Nozama: Image-based Product Recommendation System

Online shopping is becoming more and more popular nowadays due to its convenience. However, finding a product is still a time-consuming and less interesting part of the online shopping journey because users are still required to type keywords describing the products as well as getting mismatched results due to the difference between users' keywords and products' description from the sellers.

One of the solutions to solve this problem is to provide to users a more intuitive way to find relevant products using image-based instead of keyword-based approach. This project will demonstrate a website application using machine learning model to recommends a list of similar products based on a given image.

How it works

Model: VGG16 is a popular CNN model used for extracting feature vectors, but its size is too heavy (1.1GB) for a small web app. In order to deploy the project to an affordable web application environment, my CNN model should be small enough yet still maintains its effectiveness. I solve this problem by using Xception model from Google and reducing the computational power by running the k-NN algorithms within only one product category predicting from the classification task. By using Xception, the model size is significantly reduced from 1.1GB to 0.1GB, enough to deploy the model on affordable web hosting like Google Cloud Free Tier AppEngine.

Data sets: My data sets contains 40k images (5000 images for each label of 8 labels).

Predict Product Category: For the Product Classification task, I train the Xception model with transfered learning and achieve 82% validation accuracy.

Extract Feature Vectors: The AvgGlobalPooling layer is used to extract all feature vectors of my training images and the input image.

Build k-NN indexer: I use Annoy to build the index of all extracted feature vectors and the product ids. Annoy is a memory-optimized k-NN library, it is also used by Spotify to handle their recommendation tasks.

Using k-NN engine for similarity retrieval: When users input (upload) an image, the system will extract the feature vectors and predict the product category that the input image belongs to. Then the k-NN engine will retrieve the nearest 30 feature vectors and their product id in the predicted product category.

Getting Started

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes. See deployment for notes on how to deploy the project on a live system.

Prerequisites

Make sure you have python 3 and Anaconda installed on your environment.

The installing process is demonstrated in Linux and Macos environments.

This project is deployed using Google Cloud Platform (GCP). GCP provides free tier service to host a small web application without paying any fee so feel free to use it to test my project.

GCP also gives you $300 free credits to use within 1 year. That's a lot of money to use for machine learning and professional application deployment service.

Google Cloud Platform

The dataset

If you want to train your own model using my datasets, you can download my datasets from here: Download

The datasets contains 8 labels with 5,000 images each.

The datasets are provided by Standford University at this link. Please cite the following if you use the data:

Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering R. He, J. McAuley WWW, 2016 pdf

Image-based recommendations on styles and substitutes J. McAuley, C. Targett, J. Shi, A. van den Hengel SIGIR, 2015 pdf

The metadata CSV

In order to retrieve information of the products, you need to add the product's metadata to your database service. I use Google Firestore as the database service. Please refer to the notebook notebook/csv_to_firestore.ipynb to import the metadata to Google Firestore.

Download the metadata CSV here: Download

Google FireStore Documentation

Project structure

app: contains source codes of the web application. =>blueprints: contains source codes of Flask modules =>middlewares: contains source codes of the authentication service. As this feature is not mandatory, you can ignore this folder. =>models: contains source codes of the models using to communicate with Firestore to store and get data. =>static: contains all front-ends files and the trained models =>templates: contains the HTML template files =>uploads: used to store uploaded photo from the front-end

functions: contains source codes of the recommendation engine to be used as a standalone service.

notebooks: contains Jupyter notebooks for data preprocessing and machine learning tasks.

Installing

Virtualenv

pip install virtualenv

cd <git cloned folder>

virtualenv -p python3 env

source env/bin/activate

Install required libraries for the web application

cd app

pip install -r requirements.txt

Deployment

Create a GCP project

1. Sign into Google Cloud Console with your Google Account.
2. Click Select a project or the name of your existing project at the top of the page.
3. Click New project and follow the prompts on screen.

• Enable Billing. Google Cloud Platform provides $300 credit for new customers, and your usage may be eligible for Google Cloud Platform free-tier. For more information, see GCP Free Tier.
• Install Google Cloud SDK.

Authentication to GCP API

Authentication refers to the process of determining a client's identity. Authorization refers to the process of determining what permissions an authenticated client has for a set of resources. That is, authentication refers to who you are, and authorization refers to what you can do.

1. In the GCP Console, go to the Create service account key page.

   GO TO THE CREATE SERVICE ACCOUNT KEY PAGE

2. From the Service account list, select New service account.

3. In the Service account name field, enter a name.

4. From the Role list, select Project > Owner.

   Note: The Role field authorizes your service account to access resources. You can view and change this field later by using the GCP Console. If you are developing a production app, specify more granular permissions than Project > Owner. For more information, see granting roles to service accounts.

5. Click Create. A JSON file that contains your key downloads to your computer.

Setting the environment variable

If you plan to use a service account, you need to set an environment variable.

Provide authentication credentials to your application code by setting the environment variable GOOGLE_APPLICATION_CREDENTIALS. Replace [PATH] with the file path of the JSON file that contains your service account key, and [FILE_NAME] with the filename. This variable only applies to your current shell session, so if you open a new session, set the variable again.

export GOOGLE_APPLICATION_CREDENTIALS="[PATH]"

For example:

export GOOGLE_APPLICATION_CREDENTIALS="/home/user/Downloads/[FILE_NAME].json"

Create a bucket on Cloud Storage

Google Cloud Storage is a unified storage for objects with global caching; you will use it to store product images. The API is enabled by default when you create the Google Cloud Platform; follow the steps below to upload an image to Cloud Storage, which we will use in later steps:

• Open Cloud Storage Browser in Google Cloud Console.
• Click Create Bucket.
• Type in the name of the bucket. Write down this name; it uniquely identifies the bucket in Google Cloud.
• Click Create.

Cloud Firestore

Cloud Firestore is a NoSQL database. If you are not familiar with the concept, think of Cloud Firestore as a large-scale dictionary where data is stored in key-value pairs. The solution provides cheap, blazing fast storage for structured data, though it lacks the flexibility of traditional SQL databases. In Fansipan Store you will use it to store structured data, such as product details and order information.

To set up Firestore:

• Open the Firestore page in Google Cloud Console.
• In the Welcome page, Select Cloud Firestore in Native Mode.
• Select a location, such as asia-east2(Hong Kong).
• Click Create Database.

Document is the unit of data storage in Firestore, which can be grouped into collections. It is schemaless, which means that you do not have to specify the structure of data beforehand. The steps below create a document in the Products collection, which we will use later:

• Open the Firestore Data page in Google Cloud Console.
• Click Start Collection.
• Type in products_amazon as the collection ID.
• Leave the Document ID field empty. Cloud Firestore will automatically generate one for your document. The ID is the unique identifier of your document in the collection.
• Click Save.

FireBase SDK

Firebase is a Google developed mobile/web app development platform that solves common app development challenges, such as authentication, analytics, storage, etc. You need to setup Firebase SDK in order to make the web application work with Firestore.

To set up the Firebase SDK:

• Open Firebase Console and sign in with your Google account.
• Click Add project. In the Project Name drop-down menu, select your Google Cloud Platform project. This adds Firebase to your Google Cloud Platform project.
• Read the terms, and click Add Firebase.
• Click Add app. Pick Web Platform (</>). The configuration for your Firebase project will show up in a pop-up window.

Create app/static/js/initFirebase.js with following codes. Replace YOUR-API-KEY, YOUR-AUTH-DOMAIN, YOUR-DATABASE-URL, YOUR-PROJECT-ID, YOUR-STORAGE-BUCKET, and YOUR-MESSAGING-SENDER-ID with values of your Firebase configuration. This is the Firebase configuration for the HTML javascripts.

// Script for configuring Firebase.
// See https://firebase.google.com/docs/web/setup for more information.

var config = {
    apiKey: "YOUR-API-KEY",
    authDomain: "YOUR-AUTH-DOMAIN",
    databaseURL: "YOUR-DATABASE-URL",
    projectId: "YOUR-PROJECT-ID",
    storageBucket: "YOUR-STORAGE-BUCKET",
    messagingSenderId: "YOUR-MESSAGING-SENDER-ID"
};

firebase.initializeApp(config);

Create app/firebase_config.json** with following codes. Similarly, replace the placeholders with values of your Firebase configuration. This is the Firebase configuration for Firebase Admin SDK.

{
    "apiKey": "YOUR-API-KEY",
    "authDomain": "YOUR-AUTH-DOMAIN",
    "databaseURL": "YOUR-DATABASE-URL",
    "projectId": "YOUR-PROJECT-ID",
    "storageBucket": "YOUR-STORAGE-BUCKET",
    "messagingSenderId": "YOUR-MESSAGING-SENDER-ID"
}

Deployment to GCP AppEngine

app.yaml

Create app/app.yaml with following codes. Edit GCP_PROJECT, GCS_BUCKET, and FIREBASE_CONFIG.

runtime: python37
instance_class: F4_1G

env_variables:
  GCP_PROJECT: "YOUR-PROJECT"
  GCS_BUCKET: "YOUR-GCS-BUCKET"
  FIREBASE_CONFIG: "firebase_config.json"

handlers:
  - url: /static
    static_dir: static
  
  - url: /.*
    script: auto
    secure: always
    redirect_http_response_code: 301

Deploy to GCP AppEngine

cd app gcloud app deploy

Deploy the Recommendation Engine

In order to make recommendation engine to work, you need to deploy the recommendation engine to GCP Cloud Functions.

upload trained model and Annoy index files to GCP Bucket

• Download trained model and Annoy index files here: Link
• Open Cloud Storage Browser in Google Cloud Console.
• Access the Bucket you created above
• Create 3 folders models, ann_index, and ann_index/label_separated
• Upload the .h5 file to models
• Upload the .ann and .json files to ann_index/label_separated

deploy cloud function

cd <your git cloned project>

cd functions

cd recommend_v2

gcloud functions deploy recommend_v2 --runtime python37 --trigger-http --set-env-vars GSC_BUCKET=my_bucket-1 --region asia-east2 --memory 2048

After that, open app/blueprints/recommend/blueprint.py and add your Cloud Functions address to this code.

CLOUD_FUNC_RECOMMEND_URL = ''

Cloud Function Documentation

Notes for importing products data to Firestore

Please refer to the notebook notebook/csv_to_firestore.ipynb to import the metadata to Google Firestore.

Authors

• Binh Duong Thai - github

Acknowledgments

This project made possible by the following amazing library:

• Python Flask
• Bootstrap
• jQuery
• jQuery Raty
• jQuery FancyBox
• CropperJs
• TensorFlow
• Annoy

Reference for Machine Learning solution:

Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering R. He, J. McAuley WWW, 2016 pdf

Image-based recommendations on styles and substitutes J. McAuley, C. Targett, J. Shi, A. van den Hengel SIGIR, 2015 pdf

Image-based Product Recommendation System with Convolutional Neural Networks Luyang Chen, Fan Yang, Heqing Yang CS231n, 2017 pdf