ml-model-management-docker

Overview

This github repo demonstrates how to use a pickle ML mode from within a Flask (http://flask.pocoo.org/) application.

Setup

Download the trained model

The project doesn't include the trained & saved model (due to the model's size). A working model can be downloaded from here https://olifileexchange.blob.core.windows.net/ml-share/trained_model.pkl (537MB). After you downloaded the model, place the pkl file in the folder of the 'flask-web-app'.

Update the code in main.py ('flask-web-app') and replace the placeholder YOUR_MODE_FILE_NAME with the filename of your pickle model. Your main.py file should look similar to this:

with open(os.path.join(APP_ROOT, 'trained_model.pkl'), 'rb') as f:

ATTENTION
This sample was tested with a Python 3.6 environment. Other versions weren't tested.

Run

Locally

Open a commandline navigate to folder 'flask-web-app' and run the following command

python main.py

This will start the flask web application (the link will be displayewd in your commandline window). Open a browser and navigate to that url. You should see 'Hi'.

After you verfied that the app is running and works. Let's create a prediction. Use curl (https://curl.haxx.se/) or Postman (https://www.getpostman.com/) and send a POST request to the web app using the '\predict' path (e.g. http://localhost:5000/predict).

Ensure that you set the request's content-type (request header) to application/json and that you provide an image url in the body. For example:

{ 
    "image_url": "https://shop.epictv.com/sites/default/files/ae42ad29e70ba8ce6b67d3bdb6ab5c6e.jpeg" 
} 

Locally (Docker)

NOTE
Before running the app using Docker please ensure that it successfully runs locally (outside of a container).

Build the image

Open a commandline and navigate to your Flask web application folder. Execute the following command to create the Docker image.

Attention
Please replace the placeholder YOUR_IMAGE_NAME with your Docker image name. Also, notice the . character at the end of the command!

docker build -t YOUR_IMAGE_NAME:latest . 

Run the container

docker run -d -p 5000:5000

Test the app

Open a browser and navigate to http://localhost:5000.

Azure

Push the image to a Docker container registry (e.g. Docker Hub)

Since we will push the image to Docker Hub (https://hub.docker.com/), please ensure that you have a valid Docker Hub account.

Follow the steps listed here 'https://docs.docker.com/docker-cloud/builds/push-images/' to push the image to Docker Hub

Note
Alternatively, you can also publish your Docker image to a private doker registry like Azure Container registry (https://azure.microsoft.com/en-us/services/container-registry/).

Create an new 'Web App for Containers' on Azure

NOTE
For further information around 'Azure Web App for Container' please see https://azure.microsoft.com/en-us/services/app-service/containers/

Once the image is available in the container registry (e.g. Docker Hub) please follow this tutorial https://docs.microsoft.com/en-us/vsts/build-release/apps/cd/deploy-docker-webapp#create-an-azure-web-app-to-host-a-container to create a new Web App for Containers on Azure.

Since we pushed the Docker image to Docker Hub (and not to Azure Container registry) please ensure that you select 'Docker Hub' in step #3.

Most Docker images have environment variables that need to be configured. If you are using an existing Docker image built by someone else, the image may use a port other than 80. You tell Azure about the port that your image uses by using the WEBSITES_PORT app setting. Here you need to set WEBSITES_PORT to 5000

Test the app

Wait until the web app was successfully created. Then open a browser and nagivate to the Azure web app's public IP address.