Deploying a Machine Learning Model on Heroku with FastAPI

Training a Random Forest classification model to predict the income category of a person based on other personal informations.

1. Created unit tests to monitor the model performance on various slices of the data.
2. Then, deployed the model using the FastAPI package and create API tests.
3. Both the slice-validation and the API tests were incorporated into a CI/CD framework using GitHub Actions.
4. UCI census datasets was used to experience updating the dataset and model in git and DVC.

Link to the project repo

The training data is the census data available at the UCI library. It is the adult.income data from the data folder.
Link: UCI Census Data

This data versioning is tracked through DVC using AWS S3 bucket as remote storage.

A basic Random Forest classifier imported from scikit-learn library and fit onto the census data Model parameters are(other than default): { "random_state": 8, "max_depth": 16, "n_estimators":128 }

Refer to the - model card

The model versioning was tracked using dvc.
Also, the performance of the model was evaluated on a slice of data (code). The results are stored in the slice_output.txt for slices done on education and race.

To run the model trainer, evaluation code (link):

python main.py

Continuous integration was incorporated in the project using the Github actions. The action was completed only if the pytest and flake8 linter tests passed on the project without any error.

Unit tests were written for the model training and the inference API features. The tests are done using the pytest library via command:

pytest test/ -vv

A FastAPI framework was developed for the inference API using the input type-hints example from pydantic library The API main file is the inference_api.py

The API was deployed on Heroku using the main branch of the current GitHub repository with Continuous Delivery enabled.