A simple project to test deploying a trained models using Azure Functions.

The model is bears.pkl - an 87MB file. See the notebook in the panda-or-not repository for how it is trained.

The Dockerfile means the entire thing can be deployed to Azure Functions (serverless compute).

This rest of this README serves as a quick guide / reference for how to create a deployable image and run it on Azure

Local Setup

Requirements

• Linux (This guide has been tested with Ubuntu 16.04, xenial)
• Python 3.6 (only Python runtime currently supported by Azure Functions)
• Azure Functions Core Tools version 2.x
• Azure CLI
• Docker

Setup project directory

mkdir panda-or-not-serverless-azure
cd panda-or-not-serverless-azure
python -m venv .env
source .env/bin/activate

Create an Azure Function Project, and select a Python runtime

$ func init --docker

Create Function

$ func new --name panda-or-not --template "HttpTrigger"

Install dependencies

$ pip install fastai torch torchvision
$ pip freeze > requirements.txt

Update Function

Modify init.py and function.json (as in this project)

Copy model

Copy the trained model, bears.pkl, to .

Test function

Start the function on your local machine

$ func host start

Validate the function

$ curl -X POST -H "Content-Type: application/json" -d '{"url": "https://media.pri.org/s3fs-public/styles/story_main/public/images/2019/11/2019-11-19-beibeipanda.jpg"}' http://localhost:80/api/panda-or-not

Output should be something like:

request_json['url']: https://media.pri.org/s3fs-public/styles/story_main/public/images/2019/11/2019-11-19-beibeipanda.jpg,pred_class: panda_bear

Docker Setup

Build a Docker image

$ docker build -t <DOCKER_HUB_ID>/panda-or-not .

Test the Docker image

Run the Docker image on your local machine on port 8080

$ docker run -p 8080:80 -it <DOCKER_HUB_ID>/panda-or-not

Validate the service from another terminal window

$ url -X POST -H "Content-Type: application/json" -d '{"url": "https://media.pri.org/s3fs-public/styles/story_main/public/images/2019/11/2019-11-19-beibeipanda.jpg"}' http://localhost:8080/api/panda-or-not

Push Docker image to Docker Hub

Log in to Docker from the command prompt

$ docker login --username <DOCKER_HUB_ID>

Push the Docker image created earlier to Docker Hub

$ docker push <DOCKER_HUB_ID>/panda-or-not

Azure Setup

The following assumes you have a resource group named dsvmfastai.

Create Storage Account

$ az storage account create --name deeplearningprotostorage --location "West Europe" --resource-group dsvmfastai --sku Standard_LRS

Create a Linux App Service Plan

$ az appservice plan create --name deeplearningprotoservice --resource-group dsvmfastai --sku B1 --is-linux

Create the App & Deploy the Docker image from Docker Hub

$ az functionapp create --resource-group dsvmfastai --name pandaornot --storage-account  deeplearningprotostorage --plan deeplearningprotoservice --deployment-container-image-name arnts/panda-or-not

Configure the function app

$ storageConnectionString=$(az storage account show-connection-string --resource-group dsvmfastai --name deeplearningprotostorage --query connectionString --output tsv) 

$ az functionapp config appsettings set --name pandaornot --resource-group dsvmfastai --settings AzureWebJobsDashboard=$storageConnectionString AzureWebJobsStorage=$storageConnectionString

Run your Azure Function

After the previous command, it can take 15-20 minutes for the app to deploy on Azure.

$ az functionapp list --resource-group dsvmfastai

Validate the remote service

$ curl -X POST -H "Content-Type: application/json" -d '{"url": "https://media.pri.org/s3fs-public/styles/story_main/public/images/2019/11/2019-11-19-beibeipanda.jpg"}' https://pandaornot.azurewebsites.net/api/panda-or-not

Output should be

request_json['url']: https://media.pri.org/s3fs-public/styles/story_main/public/images/2019/11/2019-11-19-beibeipanda.jpg, pred_class: panda_bear