol in Local Inference Spec Page. Let us build a docker container based on run id we previously obtained in Complete the loading process.

1. First, run export MLFLOW_TRACKING_URI=<Your tracking server uri> to let the MLFlow CLI know the tracking server endpoint.
2. According to MLFlow models documentation, run mlflow models build-docker --model-uri "<Previously obtained runs:/ uri>" --name "lab7" to build the docker image.
3. Run sudo docker run -p 6002:8080 "lab7" to launch the server.
4. Run WORK_DIR/test_inference.sh to send a test inference to the server. Show TA the console output of test inference for deliverable 3.

Additional Reference

Explore MLFlow with Databricks