RankNAS: A Differential NAS based Auto Rank Search towards Video LSTM Networks on Edge
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
DescriptionIt is crucial to develop optimized lightweight video LSTM network models on edge. Recent tensor-train method can form structured space-time features into tensor, which can be further decomposed into low-rank network models for lightweight video analysis. In this paper, a differentiable neural architecture search algorithm, called RankNAS, will be proposed to automatically decide tensor rank with trade-off between accuracy and complexity, for finding optimized low-rank video LSTM compared to previous manual-ranked models. Results from experiments show that RankNAS achieves 4.84x reduction in model complexity, 1.46x reduction in run time, and 3.86% accuracy improvement compared with the manual search based methods.