AutoDL

Why we included this project

AutoDL targets a problem most teams never bother to automate: getting a decent classifier without touching a knob. You hand it a dataset in almost any form, images, video, audio, text, or tabular data, and it runs feature extraction, model selection, architecture search, and hyperparameter tuning on its own, reporting results in as little as ten seconds. It won the NeurIPS AutoDL competition outright, so its model zoo reflects what actually worked on real benchmarks, mixing classical learners like LightGBM with deep models such as ResNet and BERT. Teams that want a fast, hands-off baseline for multi-label classification across mixed modalities will find it worth studying or adapting before building their own pipeline. One caveat: this is research-grade code that has not seen much recent maintenance, so treat it as a starting point or learning resource rather than a maintained production service.