[Paper Review] ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

Ningning Ma, Xiangyu Zhang+2 more|ArXiv.org|Jul 30, 2018

Advanced Neural Network Applications53 references236 citations

TL;DR

This paper argues that speed/latency on the target platform is the direct efficiency metric for CNNs, and introduces ShuffleNet V2 guided by practical guidelines, achieving better speed-accuracy tradeoffs than prior light-weight models.

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.

Motivation & Objective

• Motivate architecture design beyond FLOPs by examining direct speed on target hardware.
• Identify practical guidelines that correlate with real-world runtime performance.
• Propose an efficient network architecture (ShuffleNet V2) adhering to these guidelines.
• Demonstrate that the new architecture achieves superior speed-accuracy tradeoffs across platforms and workloads.

Proposed method

• Evaluate runtime performance of representative networks (ShuffleNet v1 and MobileNet v2) on GPU and ARM with optimized libraries.
• Derive guidelines from controlled experiments focusing on memory access cost (MAC), group convolution, network fragmentation, and element-wise operations.
• Propose channel split and shuffle-based building blocks that maintain equal-width channels (G1) while reducing MAC (G2) and fragmentation (G3).
• Construct ShuffleNet V2 blocks that avoid excessive fragmentation and unnecessary element-wise ops (G4).
• Benchmark ShuffleNet V2 against multiple baselines across FLOPs levels and tasks (ImageNet classification; COCO detection).
• Discuss compatibility with SE modules and potential extensions to larger models.

Experimental results

Research questions

• RQ1 How do direct speed metrics on GPUs and ARM devices relate to FLOPs in light-weight CNNs?
• RQ2 What architectural guidelines maximize actual speed without sacrificing accuracy on target platforms?
• RQ3 Can a channel-split/shuffle-building-block design provide high accuracy with low memory access costs and reduced fragmentation?
• RQ4 How does ShuffleNet V2 perform compared to ShuffleNet v1, MobileNet v2, and Xception across standard FLOP budgets?
• RQ5 Is ShuffleNet V2 effective for downstream tasks like object detection (COCO) in addition to ImageNet classification?

Key findings

• Direct speed correlates with platform characteristics and memory access; FLOPs alone are insufficient.
• Balanced channel widths (equal input/output channels) minimize memory access costs and improve speed.
• Excessive group convolutions increase MAC and slow down GPUs and ARM CPUs; moderate grouping is preferable.
• Reducing network fragmentation and avoiding excessive small fragmented operators improves parallelism and speed, especially on GPUs.
• Removing certain element-wise operations (ReLU, shortcuts) in bottleneck units yields speedups without high accuracy loss in tested settings.
• ShuffleNet V2 achieves superior speed-accuracy tradeoffs compared to ShuffleNet v1, MobileNet v2, and Xception across commonly used complexity budgets; it also generalizes to COCO object detection with competitive performance.