Neural Architecture Search in Roboflow Train

Neural Architecture Search (NAS) is now available as a training engine for object detection and instance segmentation projects. Instead of manually tuning hyperparameters and running multiple separate training jobs, NAS automatically evaluates thousands of candidate architectures in a single run. This allows you to find the optimal balance of inference speed and accuracy tailored to your specific dataset and deployment hardware.

Using a unique weight-sharing strategy, NAS explores these configurations simultaneously at a fraction of the traditional compute cost. Once training is complete, you can evaluate the resulting models on a speed-vs-accuracy graph to deploy the version that best fits your performance requirements, often achieving better latency and accuracy than standard fine-tuning.

Read the announcement post

See documentation

Watch the walkthrough video

Last updated

Was this helpful?