ProxylessNAS is an open-source neural architecture search method from MIT’s HAN Lab that designs deep learning models directly for the target task and the target hardware — no proxy datasets, no guesswork about real device speed. Published at ICLR 2019, the project has about 1,450 stars on GitHub and its models are distributed through PyTorch Hub.
Classic neural architecture search was famously expensive and indirect: architectures were searched on small proxy datasets and then transferred, with latency estimated from lookup tables rather than measured on real devices. ProxylessNAS removed the proxies. It searches directly on the target dataset — ImageNet-scale training — while measuring real latency on the actual target hardware, whether that is a GPU, a CPU, or a mobile phone. A path-binarization scheme keeps memory costs manageable during search, cutting GPU hours by orders of magnitude compared with earlier methods. The resulting hardware-specialized models, including mobile variants, are integrated into PyTorch Hub for easy loading, and the approach has also been integrated into Microsoft’s NNI toolkit. The repository was last updated in 2024 and remains a standard reference for efficient, hardware-aware model design.
ProxylessNAS is free and open source under the MIT license. The code, pretrained models, and PyTorch Hub integration are all available at no cost; you only supply the compute for your own searches and training runs.
ML engineers and researchers building models for constrained hardware — phones, embedded devices, edge accelerators — who need architectures designed for their device’s real latency rather than generic models shrunk after the fact.