alibaba/lightweight-neural-architecture-search
A collection of training-free neural architecture search methods for designing efficient vision models on CPU within limited time budgets.
Velocity · 7d
+0.3
★ / day
Trend
→steady
star history
This repository implements zero-cost NAS techniques including DeepMAD and Zen-NAS, which evaluate architecture quality without full training by analyzing entropy or gradient signals. It supports joint quantization and architecture search for mixed-precision models. Applications include object detection (MAE-DET) and video action recognition, all built on PyTorch.