catalyst

Why we included this project

Catalyst is a PyTorch framework built for people who are tired of rewriting the same training loop every time they start a new experiment. Instead of hand-rolling that boilerplate, you get reusable Runner and callback components, so a single supervised-runner class handles metrics, checkpointing, and logging for you. Experiments end up reproducible by construction: the same configuration code reliably produces the same results across runs, a real time-saver for teams training several model variants a week. The bundled utilities also cover deployment steps such as ONNX export, model tracing, quantization, and pruning, narrowing the gap between a notebook model and something you can actually ship. It is a good fit for researchers and small engineering teams who want more structure than a bare training script but less machinery than a heavyweight orchestration platform.