The idea in plain language.
Train a shared family of networks, then select a version that fits the target device.
How it works
Instead of training a fresh architecture for each hardware budget, the method progressively trains a larger network to support multiple depths, widths, kernel sizes, and resolutions. A deployment process then selects a suitable subnetwork from that shared model. The goal is to reduce repeated training while adapting to different latency and resource requirements.
What to keep in mind
The available subnetworks are limited by the trained family. Selecting a good configuration still requires a realistic estimate of device performance and task accuracy.
Source: Once-for-All: Train One Network and Specialize it for Efficient Deployment. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.