// the find
microsoft/nni
An open source AutoML toolkit for automate machine learning lifecycle, including feature engineering, neural architecture search, model compression and hyper-parameter tuning.
Microsoft Research's AutoML toolkit covering hyperparameter optimization, neural architecture search, model compression, and feature engineering under one roof. It targets ML researchers and practitioners who want to automate the expensive search steps of training deep learning models. The breadth is the selling point and also the problem.
The HPO module has real depth — TPE, BOHB, Hyperband, PBT, SMAC, and several Bayesian variants are all available and well-tested, not just listed. The NAS module unifies multi-trial and one-shot strategies (DARTS, ENAS, ProxylessNAS) behind a common API so you can swap search strategies without rewriting your model definition. The training service abstraction is genuinely useful: the same experiment config runs locally, over SSH, on AML, or on Kubernetes without changing your trial code. The web UI for experiment tracking is solid — parallel coordinate plots, intermediate metric curves, and trial comparison are all there.
The project is effectively abandoned. The 'What's New' section in the README stops at June 2022, v3.0 never left RC status despite being announced in May 2022, and the last meaningful push was July 2024. If you hit a bug, you're on your own. The breadth-over-depth tradeoff hurts in practice: for HPO alone, Optuna is more actively maintained and has a larger ecosystem; for NAS, there are dedicated frameworks with more current algorithms; for compression, PyTorch's own pruning and quantization APIs have caught up. The feature engineering module is visibly the weakest pillar — it amounts to two selectors (gradient-based and GBDT) with almost no documentation. The `docs/_removed/` directory and chunked dependency files suggest the API has been substantially rewritten more than once, so tutorials you find online are likely stale.