Deploy Python Web apps

Full-stack web application for deploying a Deep Learning model, with Flask and Docker. Further discussion on operational management such as CI/CD will be added shortly.

What happens after having trained and validated your Deep Learinig model? It may be a simple MLP or the more trending Transformer, built by Tensorflow or Pytorch, for your research/application tasks. There will generally be interests for such model to be used by others, say collegues, clients, etc. The following content will give a concise overview about the best practices in deploying your model.

Deploy the learnt model as a REST API

Simply, the model can run with a simple call such as python inference.py. But the main interests lay in serving the model to the general public, as opposed to running only on your own laptop. In this spirit, we can deploy the model as a REST API, such that you can call curl to send request to the sever and obtain the response, which will be the results for inference. In doing so, you will probably do the following in the terminal for an example of image classification task. (Note the host name will be the relevant IP address in real cases. The localhost shown below means your own machine for illustration purpose.)

1

$ curl -X POST -F image=@query_image.jpg 'http://localhost:5000/predict'

Well, while this works, the caveat is that the call signature seems unnecessarily arcane. Things could be more straightforward. You don’t expect everyone to be a mechanic to drive a car, right?

Full-stack web application

By contrast, a user-friendly, intuitive, responsive, web application sounds much more attracting. Building a front-end for the API can save a lot of trouble for the end users. Besides, you get to integrate a lot of relevant sources at the same page to enrich the experience. The illustrative page below is taken from the web application I created and hosted by the AWS. Please refer to the page for additional details.

Dockerize it !

A tool that can largely stremline the development process is Docker, which serves as a platform for building, running, and shipping applications. A very practical scenario is when you want to share the application (say, still in beta stage) with other team members. You want a one-stop solution for them to quickly set up, without worrying about the operating system, the dependencies etc. Docker comes to the rescue. All they need are: download and run the image to build the application on their own machine.

Deploy to production

When deploying in production, the common practice is to use a dedicated combination of WSGI server, with a dedicated HTTP server in front of it (i.e. so-called reverse proxy). Typically, people will use Gunicorn, which is a pure Python WSGI server and nginx, which is production level HTTP server. How it works is: when serving your application with one of the WSGI servers, it’s often good practice to put a dedicated HTTP server in front of it, such that this “reverse proxy” can handle incoming requests, TLS etc. better than the WSGI server.

More conveniently, there are a number of services (e.g. hosting platforms) to facilitate such process. Typical examples include:

• PythonAnywhere
• AWS Elastic Beanstalk
• Google Cloud
• Microsoft Azure