Efficient Constraint-Aware Neural Architecture Search for Object Detection

Egor Poliakov, Wei-Jie Hung, Ching-Chun Huang

Abstract: We propose an efficient Neural Architecture Search (NAS) method, named Zero-DNAS, for object detection tasks capable of discovering a suitable architecture under given memory and FLOPs constraints. NAS aims to explore the search space automatically to discover the best-performing network architectures for a given task. However, NAS is resource-consuming and usually requires hundreds of hours of GPU computations to discover a neural network architecture with good performance. Especially for more extensive and complex computer vision tasks such as object detection, the computing time and memory when conducting NAS would increase dramatically. Practically, the conventional sampling-based NAS methods do not guarantee the best possible solutions, whereas differentiable methods require substantial memory resources, making it challenging to apply in macro search space settings. In comparison, we propose a differentiable NAS paradigm with zero-cost proxy metrics and aims to determine the architecture within the constraints of memory and FLOPS. The experiments on object detection datasets show that our proposed algorithm can discover more accurate and faster architectures in a heavy macro search space in less than 2 NVIDIA 2080TI GPU hours.