tango

Why we included this project

Anyone who has juggled a dozen ML experiments knows the pain of matching a result to the run that produced it. Tango attacks that by turning a research pipeline into discrete Python steps, wired together in a JSONnet config and executed with one CLI command. The caching is the part that earns its keep: every step's output is keyed by a hash of its inputs and version, so a rerun skips anything already sitting in the workspace instead of recomputing it. That matters when your codebase changes constantly and a full pipeline rerun would waste hours. It also has integrations for PyTorch, Hugging Face datasets, and Weights & Biases, plus distributed execution, so it drops into existing deep learning setups without forcing you to restructure anything.