yurijmikhalevich / lightning-pod-vision

An End to End ML Product Example

Lightning-Pod Vision

Overview

The core purpose of this repo is to help Lightning.ai users familiarize with the Lightning ecosystem by providing an example image classification product built with Lightning. The product is an intelligent agent behind a visual interface.

It uses Lightning, ReactJS + the Vercel ecosystem, FastAPI, Supabase, Zuplo, and Prisma.

This is a work in progress, especially the front end.

The docs can be viewed at visionpod-docs.vercel.app

The NextJS app can be viewed at visionpod.vercel.app

Programming Languages and Concepts

Languages

• Python: data engineering, and machine learning
• TypeScript and JavaScript: front-end design and visualization, and select back-end services
• Rust: exposure via project tooling (ruff, turbo) may require troubleshooting errors

Data Engineering and Database Management

• Python + Supabase-Py + Postgres: data acquisition and pushing to Supabase-Postgres

Model Development and Serving

• Lightning: developing and serving the model
• Optuna: hyperparameter optimization trials
• Weights and Biases: experiment management
• FastAPI: developing an API to serve the model with
• Locust: API load balance testing

User Interfaces

• Supabase: database, storage, and user management
• Prisma: Next-generation Node.js and TypeScript ORM
• TurboRepo + NextJS + React: front end
• Vercel: hosting the front end, and monitoring + analytics

Domain

• Topics:
  • Computer Vision
  • Image Classification
• Data Source: torchvision CIFAR10
• Domain Libraries: torchvision
• Model: Linear Encoder Decoder

Roadmap

The general outline for building this project is:

• data acquistion and storage
• do HPO trials and train from best trial config
• persist model
• build Model serving API with FastAPI
• determine feedback criteria for front end users
• design and build front end
• improve on feedback

Structure

The structure of the project is:

• docs-src is the docusaurus project
• lightning-app is the root Lightning Flow and App that will be deployed to Lightning
• next-app is the Next + React frontend deployed to Vercel
• pipeline handles data engineering for Supabase and Prisma
• requirements + requirements.txt helps CI/CD jobs install Python requirements
• research handles HPO and training
• tests are tests for visionpod
• visionpod is the python package and deep learning network

visionpod is the primary source code for the default branch. the following modules will be handled on separate production branches:

Module	Branch
docs-src	docs
lightning-app	l-app
next-app	n-app

Setup

You must have Python and NVM installed.

Installing NVM on macOS

brew install nvm
# update .zshrc or .bashrc
vim .zshrc

add the following to your profile:

to insert snippet, hit: shift + i

export NVM_DIR=~/.nvm
source $(brew --prefix nvm)/nvm.sh

to exit insert mode, hit: esc + shift + :wq!

then restart your terminal, and check the nvm version with:

nvm -v

proceed to installing node, enabling yarn, and creating a quickstart project with vite

nvm install lts
corepack enable

To setup a virtual development environment, in terminal, do:

python3 -m venv .venv/
source .venv/bin/activate
pip install -e ".[full]"
pre-commit install
deactivate
cd next-app
yarn config set nodeLinker node-module
yarn install
cd ..