ENAS oriented layer adaptive data scheduling strategy for resource limited hardware

Abstract

Efficient Neural Architecture Search (ENAS) is an effective solution for building deep Convolutional Neural Network (CNN) models automatically. However, it is confronted with challenges when concerning deploying the searched networks on embedded platforms under limited resources. The key issue is the mismatch between the traditional data scheduling and the irregularity of layers from ENAS searched networks, which results in remarkable bandwidth pressure increment, and further leads to performance degradation and power consumption increase. In this paper, three alternative data scheduling patterns are constructed for different layers from ENAS searched networks, and a layer adaptive data scheduling strategy is proposed according to the constrained resources given by embedded platforms. Additionally, an adaptive architecture is also presented to deploy the searched networks efficiently, providing 4-10x performance speedup and 2.5-6x power consumption saving.