avatarVaibhav Kumar

Summary

The provided content outlines a comprehensive guide on deploying machine learning models as APIs for free using Python, Flask, and PythonAnywhere, enabling their use across various platforms.

Abstract

The article "Deploy Machine Learning Models for Free" delves into the process of making machine learning models accessible and usable by converting them into APIs. It emphasizes the importance of deploying models to impact the general public positively, rather than limiting their use to domain experts. The guide covers the prerequisites, including the ability to create and train models in Python, and introduces essential libraries such as pickle for serialization, Flask for web development, and PythonAnywhere for hosting. It walks through setting up the environment with Python 3+ and required packages, data flow from user input to model prediction, and the use of Flask to create API endpoints. The article also addresses serialization and deserialization of models, handling arguments in Flask routes, and ensuring cross-origin resource sharing (CORS) for API accessibility. Additionally, it provides a template for a predict function, suggests a file structure for the project, and guides on hosting the Flask app on PythonAnywhere. The author concludes by offering a front-end application template and discussing the deployment of more complex models on personal servers with GPU support.

Opinions

  • The author believes that deploying machine learning models should not be intimidating and is within reach for those with web development skills.
  • It is noted that while PythonAnywhere offers a free platform for educational purposes, more powerful servers with GPU support are necessary for complex deep learning models.
  • The author shares a personal view on the inefficacy of using machine learning to predict Bitcoin prices, suggesting it as a failed attempt to demonstrate the limitations of such models.
  • The author values community feedback and invites readers to suggest improvements or point out mistakes in the deployment process.
  • There is an endorsement for using JSON format when dealing with web APIs due to its structured data sharing capabilities.
  • The author expresses past struggles with CORS errors and provides a simple solution within the Flask framework, also inviting better solutions from the community.
  • The article promotes the use of free tools and services, such as PythonAnywhere for hosting and Netlify for front-end applications, highlighting the cost-effectiveness of these solutions.

Deploy Machine Learning Models for Free

Machine learning model deployment

Introduction

Machine learning models are no good lying in the IPy notebooks or scattered python scripts. To change lives and make an impact in the world, the model needs to be deployed in a way it’s usable by the general public and not just by the domain experts. For this, the model needs to be encapsulated behind some sort of API that other applications could use to communicate with the model. It involves quite a bit of web development and people often get intimidated by it. However, it is not that difficult. Read through this post to discover it for yourself!

The aim of this post is to deploy a machine learning model as a general API over the web that can be used over various platforms such as:

  1. Websites
  2. Bots for apps such as telegram, slack, discord and many more
  3. Android/iOS apps
  4. Alexa skills
  5. Any other platform that allows you to send basic GET/POST requests over HTTP/HTTPS

Host the API for FREE*!

*Unless your model uses CUDA which requires nvidia GPUs this method is free.You need to buy a server with an nvidia GPU in case you want to use CUDA or any other GPU intensive library. In case you do have a server, I’ve covered introductory guidelines at the end of this post.

Prerequisites

Readers need to be able to create machine learning models, train them and later use them to predict results in python. This post is aimed to only provide insights on deploying a machine learning model into production for small-scale use.

Following are the few libraries and resources which will be used:

  • pickle: A native python library to save (serialize) and load (de-serialize) python objects as files on the disk
  • flask: A python based easy to use web framework. Here are the docs and a good series of video tutorials by sentdex
  • pythonanywhere: A free to use educational website that allows hosting python flask and provides a complete python development environment.

Environment setup

Make sure to use python 3+

pip install the following packages:

  1. flask (our python-web framework)
  2. flask_cors: for CORS headers
  3. jsonify: to return JSON files using flask
  4. Other learning libraries as per needs like numpy, pandas, sklearn etc

pip install flask flask_cors jsonify numpy pandas

Data flow

Following is a diagrammatic representation of the data flowing between the user and server space.

API Data flow

Here’s the summary of what needs to be done:

  1. Train the model using Jupyter notebook
  2. Save the trained model object as a pickle file (serialization)
  3. Create a flask environment that will have an API endpoint which would encapsulate our trained model and enable it to receive inputs (features) through GET requests over HTTP/HTTPS and then return the output after de-serializing the earlier serialized model
  4. Upload the flask script along with the trained model on pythonanywhere
  5. Make requests to the hosted flask script through a website, bot, android app or any other application capable of sending HTTP/HTTPS requests

Machine learning model

The purpose of this post is the deployment of a model and not its creation. Let's just assume we have a model which takes in date as input and predicts the price of bitcoin. The model is already trained over historical price data.

Model data flow

If you are interested in knowing about this model you may refer to my github repository. Although, beware! Bitcoin’s price can never be predicted using machine learning and this is just a failed attempt (that too a very stupid one) to only consolidate this point. Efficacy of the model is not the concern for this blog post so let’s just deal with this model for now.Serializing and de-serializing the model

Pickle

A module called pickle helps perform serialization and deserialization in python.

Serializing

In simple words serializing is a way to write a python object on the disk that can be transferred anywhere and later de-serialized (read) back by a python script.

Serialization, De-serialization

Here’s how it’s done:

After running the above code, a file called “mlp_nn.pkl” is created which is the trained model that can be transferred anywhere and used after de-serialization as follows:

De-serializing

Flask setup

Let’s first set up the flask server on the local host and later deploy it on pythonanywhere for free.

Setting up flask-app over the localhost:

Make sure flask is installed. pip install flask

The following script starts the flask server on localhost and default port (5000) making the URL: http://127.0.0.1:5000/

Just paste http://127.0.0.1:5000/ on browser and press enter to see the server working.

  • App Route

app.route decorator is used for specifying the flask app route over the web.

“/” simply means the home that is “http://127.0.0.1:5000/”

“/price/” means http://127.0.0.1:5000/price/

  • Arguments

Arguments can be passed over these routes as follows:

http://127.0.0.1:5000/price/?date=12&month=3&year=2019

A general template would look like this:

https://[HOME]/[ROUTE]/?[var1_name]=[var1_value]&[var2_name]=[var2_value]

These arguments are parsed in the flask script as follows:

  1. request.args.get(‘date’) will return 12
  2. request.args.get(‘month’) will return 3
  3. request.args.get(‘year’) will return 2019

When you enter the URL into browser it actually sends a GET request over HTTP/HTTPS to the specified url.

In our case, we sent a GET request with arguments having date, month and year to the flask server to get back the prediction of our model.

  • jsonify

flask.jsonify() function will return a python dictionary as JSON. JSON format can be thought of as a python dictionary that can be nested any number of times and passed over the web. It’s a great way to structure data shared in between applications. It’s always advised to make use of JSON format when dealing with web APIs.

  • CORS

flask_cors package is used to fix CORS headers used for AJAX calls making cross-origin AJAX possible.

travelingYou can either read more about CORS here or just use it as it is. You CAN NOT ignore it because without correct CORS headers our API would be unusable by almost any other external applications. I personally struggled a lot with the CORS errors and had to read flask and some web development documentation to come up with this simple solution. If you have a better way to deal with CORS headers using flask please let me know in the responses.

Predict function

The code above has a predict function in the class my_bitcoin_predictor that needs a little explanation

Predict function can take input features and output the prediction of the model. This is the function actually deployed over the web using python flask. Instead of receiving the input features from the same environment in a python script or IPy notebook the features are now received through web servers like flask over the web to return back the predicted result.

For the sake of simplicity, I am skipping a lot of parts of my_bitcoin_predictor class and just writing the important bits.

File structure

A simple file structure looks like this:

Bitcoin predictor
 |
 +-- flask_app.py
 |    
 +-- predictor.py
 |
 +-- mlp_nn.pkl
 | 
 +-- data
 |  |  
 |  +-- bitcoin_price_historical_data.csv

I haven’t covered the part where the model gets trained by the historical data

The folder called Bitcoin predictor is the one needed to be uploaded on pythonanywhere

Hosting the flask app

flask app hosted on the localhost can’t be shared with others since it’s “local”. Host the script to a free python hosting website called pythonanywhere by following a few simple steps:

Step 1. Sign up for a new account.

For now, let's stick with the free account. You can buy better hosting plans for more diverse and complex needs. (There’s a nice tour for new users on this website which will help you)

Step 2. Add a new web app

Choose flask and whichever python version you want. I will be using the latest Python 3.7 for the sake of this tutorial. After creating the web app, you will get a URL that points to your flask endpoint. By default, it will display a message saying “Hello from Flask!”. Your endpoint looks something like this: [username].pythonanywhere.com

Add new web app

Step 3. Install dependencies

Your ML model must be using various external libraries such as sklearn, numpy, pandas etc. Install all of them in your pythonanywhere environment. To do so, open a new bash console and install libraries using pip. Make sure to install libraries with pip using the user option because we don’t get the superuser rights. Flask will already be installed, no need to install it again.

pip install --user flask_cors numpy pandas

Install dependencies

Step 4. Upload the files

Inside the default folder — /mysite/ you need to upload your complete folder. You can do it either using the files page on the website or using the bash console by using wget command to download your files.

Upload the files

Step 5. Reload the web app

Your endpoint will now act as an API to facilitate other applications.

Reload the web app

An example front-end application

You need to know basic HTML, CSS and JavaScript for this.

An example website

It’s observed that a lot of the machine learning enthusiasts don’t like web development which is why I have created a simple template that can be used to receive the input features through an HTML form and make a request to the API to fetch results. Even if you aren’t proficient with HTML-CSS-JavaScript you can easily edit this template.

Here’s the GitHub repository for the same.

For the example in discussion the edited JavaScript template would like this:

Host this website online and we’re done! Anyone anywhere can use this website to submit the features and get the predicted results from the ML model sitting at the back-end under various layers of data.

*I prefer uploading my website on GitHub and then making it live using netlify. All for free!

Hosting on personal server

pythonanywhere is only for educational purposes and lacks a GPU. For complex deep learning models relying heavily on CUDA and GPU based computation a stronger server is needed. Flask is made functional on a new server from scratch using a WSGI server like gunicorn behind a web server like nginx. A process control system like supervisor can also be used.

All of this requires some web development knowledge and writing about it would make this post too verbose. This webpage helped me setup flask on my server for the first time. It’s really simple and very well written.

You may also refer my image classification model that uses CUDA and was hosted on localhost. It shows how you can deal with image files using flask and pass it on to the model. I will be writing a complete post devoted to it soon.

Conclusion

Hosting and sharing machine learning models can be really easy. Creating android apps, chatbots and many more applications relying on the machine learning models back-end can now be created with great ease.

Thanks for reading. Please let me know in the responses if I made any mistakes or if something could’ve been done in a better way.

Python
Flask
Machine Learning
API
Data Science
Recommended from ReadMedium
avatarAlexander Nguyen
The resume that got a software engineer a $300,000 job at Google.

1-page. Well-formatted.

4 min read
avatarOm Kamath
Did Google Just Kill Streamlit?

Checking out Google’s Mesop Framework to see if it’s better than Streamlit.

5 min read
avatarAbhay Parashar
17 Mindblowing Python Automation Scripts I Use Everyday

Scripts That Increased My Productivity and Performance

14 min read
avatarPrasad Mahamulkar
Machine Learning Operations (MLOps) For Beginners

End-to-end Project Implementation

19 min read
avatarZoumana Keita
AI Agents — From Concepts to Practical Implementation in Python

This will change the way you think about AI and its capabilities

10 min read
avatarVishal Rajput
Prompt Engineering Is Dead: DSPy Is New Paradigm For Prompting

DSPy Paradigm: Let’s program — not prompt — LLMs

11 min read