Research Playground

Inspired by ashleve/lightning-hydra-template and NVIDIA/NeMo

It's a library I created for efficient ML/DL research on various tasks.

Key features:

🚀 Production-ready training pipelines
🧠 Actual model implementations
⚡️ Easy configuration management with Hydra
📊 Experiment tracking with Weights & Biases
🔧 Modular architecture for quick prototyping
🐳 Docker support for reproducible environments
☸️ Multi-GPU training with K3s and Terraform (soon)

Getting Started

Install asdf to manage different tools' runtime versions.
Update .env.example to your needs.
Setup your training Hydra config in configs/experiments/ folder.

Choose between local development outside or inside Docker container.

Outside of Docker (not recommended):

make init-local
poetry shell && python3 src train --experiment <experiment_name>

Inside Docker:

make init && make build && make run
python3 src train --experiment <experiment_name>

Notes

Use make help to see all available commands.
Use python3 src --help to see all available CLI arguments.