udacity-nanodegree-mldevops-project3

Third project of Udacity ML DevOps Engineer nanodegree: Deploying a Machine Learning Model on Heroku with FastAPI

Overview

The main goal of this project is to develop skills in the following tools:

• DVC: Open source version control system for ML projects. Ability to track data and models.
• FastAPI: Fast web framework for building APIs with Python.
• Heroku: Platform as a service (PaaS) that allows free deployment of applications in the cloud.

Dataset

Here we will be working with the US census dataset. The goal is to predict whether a person earns more than $50k based on certain census information (such as age, gender, education and race).

EDA

Some basic analysis of the raw data is contained in the EDA notebook in the notebooks folder. Note that most of the output is not visible from pandas-profiling. You will need to open the notebook and rerun it to see the output.

Data cleaning and processing

The code for the taking he raw data and performing basic data cleaning was developed in the clean-data.ipynb notebook. The code was then productionized in the mlp package under the processing submodule. The code for preparing the data for training was provided by Udacity and is included in the same module.

Modeling

Finding the absolute best model was not the main goal of the project. However, I ran some analysis and a hyperparameter search for a Random Forest classifier. This can be found in the notebooks folder.

Instructions

Below you will find instructions on how you can rerun things in the repository.

Clone repo

Run the following to clone this repository to your local directory:

git clone https://github.com/robsmith155/udacity-nanodegree-mldevops-project3.git

Python environment

To create the Python virtual environment, run the following (assumind Conda is installed):

conda env create -f environment.yml

Setup pre-commit

We will use pre-commit to run Git hook scripts on every commit to automatically run code linting and checking. Here we use isort to automatically sort imports, black to style the code and flake8 to check the code.

Pre-commit will set up these Git hooks for you. Change directory to the root of the repo. Make sure that the virtual environment is active and run:

pre-commit install

Now when you make a Git commit it should run the Git hooks for isort, black and flake8. Note that this will only check the files that are being committed. If any changes are made to these, the Git commit will not be completed and you will need to add the changed files again and commit

Alternatively you can run the hooks on all files prior to staging and commiting as follows:

pre-install run --all

Pytest

To run all tests in the repo, simply run:

pytest -vv

Data

In the project conducted for Udacity the raw data and all subsequent outputs and models were stored in an AWS S3 bucket. However, you will not have access to this so you need to download the raw data from the UCI website and put it in the data folder. You can then rerun the pipeline detailed in the next section.

DVC

DVC remote

In this project I used an S3 bucket provided by Udacity to store the data and models. However, this will likely be deleted after completing the course and nobody else has access to it.

Instead, you will need to create your own remote storage for DVC. This can just be a local directory. To do this:

dvc remote add -d localremote <PATH_OF_LOCAL_FOLDER>

Rerun pipeline

Now you can rerun the whole pipeline taking us from raw data to trained model using the DVC pipeline contained in dvc.yaml. Run:

dvc repro

FastAPI

Here we use FastAPI to create a web framework for deploying the model.

Run locally

To run the app locally, you can serve the app as follows from the root of the repo:

uvicorn app:app --reload

Note that the --reload argument means that uvicorn will automatically redeploy if you make changes to the code. By default, the route of the app will be available at http://127.0.0.1:8000/. To see the FastAPI docs use

Heroku deployment

The app has been deployed to Heroku. You can find the app at: https://robsmith155-salary-prediction.herokuapp.com/

The FastAPI docs can be accessed here.

An example of sending a POST request to the deployed model on Heroku can be found in the example_heroku_request.py file in the repo. With the virtual environment active, just run:

python example_heroku_request.py