silasxue/cbirapp

Content-Based Image Recommender App on PCF

This repo contains the code required to build an image recommender app on PCF. The image recommender modeling pipeline is built within the Pivotal Greenplum Database. The app that is hosted on PCF allows users to input an image URL and then invokes the modeling pipeline in the Greenplum Database to retrieve visually similar images. For the modeling pipeline, check out: https://github.com/gautamsm/data-science-on-mpp/tree/master/computer_vision

The app is an instantiation of the Boilerplate for Flask Apps for Data Science on PCF (https://github.com/vatsan/dspcfboilerplate)

Pre-requisites

The conda_requirements.txt file lists all the python packages that are available via conda and are pre-requisites for this app. The requirements.txt file lists all python packages that are only available through pip and are pre-requisities for this app.

Code Organization

dspcfboilerplate/ # root level folder containing all package files & app files
    README.md # This file
    conda_requirements.txt # file containing all `conda` packages needed by this app.
    requirements.txt # file containing all `pip` packages needed by this app (not available via `conda`)
    deploy # bash script to deploy this app (either locally or on PCF)
    setup.py # python packaging tools
    MANIFEST.in #Manifest file for python packaging (what files to include into the python package)
    LICENSE.txt #License for this app    
    predim/
        server.py #Main module containing all controller code
        static/
           css/ #All user specified css. These will be bundled & minified into "gen/user_css.css" by Flask-Assets.    
           data/   
           img/    
           js/ #All user specified javascript. These will all be bundled & minified into "gen/user_js.js" by Flask-Assets.    
           vendor/ #All bootstrap.js related files (css & javascript)
        templates/ #HTML templates
           layout.html  #Base layout from which every page will inherit. This also contains javascript & css inserts
           home.html #home page template
           about.html #about page template
           contact.html #contact page template
           cbirapp_home.html #cbir home page

Starting the app locally

1. Create a file dspcfapp/user.cred with the relevant database access credentials like the following. This file will not be added to your version control (the .gitignore file will filter it out):

    [database_creds]
    host: <YOUR HOSTNAME>
    port: <YOUR PORT>
    user: <YOUR USERNAME>
    database: <YOUR DATABASE>
    password: <YOUR PASSWORD>
   

2. Ensure your local machine can talk to the environment where the data resides in (ex: you may need to connect to a VPN if your data resides on a BDS cluster behind a firewall).

3. Run the following from the root directory

    ./deploy
   

This will bring up the app on http://localhost:9090

Pushing the app to PCF

1. Push the app to your PCF instance (assuming you've set one up)

    dsmiot [master●●] cf push dspcfboilerplate  -f manifest.yml    
   

2. Create User Provided Service for database credentials (first time only)

    dsmiot [master●●] cf cups dspcfboilerplatecreds -p '{"host":"<HOST>","user":"<USER>","password":"<PASSWORD>", "database":"<DATABASE>", "port":"<PORT>" }'
   

3. Bind the User Provided Service to the app (first time only)

    dsmiot [master●●] cf bind-service dspcfboilerplate dspcfboilerplatecreds
   

The app should now be accessible on your PCF instance.

Screenshots

Here is a sample screenshot of the image recommender app in action: