Skip to main content

Getting Started

This guide aims to get an instance of ModelBox started and have you train a model, and then look at the metadata and metrics collected by ModelBox to analyze the experiment, models and other artifacts.

There are three ways to get started with ModelBox. The docker-compose method is preferred if you have Docker installed on your machine. If you don't have Docker, GitPod would be the second best alternative. Lastly, you could download the ModelBox binary and run it locally as well, either with ephemeral storage or use the various storage dependencies.

Docker Compose​

This is the quickest way to get started if you have docker and docker-compose installed. 

docker compose --profile local up

This starts the ModelBox server with all the dependencies and a container with a Jupyter Notebook that demonstrates how to integrate a Pytorch trainer with ModelBox.

The ModelBox server hosts the API at the address - 172.21.0.2:8085

The Jupyter notebook with the tutorials is available at the address - https://localhost:8888

Train a Pytorch Model by following the notebook

Local Server with Ephemeral Metadata Storage​

  1. Install the dependencies

  2. Build ModelBox

  3. Generate the Server and client configs.

  4. Train a Model

Local Server with Local Datastores​

  1. Follow steps 1-3 from the above section which demonstrates how to run modelbox locally.

  2. Install Postgres/MySQL Server.

  3. Decide which metrics backend to use.

  4. Decide which blob storage backend to use.

  5. Start the server and train a model.