NetAdapt: Platform-Aware Neural Network Adaptation for Mobile Applications

TL;DR

NetAdapt is an automated neural network adaptation algorithm that optimizes deep networks for mobile platforms by directly measuring and minimizing latency and energy, leading to better accuracy-latency trade-offs.

Contribution

It introduces a platform-aware adaptation method that uses empirical direct metrics for optimizing neural networks on mobile devices, outperforming existing indirect metric-based approaches.

Findings

Achieves up to 1.7× speedup in inference latency on ImageNet.

Outperforms state-of-the-art algorithms in accuracy-latency trade-offs.

Effectively adapts networks for both CPU and GPU mobile platforms.

Abstract

This work proposes an algorithm, called NetAdapt, that automatically adapts a pre-trained deep neural network to a mobile platform given a resource budget. While many existing algorithms simplify networks based on the number of MACs or weights, optimizing those indirect metrics may not necessarily reduce the direct metrics, such as latency and energy consumption. To solve this problem, NetAdapt incorporates direct metrics into its adaptation algorithm. These direct metrics are evaluated using empirical measurements, so that detailed knowledge of the platform and toolchain is not required. NetAdapt automatically and progressively simplifies a pre-trained network until the resource budget is met while maximizing the accuracy. Experiment results show that NetAdapt achieves better accuracy versus latency trade-offs on both mobile CPU and mobile GPU, compared with the state-of-the-art automated network simplification algorithms. For image classification on the ImageNet dataset, NetAdapt achieves up to a 1.7$\times$ speedup in measured inference latency with equal or higher accuracy on MobileNets (V1&V2).