CoChat

ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

Open access
Ma, Ningning, Zhang, Xiangyu, Zheng, Hai-Tao, Sun, Jian · 2018 · Lecture notes in computer science

The DOI matches its Crossref record; 60 of the 60 references it cites were found in a registry or an index; 21 of the 35 sources its claims cite were checked for support.

Abstract

This work proposes to evaluate the direct metric on the target platform, beyond only considering FLOPs, and derives several practical guidelines for efficient network design, called ShuffleNet V2.

Overview

CoChat Summary

TL;DR

The paper argues that CNN architecture design should optimize direct speed/latency on the target platform rather than only FLOPs, because memory access cost, parallelism, and platform-specific optimization also determine runtime. It derives four practical guidelines and introduces ShuffleNet V2, which achieves better speed–accuracy tradeoffs than ShuffleNet v1, MobileNet v2, Xception, and others on GPU and ARM.

Key findings

  • Four guidelines are proposed: equal channel width minimizes memory access cost (MAC); excessive group convolution increases MAC; network fragmentation reduces parallelism; element-wise operations are non-negligible. Validation shows, for example, that using 8 groups is more than 2× slower than 1 group on GPU and up to 30% slower on ARM, and a 4-fragment structure is 3× slower than a 1-fragment structure on GPU.
  • ShuffleNet V2 outperforms prior lightweight models: at 40M FLOPs it is 3.5% and 3.7% more accurate than ShuffleNet v1 and MobileNet v2, respectively. At 500M FLOPs it is 58% faster than MobileNet v2, 63% faster than ShuffleNet v1, and 25% faster than Xception on GPU.
  • On ImageNet classification, ShuffleNet V2 models report top-1 error / GPU speed / ARM speed of 39.7% / 417 / 57.0 at 41 MFLOPs (0.5×), 30.6% / 341 / 24.4 at 146 MFLOPs (1×), 27.4% / 255 / 11.8 at 299 MFLOPs (1.5×), and 25.1% / 217 / 6.7 at 591 MFLOPs (2×).
  • On COCO object detection, ShuffleNet V2 achieves mmAP of 22.5, 29.0, 31.8, and 33.3 at 40M, 140M, 300M, and 500M FLOPs, respectively, and the ShuffleNet V2* variant reaches 23.7, 29.6, 32.2, and 34.2. For large models, ShuffleNet v2-50 at 2.3G FLOPs has 22.8% top-1 error versus 25.2% for ShuffleNet v1-50 and 24.0% for ResNet-50 at 3.8G FLOPs.

Methods

  • Controlled runtime experiments on two platforms: a single NVIDIA GeForce GTX 1080Ti with CUDNN 7.0, and a Qualcomm Snapdragon 810 with a Neon-based single-thread implementation.
  • Ablation benchmarks isolate design factors: channel ratios, group numbers, fragmented vs. non-fragmented blocks, and ReLU/shortcut element-wise operations, with FLOPs fixed across variants where applicable.
  • Architecture evaluation uses ImageNet 2012 classification at roughly 40, 140, 300, and 500+ MFLOPs, plus COCO object detection with Light-Head RCNN, comparing ShuffleNet V2 against ShuffleNet v1, MobileNet v2, Xception, DenseNet, CondenseNet, IGCV2/IGCV3, and auto-generated models.

Limitations

  • Stated: the MAC analysis assumes cache is large enough to store entire feature maps and parameters, so real MAC may deviate from the theoretical bound; the guidelines are platform-independent but validated only on GPU and ARM, and the paper argues the direct metric should be evaluated on the target platform.
  • Stated: COCO detection is not tested on ARM because PSRoI Pooling is unavailable there; speed measurements for CondenseNets are omitted due to lack of an efficient implementation; the reimplemented MobileNet v2 results are not as good as those reported in the original paper.
  • Evident from the text: evaluation is limited to ImageNet classification and COCO detection on two hardware platforms, with specific batch sizes and single-thread ARM testing, so generalization to other devices, tasks, and deployment settings is not established.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/978-3-030-01264-9_8

Cites

60 of 60 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

4 Valid · 7 Not valid · 8 Not checked · 58 Own results and interpretations

28 more claims

Create a free account to see all 77 claims and to verify the 1 reference still open.

See all claims — free account
  • It may state two separate claims as one. 36
  • Probably not a substantial claim. 11
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

  • Deep Residual Learning for Image Recognition

    He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 5×

    “For validation, we experimented with the “bottleneck” unit (1 × 1 conv followed by 3×3 conv followed by 1×1 conv, with ReLU and shortcut connection) in ResNet [5].”
    BackgroundMethodInfluentialSelf-citation
  • Very Deep Convolutional Networks for Large-Scale Image Recognition

    Simonyan, K., & Zisserman, A. (2014) · arXiv · doi:10.48550/arxiv.1409.1556 · Cited by 114,148 · cited 1×

    “Since the milestone work of AlexNet [15], the ImageNet classification accuracy has been significantly improved by novel structures, including VGG [25], GoogLeNet [28], ResNet [5,6], DenseNet [11], ResNeXt [33], SE-Net [9], and automatic neutral architecture search [18,21,39], to name a few.”
    Method
  • ImageNet classification with deep convolutional neural networks

    Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 1×

    “Since the milestone work of AlexNet [15], the ImageNet classification accuracy has been significantly improved by novel structures, including VGG [25], GoogLeNet [28], ResNet [5,6], DenseNet [11], ResNeXt [33], SE-Net [9], and automatic neutral architecture search [18,21,39], to name a few.”
    Method
  • ImageNet: A large-scale hierarchical image database

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009) · 2009 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2009.5206848 · Cited by 65,333 · cited 1×

    “Our ablation experiments are performed on ImageNet 2012 classification dataset [4,23].”
    Method
  • Densely Connected Convolutional Networks

    Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K. Q. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.243 · Cited by 48,566 · cited 9×

    “In DenseNet [11], to analyze the feature reuse pattern, the l1-norm of the weights between layers are plotted, as in Fig.”
    BackgroundMethodResultInfluential
  • Going deeper with convolutions

    Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., & Rabinovich, A. (2015) · 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2015.7298594 · Cited by 48,562 · cited 2×

    “Since the milestone work of AlexNet [15], the ImageNet classification accuracy has been significantly improved by novel structures, including VGG [25], GoogLeNet [28], ResNet [5,6], DenseNet [11], ResNeXt [33], SE-Net [9], and automatic neutral architecture search [18,21,39], to name a few.”
    Method
  • Microsoft COCO: Common Objects in Context

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., & Zitnick, C. L. (2014) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-10602-1_48 · Cited by 45,460 · cited 1×

    “To evaluate the generalization ability, we also tested COCO object detection [17] task.”
    Method
  • ImageNet Large Scale Visual Recognition Challenge

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380 · cited 1×

    “Our ablation experiments are performed on ImageNet 2012 classification dataset [4,23].”
    Method
  • Squeeze-and-Excitation Networks

    Hu, J., Shen, L., & Sun, G. (2018) · 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2018.00745 · Cited by 35,667 · cited 5×

    “When equipped with Squeezeand-excitation (SE) module [9], the classification accuracy of ShuffleNet v2 is improved by 0.”
    BackgroundMethodInfluential
  • Rethinking the Inception Architecture for Computer Vision

    Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., & Wojna, Z. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.308 · Cited by 32,138 · cited 1×

    “In the GoogLeNet series [27, 29, 28, 13] and auto-generated architectures [39, 21, 18]), a “multi-path” structure is widely adopted in each network block.”
    Background

Show all 60 references

Cited by

See all 7,356

Metadata

Authors
Ma, N., Zhang, X., Zheng, H.-T., & Sun, J.
Affiliations
  • Megvii (China) (CN)
  • Tsinghua University (CN)
Corresponding author
  • Ma, Ningning
  • Zhang, Xiangyu
Venue
Lecture notes in computer science
Published
2018
Pages
122-138
ISSN
  • 0302-9743 (ISSN-L)
  • 1611-3349
OpenAlex
W2883780447
Cited by
7,356
References
39
Citations by year
2026: 804 · 2025: 1317 · 2024: 1402 · 2023: 1328 · 2022: 972 · 2021: 753 · 2020: 518 · 2019: 251 · 2018: 11
FWCI
115.52
Type
conference paper
Language
en
Open access
Licence
Locations
  • Lecture notes in computer science (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
National Natural Science Foundation of China (61773229)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
Keywords
ShuffleNetV2, speed-accuracy trade-off, inference speed, memory access cost, lightweight neural networks

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol and qwen3.8-omni-flash).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. https://app.cochat.ai/papers/10.1007/978-3-030-01264-9_8
The paper
Ma, N., Zhang, X., Zheng, H.-T., & Sun, J. (2018). ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design. Lecture notes in computer science. https://doi.org/10.1007/978-3-030-01264-9_8

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.