High Performance SqueezeNext: Real time deployment on Bluebox 2.0 by NXP

Abstract

DNN implementation and deployment is quite a challenge within a resource constrained environment on real-time embedded platforms. To attain the goal of DNN tailor made architecture deployment on a real-time embedded platform with limited hardware resources (low computational and memory resources) in comparison to a CPU or GPU based system, High Performance SqueezeNext (HPS) architecture was proposed. We propose and tailor made this architecture to be successfully deployed on Bluexbox 2.0 by NXP and also to be a DNN based on pytorch framework. High Performance SqueezeNext was inspired by SqueezeNet and SqueezeNext along with motivation derived from MobileNet architectures. High Performance SqueezeNext (HPS) achieved a model accuracy of 92.5% with 2.62MB model size at 16 seconds per epoch model using a NVIDIA based GPU system for training. It was trained and tested on various datasets such as CIFAR-10 and CIFAR-100 with no transfer learning. Thereafter, successfully deploying the proposed architecture on Bluebox 2.0, a real-time system developed by NXP with the assistance of RTMaps Remote Studio. The model accuracy results achieved were better than the existing CNN/DNN architectures model accuracies such as alexnet_tf (82% model accuracy), Maxout networks (90.65%), DCNN (89%), modified SqueezeNext (92.25%), Squeezed CNN (79.30%), MobileNet (76.7%) and an enhanced hybrid MobileNet (89.9%) with better model size. It was developed, modified and improved with the help of different optimizer implementations, hyper parameter tuning, tweaking, using no transfer learning approach and using in-place activation functions while maintaining decent accuracy.