A Hands-on Tutorial

Train and Deploy a real-time ML model

CometML + Cerebrium = 🚀

Contents

• The problem

• Run the whole thing in 5 steps

• Lectures

  • 1. Model training
  • 2. Model deployment as REST API
  • 3. Test API endpoint

• Next steps

The problem

Training models inside notebooks is easy. Unfortunately, this is not enough when you want to build complete ML solutions for real-world problems.

In this hands-on tutorial you will learn how to

• train an ML model that predicts crypto prices, and
• deploy this model as a REST API

We will use Serverless ML tools to

• track experiment runs and publish our best model to the registy, with CometML.
• deploy the model as a REST API, with Cerebrium.

Without further ado, let's get to work!

Run the whole thing in 5 steps

1. Create a Python virtual environment with all project dependencies with

   $ make init
   

2. Set your API keys for CometML and Cerebrium as environment variables.

   $ . ./set_environment_variables.sh
   

3. Download historical data from Coinbase and save it locally to disk

   $ make data
   

4. Train ML model

   $ make train
   

5. Deploy the model

   $ make deploy
   

6. Test the endpoint works

   $ make test-endpoint
   

Lectures

1. Model training

2. Model deployment as REST API

3. Test API endpoint

4. Automation

Next steps

TODOs

• Get data from Coinbase
• Transform ts data into supervised ML data
• Engineer a few features using technical indicators
• Train a decent model with linear regression
• Fix bug when computing tech indicators from raw prices
• Deploy the modelLightGBMRegressor
• Webhooks
  • Add secrets to GitHub
  • Trigger workflow from Comet ML model registry
• Optimize model hyper-parameter with out-of-time cross validation use