ML Diamond Pricing REST backend

Flask REST-Api for getting ML-calculated diamond price predictions based on the provided parameters.

Scrape diamond sales from the big online sellers of loose diamonds, train a price prediction model, and evaluate pricing on unseen constellations of diamond properties. Don't get ripped off by your local jeweller!

I recommend to exclusively train with and consider GIA certified diamonds for more consistent pricing.

Endpoints

(GET) /single_inference

Parameters:

Name	Type	Description
Shape	String	Diamond shape, Ex: "Round"
Color	String	Color grade, Ex: "F"
Clarity	String	Clarity, Ex: "VVS1"
Cut	String	Cut grade, Ex: "Ideal"
Carat	Double	Carat weight, Ex: 1.12

JSON response:

{
    'price_prediction': [predicted price]
}

(GET) /batch_inference

Parameters:

Name	Type	Description
batch	array	List of diamond objects containing all information described in the /single_inference endpoint

JSON response:

{
    'diamond_1': {
        'details': {...},
        'price_prediction': [predicted price]
    },
    ...
}

(POST) /train_model

Parameters: None

Response: {'success': [True or False]}

Prerequisites

For training the model

• A database of diamond sales saved as database.csv in ./data/, preferrably at least 10'000 sales. For structure, see example in scraper.py.

• Tensorflow Keras, sklearn, flask packages installed

For inferring diamond prices

• regression_estimator.h5 in ./model/. This is exported automatically when training the model.

How to run locally

• Clone the repo
• Run main.py
• Send requests to localhost:5000 and enjoy!

Performance

Model performance is very much dependent on the training data available. I have uploaded a small sample dataset of 20'000 diamond sales which you are welcome to use, but for increased performance I would recommend scraping your own larger dataset than that.

Example

This prediction model was trained on 50'000 sales.

Request: GET with payload {Shape: "Emerald", Color: "J", Clarity: "VS2", Cut: "Ideal", Carat: "1.5"}

Response: {price_prediction: 6865.705}

Reality:

This prediction is within +-1% of the only actual diamond I found with these exact properties!