ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

TL;DR

This paper introduces ShuffleNet V2, an efficient CNN architecture guided by practical platform-aware design principles, emphasizing direct speed metrics over traditional FLOPs-based metrics, and demonstrates superior speed-accuracy tradeoffs.

Contribution

The paper provides practical guidelines for efficient CNN design considering real-world speed metrics and presents ShuffleNet V2, a new architecture optimized for these principles.

Findings

ShuffleNet V2 achieves state-of-the-art speed and accuracy tradeoff.

Platform-aware design improves real-world network efficiency.

Guidelines derived from experiments enhance CNN architecture design.

Abstract

Currently, the neural network architecture design is mostly guided by the \emph{indirect} metric of computation complexity, i.e., FLOPs. However, the \emph{direct} metric, e.g., speed, also depends on the other factors such as memory access cost and platform characterics. Thus, this work proposes to evaluate the direct metric on the target platform, beyond only considering FLOPs. Based on a series of controlled experiments, this work derives several practical \emph{guidelines} for efficient network design. Accordingly, a new architecture is presented, called \emph{ShuffleNet V2}. Comprehensive ablation experiments verify that our model is the state-of-the-art in terms of speed and accuracy tradeoff.