PertNAS: Architectural Perturbations for Memory-Efficient Neural Architecture Search
Afzal Ahmad, Zhiyao Xie, Wei Zhang · 2023
Differentiable Neural Architecture Search (NAS) relies on aggressive weight-sharing to reduce its search cost. This leads to GPU-memory bottlenecks that hamper the algorithm’s scalability. To resolve these bottlenecks, we propose a perturbations-based evolutionary approach that significantly reduces the memory cost while largely maintaining the efficiency benefits of weight-sharing. Our approach makes minute changes to compact neural architectures and measures their impact on performance. In this way, it extracts high-quality motifs from the search space. We utilize these perturbations to perform NAS in compact models evolving over time to traverse the search space. Our method disentangles GPU-memory consumption from search space size, offering exceptional scalability to large search spaces. Results show competitive accuracy on multiple benchmarks, including CIFAR10, ImageNet2012, and NASBench-301. Specifically, our approach improves accuracy on ImageNet and NASBench-301 by 0.3% and 0.87%, respectively. Furthermore, the memory consumption of search is reduced by roughly 80% against state-of-the-art weight-shared differentiable NAS works while achieving a search time of only 6 GPU hours.